This page intentionally left blank
Two-Dimensional Wavelets and their Relatives
Two-dimensional wavelets offer a num...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

This page intentionally left blank

Two-Dimensional Wavelets and their Relatives

Two-dimensional wavelets offer a number of advantages over discrete wavelet transforms when processing rapidly varying functions and signals. In particular, they offer benefits for real-time applications such as medical imaging, fluid dynamics, shape recognition, image enhancement and target tracking. This book introduces the reader to 2-D wavelets via 1-D continuous wavelet transforms, and includes a long list of useful applications. The authors then describe in detail the underlying mathematics before moving on to more advanced topics such as matrix geometry of wavelet analysis, three-dimensional wavelets and wavelets on a sphere. Throughout the book, practical applications and illustrative examples are used extensively, ensuring the book’s value to engineers, physicists and mathematicians alike. Jean-Pierre Antoine is a Professor of Mathematical Physics at the Institut de Physique

Th´eorique, Universit´e Catholique de Louvain. Romain Murenzi is currently Minister of Education, Science, Technology, and Scientific

Research of the Republic of Rwanda, on leave of absence from the Department of Physics, Clark Atlanta University, Atlanta, Georgia. Pierre Vandergheynst is a Professor at the Signal Processing Institute, Swiss Federal

Institute of Technology, Lausanne. Syed Twareque Ali is a Professor at the Department of Mathematics and Statistics, Concordia University, Montr´eal.

Two-Dimensional Wavelets and their Relatives Jean-Pierre Antoine Institut de Physique Th´eorique, Universit´e Catholique de Louvain

Romain Murenzi CTSPS, Clark Atlanta University, Ministry of Education, Science, Technology and Scientific Research, Rwanda

Pierre Vandergheynst Signal Processing Laboratory, Swiss Federal Institute of Technology

Syed Twareque Ali Department of Mathematics and Statistics, Concordia University

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge , UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521624060 © J.-P. Antoine, R. Murenzi, P. Vandergheynst and S. Twareque Ali 2004 This publication is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format 2004 - -

---- eBook (EBL) --- eBook (EBL)

- -

---- hardback --- hardback

Cambridge University Press has no responsibility for the persistence or accuracy of s for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Prologue

page ix

1

Warm-up: the 1-D continuous wavelet transform

1.1 1.2 1.3 1.4 1.5 1.6 1.7

What is wavelet analysis? The continuous wavelet transform Discretization of the CWT, frames Ridges and skeleton The discrete WT: orthonormal bases of wavelets Generalizations Applications of the 1-D CWT

1 5 10 14 19 24 29

2

The 2-D continuous wavelet transform

32

2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8

Derivation Basic properties of the 2-D CWT Implementation and interpretation of the 2-D CWT Discretization, frames Comparison with the 2-D discrete wavelet transform Bridging the gap: continuous wavelet packets and fast algorithms Steerable filters Redundancy: plus and minus

32 36 41 54 68 78 93 96

3

Some 2-D wavelets and their performance

97

3.1 3.2

Which wavelets? Isotropic wavelets

97 99

v

1

vi

Contents

3.3 3.4

Directional wavelets Wavelet calibration: evaluating the performances of the CWT

103 118

4

Applications of the 2-D CWT. I: image processing

125

4.1 4.2 4.3 4.4 4.5 4.6 4.7

Contour detection, character recognition Object detection and recognition in noisy images Image retrieval Medical imaging Detection of symmetries in patterns Image denoising Nonlinear extensions of the CWT

125 134 145 150 150 162 163

5

Applications of the 2-D CWT. II: physical applications

175

5.1 5.2 5.3 5.4 5.5 5.6

Astronomy and astrophysics Geophysics Applications in fluid dynamics Fractals and the thermodynamical formalism Texture analysis Applications of the DWT

175 192 197 205 210 212

6

Matrix geometry of wavelet analysis. I

214

6.1 6.2 6.3

Group theory and matrix geometry of wavelets Phase space analysis The case of Gabor wavelets

214 230 238

7

Matrix geometry of wavelet analysis. II

247

7.1 7.2 7.3 7.4

A group-adapted wavelet analysis The 2-D continuous wavelet transform 2-D wavelets on phase space The affine Poincar´e group

247 259 268 276

vii

Contents

8

Minimal uncertainty and Wigner transforms

281

8.1 8.2 8.3 8.4

Phase space distributions and minimal uncertainty gaborettes Minimal uncertainty wavelets Wigner functions Wigner functions for the wavelet groups

281 284 287 291

9

Higher-dimensional wavelets

300

9.1 9.2 9.3

Three-dimensional wavelets Wavelets on the 2-sphere and other manifolds Wavelet approximations on the sphere

300 308 332

10

Spatio-temporal wavelets and motion estimation

343

10.1 10.2 10.3 10.4 10.5

Introduction Spatio-temporal signals and their transformations The transformation group and its representations The spatio-temporal wavelet transform A motion estimation (ME) algorithm

343 344 348 352 356

11

Beyond wavelets

373

11.1 11.2 11.3 11.4

New transforms: ridgelets, curvelets, etc. Rate-distortion analysis of anisotropic approximations Sparse approximations in redundant dictionaries Algebraic wavelets

374 383 389 398

Epilogue Appendix Some elements of group theory References Index

413 415 431 455

Prologue

Wavelets are everywhere nowadays. Be it in signal or image processing, in astronomy, in fluid dynamics (turbulence), in condensed matter physics, wavelets have found applications in almost every corner of physics. In addition, wavelet methods have become standard in applied mathematics, numerical analysis, approximation theory, etc. It is hardly possible to attend a conference on any of these fields without encountering several contributions dealing with them. Correspondingly, hundreds of papers appear every year and new books on the topic get published at a sustained pace, with publishers strongly competing with each other. So, why bother to publish an additional one? The answer lies in the finer distinction between various types of wavelet transforms. There is, indeed, a crucial difference between two approaches, namely, the continuous wavelet transform (CWT) and the discrete wavelet transform (DWT). Furthermore, one has to distinguish between problems in one dimension (signal analysis) and problems in two dimensions (image processing), since the status of the literature is very different in the two cases. Take first the one-dimensional case. Beginning with the classic textbook of Ingrid Daubechies [Dau92], several books, such as those of M. Holschneider [Hol95], B. Torr´esani [Tor95] or A. Arn´eodo et al. [Arn95], cover the continuous wavelet transform, in a more or less mathematically oriented approach. On the other hand, the discrete wavelet transform is treated in many textbooks, more in the signal processing style, such as M. V. Wickerhauser [Wic94], M. Vetterli and J. Kovaˇcevi´c [Vet95], P. Wojtaszczyk [Woj97], or S. G. Mallat [Mal99], whereas others emphasize the algorithmic aspects, sometimes in a rather abstract way, for example, C. K. Chui [Chu92] or Y. Meyer [Mey94] (of course, there are many more on the market). Altogether these books tell a fascinating story, that is ideally depicted in the highly popular volume of B. Burke Hubbard [Bur98], which is based on interviews by the author with all the founding “fathers” of the theory (J. Morlet, A. Grossmann, I. Daubechies, Y. Meyer, etc.). It is a fact that DWT-inspired methods (multiresolution, lifting scheme, etc., that we shall describe in due time) constitute the overwhelming majority among the wavelet community, under the joint influence of electrical engineering (signal processing with ix

x

Prologue

filters and subband coding) and applied mathematics (numerical and algorithmic methods). Yet the CWT and, more generally, redundant representations of signals, offer distinct advantages in certain cases, as we shall see later. In two dimensions, that is, application to image processing, the situation is clearer. Discrete methods are somewhat trivial, since the basic structure is that of a tensor product, 2-D = 1-D ⊗ 1-D, enforcing a Cartesian geometry (x and y coordinates). Thus most textbooks on the DWT will cover, although briefly in general, the 2-D case as a straightforward extension of the 1-D setup. As for the 2-D CWT, it receives at best a cursory treatment in most cases. The raison d’ˆetre of the present volume is precisely to fill this gap in the literature and give a thorough treatment of the 2-D CWT and some of its applications in image processing and in various branches of physics. As a byproduct, we will also discuss in detail several extensions, such as 3-D wavelets, wavelets on the sphere or wavelets in space-time.

A historical note Before entering the subject proper, it may not be uninteresting to give some details on its origin, without pretension to completeness, of course; we are not historians. The first extension of the wavelet transform to imaging is due to Mallat [259,260], who developed systematically a 2-D discrete (but redundant) WT, combining the traditional concept of filter bank and the analogy with human vision. In fact, most of the concepts are indeed already present in the pioneering work of Marr [Mar82] on vision modeling, in particular the idea of multiresolution. Indeed, when we look at an object, our visual system works by registering first a global, low-resolution, image and then focusing systematically to finer and finer details. Thus, contrary to the 1-D case, the 2-D discrete WT preceded the continuous version. The 2-D continuous WT was born in a quite different way. The story starts in the coffee room of the Institut de Physique Th´eorique in UCL, Louvain-la-Neuve (LLN), in Spring 1987. Alex Grossmann from Marseille, one of the founding fathers of wavelets, was visiting J.-P. A., indeed they had already started to collaborate on the application of 1-D wavelets in NMR spectroscopy. Thus the two were discussing a possible Ph.D. topic for a young African student, called Romain Murenzi (R.M.). The latter had just concluded a Master’s thesis on five-dimensional quantum field theory, a subject hardly practical for a developing country! So the idea came up, why not try to do in two dimensions what had been so successful in 1-D, namely, wavelet analysis? The topic seemed tractable, involving moderate amounts of mathematics and some simple computing technology, and if it worked out, there could be very interesting practical applications. The problem was that nobody knew how to do it! The next summer, R.M. went down to Marseille and started to work with Grossmann and Ingrid Daubechies who happened to be there too. And when he came back 3 months later, the solution was clear. The key

xi

Prologue

is to start from the operations that one wants to apply to an image, namely, translations in the image plane, rotations for choosing a direction of sight, and global magnification (zooming in and out). The problem is to combine these three elements in such a way that the wavelet machine could start rolling (there are mathematical conditions to satisfy here). The result of R.M. was that the so-called similitude group yields a solution (actually, the only one). There remained to put it all together, to turn the mathematical crank and to apply the resulting formalism to a real problem, namely, 2-D fractals (the outcome of a visit of R.M. to Arn´eodo in Bordeaux), and the Ph.D. thesis was within reach [Mur90]. Several papers followed [12,13], more M.Sc. or Ph.D. students got involved over the years. We may cite Pierre Carrette, St´ephane Maes, Canisius Cishahayo, Pierre Vandergheynst, Y´eb´eni B. Kouagou, Laurent Jacques, Laurent Demanet. Each of them has brought his contribution to the edifice, small or big, but always useful. This is probably a good place for asking, why wavelets? After all, there are plenty of methods available for processing images. What is new here? A key fact is probably that wavelets are somehow a byproduct of quantum thinking. More precisely, it is an application of the quantum idea of a probe for testing an object, the result being given by the scalar product of the two functions (indeed the framework is a Hilbert space, that of finite energy signals). To get the transform, the probe is translated and scaled (zoom), and turned around in the 2-D case, and the result is plotted as a function of the corresponding parameters. (Actually the same could be said of the so-called Gabor or Windowed Fourier transform.) One gets in this way a highly flexible and efficient tool for signal/image processing, that sheds a different light and offers an alternative approach to many standard problems, in particular those involving the detection of singularities or discontinuities in signals. As somebody once remarked, wavelets do not solve all the problems, but they often help asking the right questions. Another sign of the quantum influence is the crucial role played by a unitary group representation, a tool largely absent in classical physics – and thus from signal processing as well. And it is no accident, in our opinion, that the crucial steps in developing wavelets were made by Alex Grossmann and Ingrid Daubechies, both educated as theoretical (quantum) physicists. Otherwise, it might have taken much longer for electrical engineers and mathematicians to meet!

About the contents of the book Now it is time to give some indications on the contents of the book. One can divide it into several stages. In a first part (Chapters 1–3), we develop systematically the continuous wavelet transform, first in one dimension (briefly), then in two dimensions. The emphasis here is on the practical use of the tool, with a minimum of mathematics. Then we devote two long chapters, 4 and 5, to applications. Three short chapters, 6–8,

xii

Prologue

set the general mathematical scene. This allows us, in Chapters 9 and 10, to describe wavelets in more general settings (3-D, sphere, space–time). In Chapter 11, finally, we discuss some recent developments that actually go beyond wavelets. This gradual structure is one of the original aspects of the book, in comparison with those on the market. Let us go into more details. As a warming up exercise, we begin, in Chapter 1, with a rather concise overview of the 1-D WT. This allows the reader to develop a feeling about the wavelet transform and to understand its success in signal processing. All aspects will be touched upon: the continuous WT, multiresolution and the discrete WT, various generalizations of the latter, some applications. One of the leitmotives is the role of redundancy, especially with respect to stability of the representation. Chapter 2, which forms the hard core of the first part, presents in a systematical way the theory of the 2-D CWT. As said above already, the starting point is to decide which elementary operations one wants to apply to an image. Choosing translations in the image plane, rotations (direction of sight), and global magnification (zooming in and out), together with the probe idea, leads uniquely to the 2-D CWT. We study in detail its basic properties: energy conservation, reconstruction formula, reproducing property, covariance under the chosen operations. Then we describe the interpretation of the WT as a singularity scanner and as a phase space representation of signals. Since the WT of a 2-D image is a function of four variables, visualizing it inevitably becomes problematic. Hence the need to reduce the number of parameters, either by fixing some of them, or integrating over them. This introduces a tool that will prove very useful in the applications, namely, the various partial energy densities, that is, the function obtained by integrating the squared modulus of the CWT over a subset of the parameters. In other words, various types of wavelet spectra, the analogs of the familiar power spectrum of a signal. As is well known in 1-D, the CWT is highly redundant, as one can expect from a transform that doubles the number of variables: one to two in 1-D, two to four in 2-D. This fact may be exploited in two ways. Either one limits oneself to a small subset of the transform, where most of the energy is concentrated, and thus one is led to the notions of local maxima, ridges and skeleton; or one discretizes the CWT and obtains wavelet frames. Such a representation is still redundant, but much less than the full CWT, and in many instances is a good substitute for a genuine orthonormal basis. An alternative is the so-called dyadic WT, originally due to Mallat, in which only the scale variable is discretized. Together with the latter, we also describe briefly the standard DWT, based on the multiresolution idea, and several generalizations, mostly the so-called lifting scheme. We conclude the chapter with a thorough discussion of a different scheme, called directional dyadic wavelet frames. Here, as in 1-D, there are two conflicting requirements: redundancy of the transform, which brings stability, and computing economy, that seeks fast algorithms. The formalism described here offers a good compromise.

xiii

Prologue

When it comes to treating a precise problem, the first question to ask is, which wavelet should one use? Thus there is a need for a sizable collection of them, well documented and calibrated. The aim of Chapter 3 is to provide this. The crucial distinction here is whether directions in the image are relevant or not. If they are not, a pointwise analysis suffices, and one can use rotation invariant (isotropic, radial) wavelets, the best known being the Mexican hat or LOG wavelet (already introduced by Marr [Mar82]). On the contrary, if directions must be detected, one needs a wavelet with a good orientation selectivity. The most efficient result is obtained with the so-called directional wavelets. These are filters living in a convex cone, with apex at the origin, in Fourier space. Examples are the 2-D Morlet wavelet and the family of conical wavelets. All these wavelets, and some more, are discussed in detail in Chapter 3, and their performances determined quantitatively. At this stage, the tool is ready and we turn to applications. Many of them are not easy to find, because they have appeared only in conference proceedings or in (unpublished) Ph.D. theses. For that reason, we have decided to present them in a rather detailed fashion, always giving original references, including personal websites when available. In each case, we emphasize the rationale for using wavelets in the particular problem at hand, rather than go into the technicalities. It is convenient (although not always unambiguous) to distinguish between two different fields of applications, image processing and physics. To the first type, the subject matter of Chapter 4, belong contour detection and character recognition; automatic target detection and recognition (for instance, in infrared radar imagery); image retrieval from data banks; medical imaging; detection of symmetries in patterns, in particular quasicrystals and other quasiperiodic patterns; and image denoising. The chapter concludes with two nonlinear extensions of the CWT, which both have important applications. The first one is contrast enhancement in images through an adaptive normalization. This technique, based on analogy with our visual system, may be of interest in medical imaging. Indeed typical images, such as those obtained by radiography or by NMR imaging, have rather weak contrast, which makes their interpretation sometimes difficult. The other problem we deal with is watermarking of images, which consists in adding an invisible “signature” (the watermark) to an image, that only the owner can recognize and is robust to manipulations. Clearly the field of image copyright offers a good market for such techniques. The novel method we present is based on the contrast analysis described previously, exploiting directional wavelets, and it turns out to be particularly efficient. The second class of applications, described in Chapter 5, concerns various fields of physics. Characteristically, they all belong to classical physics, as opposed to quantum physics, because the former relies much more on images. Indeed, there are very few applications of wavelet analysis in quantum problems. The first domain on which 2-D wavelets have made a substantial impact is astronomy and astrophysics, for several reasons. The Universe has a marked hierarchical structure.

xiv

Prologue

Nearby stars, galaxies, quasars, galaxy clusters and superclusters have very different sizes and live at very different distances. Thus the scale variable is essential and a multiscale analysis is in order. This, of course, suggests wavelet analysis, and indeed many authors have used it in problems such as determination of the large-scale structure of the Universe, galaxy or void counting, or analysis of the cosmic background radiation. In addition, we describe more in depth two applications of our own, namely, the detection of various magnetic features of the Sun, from satellite images, and the detection of distant gamma-ray sources in the Universe. In the latter case, difficult statistics problems arise, because of the extreme weakness of the signal (such a source emits very few high energy photons). The next topic is Earth physics: fault detection in geophysics, seismology, climatology (notably, thunderstorm prevision). A number of successful applications pertain to fluid dynamics, from the detection of coherent strucures in fully developed 2-D turbulence (a domain pioneered forcefully by Marie Farge [164]) to the measurement of the velocity field in a turbulent fluid, or the disentangling of a 2-D (or 3-D) wave train. Next comes the world of fractals. These are structures that are solely characterized by their behavior under a scaling transformation: ideal ground for wavelets! However, the self-reproducing properties of physical fractals are in general only approximate, so that methods from statistical mechanics are needed. Thus, a thermodynamical formalism has been designed by Arn´eodo and his group in Bordeaux for treating such problems, and we give a brief account of it. Finally we touch upon the problem of shape recognition, where wavelet descriptors have proven useful too. At this point, the book undergoes a sort of phase transition. Up to here, everything was done by hand, so to speak. The properties of the CWT have been derived by explicit calculations and very few mathematical prerequisites have been asked for. But now it is time to look over the hill and notice that the whole theory is firmly grounded in group theory. Indeed the wavelet transform and all its properties may be entirely derived from an appropriate representation of the affine group, both in one and in two dimensions. A mathematical condition, called square integrability of the representation, ensures the validity of the derivation, in particular the possibility of inverting the wavelet transform, that is, of obtaining reconstruction formulas. We devote two rather short chapters, 6 and 7 to these developments, with a double benefit. First, on the pedagogical level, we want to convince the reader that the group-theoretical approach is not only mathematically correct and pleasant, it is also natural and easy. It allows us indeed to understand in a simple and unified language the deeper mathematical structures involved. It is also quite efficient, in that it yields a general formalism (in fact, a special case of the coherent state formalism, well known in quantum physics, in particular, in quantum optics) that permits us to extend the CWT to more general manifolds, such as R3 , the two-sphere, or space–time, all generalizations that will be discussed in later chapters. Of course, we do not expect our reader to be fully conversant with group theory, and we will define all the needed ingredients along the way. Actually we will essentially restrict our treatment

xv

Prologue

to 2 × 2 or 3 × 3 matrices, without resort to abstract notions. Nevertheless, we found it convenient to gather all the group-theoretical information in a separate appendix. We begin, in Chapter 6, by revisiting the 1-D CWT in the light of the so-called ax + b or restricted affine group of the line, that is, the set of all translations and positive dilations. It turns out that the CWT may also be interpreted as a phase space representation of signals, in the sense of Hamiltonian mechanics, and the group-theoretical language makes this evident. The same treatment is then applied to the Gabor transform, also called Short Time or Windowed Fourier transform, simply replacing the affine group by the Weyl–Heisenberg group, that is, the group of phase space translations (this point of view has also been emphasized by Daubechies [Dau92]). Next, in Chapter 7, we repeat the procedure in two dimensions. Here the relevant group is the similitude group SIM(2), which consists of translations, rotations and dilations of the plane, that is, precisely all the transformations we have chosen to apply to images. Here, as in the 1-D case, the basic tool is a representation of the group by unitary operators acting in the space of finite energy signals, a natural representation that possesses the property of square integrability, meaning roughly that its matrix elements are square integrable functions of the group parameters. Here too, the CWT is a phase space realization of signals, and we spend some time exploring the consequences of this fact. In a third chapter with a mathematical flavor, Chapter 8, we discuss two less known properties of wavelets. First, some of them have minimal uncertainty, in the sense that they saturate some uncertainty relations linked to the Lie algebra of the wavelet group, exactly as Gaussians saturate those associated to the canonical commutation relations. Then we explore the relationship between wavelet transforms and the Wigner transform, well–known in physics and in radar theory (under the name of the closely related ambiguity function). The next two chapters are devoted to various extensions of the standard CWT, that can be derived with help of the general formalism just developed. First we treat, in Chapter 9, the higher dimensional cases. We begin with the 3-D CWT, which is a straightforward extension of the 2-D case. Then we examine in depth the CWT over the 2-sphere. Here, of course, there is a strong motivation from several domains, from geophysics to astrophysics. The former is clear. As for the latter, when one considers the whole Universe, as in the problem of gamma source detection mentioned above, it is necessary to take the curvature into account. However, there is an equally appealing aspect in the mathematics of the subject. Indeed, the group to consider here is the conformal group of the sphere S 2 , which is nothing but the proper Lorentz group SOo (3, 1). The same group is also the conformal group of the plane R2 , for instance, the tangent plane at the North Pole. The sphere and its tangent plane are mapped onto each other by the stereographic projection from the South Pole and its inverse. This operation is in fact the key to the construction of a spherical CWT. Indeed, the operations to be performed on spherical signals are motions on the sphere, given by rotations, and local dilations around a given point. In order to define

xvi

Prologue

these, one first defines dilations around the North Pole by lifting the corresponding ones in the tangent plane by inverse stereographic projection. Then, dilations around any other point of the sphere are obtained by combining the previous ones with an appropriate rotation. As a consequence, the parameter space of the spherical CWT is not the Lorentz group itself, but a homogeneous space of it, containing only rotations and the dilations just defined, that is, the quotient of SOo (3, 1) by a certain subgroup. Therefore, one needs the general formalism described in Chapter 7 in order to get a genuine spherical CWT. As an additional benefit, one recovers the natural link between the sphere and its tangent plane: the spherical CWT tends to the usual plane CWT when the radius of the sphere increases to infinity (the so-called Euclidean limit). It is gratifying that this aspect too is entirely described by the group-theoretical machinery, in terms of an operation called group contraction. Another byproduct of our spherical CWT is the possibility of designing good wavelet approximations of integrable functions on the sphere, another result previously known in the plane case. Here again practical applications are at hand, in the context of the so-called Geomathematics advertised by Freeden and his school [Fre97]. Then we turn, in Chapter 10, to the extension of the CWT to space–time. The problem of interest here is, of course, motion estimation, more precisely, detection, tracking, and identification of objects in (relative) motion. Examples include traffic monitoring, autonomous vehicle navigation, and tracking of ballistic missile warheads. This is a difficult problem, since the data is huge and often very noisy. As a consequence, most algorithms tend to lose track of the targets after a while, particularly if the latter changes its appearance (e.g., a maneuvering aeroplane) or in the case of an occlusion (one moving object hides another one). From the wavelet point of view, one designs a spatio-temporal CWT, whose parameters are space and time translations, rotations, global space–time dilations, that catch the size of the target, and a speed tuning parameter that measures its speed. The usual formalism goes through almost verbatim and allows one to design an efficient algorithm for motion estimation. One key ingredient again is the successive use of several partial energy densities. In the final Chapter 11, we turn to another kind of generalizations, namely, transforms specially adapted to the detection and modeling of lines and curves, called the ridgelet and the curvelet transforms. The motivation for these new transforms, and their superiority over standard wavelets, is that they take much better into account the geometry of the object to be analyzed. A curve in the plane is more 1-D than 2-D, and the conventional 2-D CWT simply ignores this fact – hence it is unnecessarily costly. Here, of course, one experiences the much bigger richness of the 2-D world, in particular, concerning singularities of functions. These transforms naturally lead to new approaches to image compression and various nonlinear approximations, that we also describe. We conclude the chapter and the book with a topic called ‘algebraic wavelets’. These are wavelets adapted to self-similar tilings on the line or the plane obtained by

xvii

Prologue

replacing the usual natural numbers by a different system of numeration, for instance, √ the golden mean τ = 12 (1 + 5). This is actually a generalization of the discrete WT, but it provides another example of wavelets adapted to a specific geometry, hence it is not out of place in this volume, and we found it interesting to give a short account of it, both in 1-D and in 2-D. In the latter case, typical examples are the famous Penrose tilings of the plane, with pentagonal symmetry, and this brings us back to the study of aperiodic patterns and to quasicrystals! The conclusion of the whole story is definitely optimistic. Wavelets, and in particular the continuous WT, have proven to be a versatile and extremely efficient tool for image processing, provided one uses the right wavelet on the right problem. Their future is undoubtly bright, in many fields of science and technology. Before concluding this introduction, several technical remarks are in order. First, most examples that are not reproduced from original papers have been computed using our own wavelet toolbox, called the YAW (Yet Another Wavelet) Toolbox, and freely accessible on the Louvain-la-Neuve website . Next, we have found it useful to split the references into two sections, devoted to books and Ph.D. theses, and regular journal articles (with a different presentation, viz. [Ald96] and [2], respectively). As we have already said, theses are an extremely rich source of information, although they are often only accessible on the web. In general, we have tried to trace most of the results to the original papers. Of course, there are omissions and misrepresentations, due to our ignorance and prejudices. We take responsibility for this and apologize in advance to those authors whose work we might have mistreated.

Acknowledgements The present volume results from some fifteen years of continuing research interest in wavelets in Louvain-la-Neuve, starting with a collaboration between Alex Grossmann (Marseille) and J.-P. A. Throughout these years, all four authors have lectured on wavelets in places as diverse as Louvain-la-Neuve, Paris, Atlanta, Zakopane, Havana, Cotonou, Amsterdam, Lyon, Brussels, many papers have been written, and the book reflects all the experience thus acquired. Many students have been involved and they all deserve thanks for their contribution to the edifice. But most of all, we have to express our gratitude to three people, Bernard Piette, Alain Coron, and Laurent Jacques, for the project could never have been completed without their computer skills and tireless help. As expected, the elaboration process involved many reciprocal visits, and we have to thank our respective institutions (UCL, CTSPS at Clark Atlanta U., EPFL, Concordia U.) for their hospitality and support, as well as the various funding agencies that made these travels possible. Finally, we thank many colleagues for stimulating

xviii

Prologue

discussions, such as Roberto Cesar, Alain Arn´eodo, Matthias Holschneider, Fran¸coise Bastin, Fabio Bagarello, Emmanuel Van Vyve, to name a few. Special thanks are due to Bruno Torr´esani, whose friendly comments and criticisms have helped us considerably, and who in addition proofread a large part of the manuscript. Jean-Pierre Antoine (Louvain-la-Neuve) Romain Murenzi (Atlanta and Kigali) Pierre Vandergheynst (Lausanne) Syed Twareque Ali (Montr´eal)

1

Warm-up: the 1-D continuous wavelet transform

1.1

What is wavelet analysis? Wavelet analysis is a particular time- or space-scale representation of signals that has found a wide range of applications in physics, signal processing and applied mathematics in the last few years. In order to get a feeling for it and to understand its success, we consider first the case of one-dimensional signals. Actually the discussion in this introductory chapter is mostly qualitative. All the mathematically relevant properties will be described precisely and proved systematically in the next chapter for the two-dimensional case, which is the proper subject of this book. It is a fact that most real life signals are nonstationary (that is, their statistical properties change with time) and they usually cover a wide range of frequencies. Many signals contain transient components, whose appearance and disappearance are physically very significant. Also, characteristic frequencies may drift in time (e.g., in geophysical time series – one calls them pseudo-frequencies). In addition, there is often a direct correlation between the characteristic frequency of a given segment of the signal and the time duration of that segment. Low frequency pieces tend to last for a long interval, whereas high frequencies occur in general for a short moment only. Human speech signals are typical in this respect: vowels have a relatively low mean frequency and last quite a long time, whereas consonants contain a wide spectrum, up to very high frequencies, especially in the attack, but they are very short. Clearly standard Fourier analysis is inadequate for treating such signals. Strictly speaking, it applies only to stationary signals, and it loses all information about the time localization of a given frequency component. In addition, it is very uneconomical. When the signal is almost flat, and thus uninteresting, one still has to sum an infinite alternating series to reproduce it. Worse yet, Fourier analysis is highly unstable with respect to perturbation, because of its global character. For instance, if one adds an extra term, with a very small amplitude, to a linear superposition of sine waves, the signal will barely be modified, but the Fourier spectrum will be completely perturbed. This does not happen if the signal is represented in terms of localized components. Indeed, as we shall see shortly, the basic idea of the wavelet transform is to decompose a signal locally

1

2

Warm-up: the 1-D continuous wavelet transform

into contributions living at different scales. This is a marked contrast with the Fourier components, which are sinusoidal waves repeating themselves indefinitely. As such, it is difficult to give them any physical reality. If a piece of audio signal is identically zero, it is because no sound is emitted, not because the Fourier components necessary to represent the zero signal interfere destructively. These components are a mathematical construction, rather than a genuine physical phenomenon. To quote J. Ville [364]: Si nous consid´erons en effet un morceau de musique . . . et qu’une note, la par exemple, figure une fois dans le morceau, l’analyse harmonique [de Fourier] nous pr´esentera la fr´equence correspondante avec une certaine amplitude et une certaine phase, sans localiser le la dans le temps. Or, il est e´ vident qu’au cours du morceau il est des instants o`u l’on n’entend pas le la. La repr´esentation est n´eanmoins math´ematiquement correcte, parce que les phases des notes voisines du la sont agenc´ees de mani`ere a` d´etruire cette note par interf´erence lorsqu’on ne l’entend pas et a` la renforcer, e´ galement par interf´erence, lorsqu’on l’entend; mais s’il y a dans cette conception une habilet´e qui honore l’analyse math´ematique, il ne faut pas se dissimuler qu’il y a e´ galement une d´efiguration de la r´ealit´e: en effet, quand on n’entend pas le la, la raison v´eritable est que le la n’est pas e´ mis.

That is, If we consider a piece of music . . . and if a note, an A for instance, appears once in that piece, Fourier analysis will yield the corresponding frequency with a certain amplitude and a certain phase, without localizing the A in time. Clearly the A will not be heard at certain instants. Yet the representation is mathematically correct, because the phases of the neighboring notes conspire to suppress the A by interference when it is not heard and to enhance it, again by interference, when it is heard. However, although this conception shows a skillfulness that honors mathematical analysis, one should not hide the fact that it also distorts reality: indeed, when the A is not heard, the true reason is that the A is not emitted.

Another eloquent comment along the same line by L. de Broglie may be found, together with the one above, in [Fla93; p.9]. Facing these problems, signal analysts turn to time–frequency representations. The idea is that one needs two parameters: one, called a, characterizes the frequency, the other one, b, indicates the position in the signal. This concept of a time–frequency representation is in fact quite old and familiar. The most obvious example is simply a musical score (see Figure 1.1). Clearly, it is not sufficient to give the pitch of a given note, that is, the frequency to which it corresponds, it is also important to know when to play it (time information)! Let s(x) be a finite energy signal, that is, a square integrable function s ∈ L 2 (R, d x). In most cases, x will be a time variable and the (Fourier) conjugate quantity a frequency, 2 4 Fig. 1.1. A traditional time–frequency representation of a signal (from Mozart’s Don Giovanni,

Act 1).

3

1.1 What is wavelet analysis?

but in general x simply represents position in the signal. Thus, following [Dau92], we prefer to keep a neutral notation (x, ξ ) for the couple of conjugate variables, instead of the more familiar (t, ω). Accordingly, the Fourier transform of the signal s is defined by ∞ 1 s(ξ ) = √ d x e−iξ x s(x). (1.1) 2π −∞ If one requires the transform to be linear, a general time–frequency transform of the signal s will take the form: ∞ d x ψb,a (x) s(x) , (1.2) s(x) → S(b, a) = −∞

where ψb,a is the analyzing function. Within this class, two time–frequency transforms stand out as particularly simple and efficient: the windowed (or short time) Fourier transform (WFT) and the wavelet transform (WT). For both of them, the analyzing function ψb,a is obtained by acting on a basic (or mother) function ψ, in particular, b is simply a time translation. The essential difference between the two is in the way the frequency parameter a is introduced: (1) Windowed Fourier transform: ψb,a (x) = ei(x−b)/a ψ(x − b).

(1.3)

Here ψ is a window function and the a-dependence is a modulation (1/a ∼ frequency); the window has constant width, but the smaller a, the larger the number of oscillations in the window (see Figure 1.2 (left)). (2) Wavelet transform: x −b 1 ψb,a (x) = √ ψ . (1.4) a a The action of a on the function ψ (which must be oscillating, see below) is a dilation (a > 1) or a contraction (a < 1): the shape of the function is unchanged, it is simply spread out or squeezed (see Figure 1.2 (right)). In particular, the effective support of ψb,a varies as a function of a. The windowed Fourier transform was originally introduced by Gabor (actually in a discretized version), with the window function ψ taken as a Gaussian; for this reason, it is sometimes called the Gabor transform. With this choice, the function ψb,a is simply a canonical (harmonic oscillator) coherent state [Kla85], as one sees immediately by writing 1/a = p. Since the new variables are the time (position) b and the frequency 1/a, the Gabor transform yields a genuine time–frequency representation of the signal. As for the wavelet transform, the variables are b and the scale a (or pitch in the case of music), hence we shall speak rather of a time-scale representation. We may remark here that the resemblance between the windowed Fourier transform and the wavelet transform is not accidental. They are both particular instances of a large

4

Warm-up: the 1-D continuous wavelet transform

1/a ~ ~ frequency ψb,a(x)

high

medium

low

a1

x Fig. 1.2. The function ψb,a (x) for different values of the scale parameter a, in the case of the windowed Fourier transform (left) and the wavelet transform (right). The quantity 1/a, which corresponds to a frequency, increases from bottom to top.

class of integral transforms constructed by the formalism of coherent states [Ali00]. This general analysis, however, has a more mathematical flavor and is not needed in a first approach, although it clarifies and unifies the picture considerably. Therefore, we postpone it to Chapter 6, since we want to emphasize first the practical aspects of the wavelet transform. One should note that the assumption of linearity is nontrivial, for there exists a whole class of quadratic or, more properly, sesquilinear time–frequency representations. The prototype is the so-called Wigner–Ville transform, introduced originally by E.P. Wigner [373] in quantum mechanics (in 1932!) and extended by J. Ville [364] to signal analysis: +∞ x x Ws (b, ξ ) = d x e−iξ x s(b − ) s(b + ), ξ = 1/a. (1.5) 2 2 −∞ Note that the signal s(x) is usually a real function, but, in quantum mechanics, s(x) represents a wave function, and is thus in general complex. This transform is entirely intrinsic to the signal, since it does not contain any extra function (wavelet, window)

5

1.2 The continuous wavelet transform

that inevitably influences the result. On the other hand, it is quadratic, which implies the appearance of interference terms whenever the signal is a superposition of two components. In order to minimize these as much as possible, one usually smoothes the Wigner–Ville transform with some function , thus obtaining a whole class of quadratic transforms, called Cohen’s class [109,Fla93], of the general form: Cs (b, ξ ) = db dξ (b − b , ξ − ξ ) Ws (b , ξ ). (1.6) R2

An example is the so-called smoothened pseudo-Wigner–Ville distribution, +∞ +∞

SPWs (b, ξ ) = db g(b − b ) d x h(x) e−iξ x s(b − x/2) s(b + x/2), (1.7) −∞

−∞

corresponding to a factorizable kernel (b, ξ ) = (2π )−1/2 g(b) h(ξ ), where h denotes the Fourier transform of h. Further information about quadratic transforms may be found in [Fla93], and as a general survey for time–frequency methods, we refer to [Gro01].

1.2

The continuous wavelet transform Actually one should distinguish two different versions of the wavelet transform, the continuous WT (CWT) and the discrete (or more properly, discrete time) WT (DWT) [Dau92,Hol95]. The CWT plays the same rˆole as the Fourier transform and is mostly used for analysis and feature detection in signals, whereas the DWT is the analog of the Discrete Fourier Transform (see for instance [Bur98] or [326]) and is more appropriate for data compression and signal reconstruction. The situation may be caricatured by saying that the CWT is more natural to the physicist, while the DWT is more congenial to the signal analyst and the numericist. The continuous wavelet transform is the main topic of this book. Nevertheless, for the sake of comparison, we will give short overviews of the discrete WT, both in one and two dimensions. The two versions of the WT are based on the same transformation formula, which reads, from (1.2) and (1.4): ∞ x −b −1/2 S(b, a) = |a| s(x), (1.8) dx ψ a −∞ where a = 0 is a scale parameter and b ∈ R a translation parameter (one often imposes only a > 0, which is more natural, but makes formulas slightly more complicated; see Chapter 6). Equivalently, in terms of Fourier transforms: ∞ 1/2 ) S(b, a) = |a| s(ξ ) eiξ b . dξ ψ(aξ (1.9) −∞

6

Warm-up: the 1-D continuous wavelet transform

In these relations, s is a square integrable function, representing a finite energy signal, and the function ψ, the analyzing wavelet, is assumed to be well localized both in the space (or time) domain and in the frequency domain. In addition ψ must satisfy the following admissibility condition, which guarantees the invertibility of the WT: ∞ )|2 |ψ(ξ cψ ≡ 2π dξ < ∞. (1.10) |ξ | −∞ In most cases, this condition may be reduced to the (only slightly weaker) requirement that ψ has zero mean: ∞ = 0 ⇐⇒ d x ψ(x) = 0. (1.11) ψ(0) −∞

Intuitively, it expresses the fact that a wavelet must be an oscillating function, real or complex (“little wave”). This is often thought to be the origin of the term “wavelet”, but it is not the case historically. Indeed the word was widely in use in the geophysics community, with quite a different meaning, when it was introduced by Grossmann and Morlet [205,206] in the present sense, under the name “wavelets of constant shape” – but, of course, this lengthy nomenclature did not survive the very first founding paper! ) is real and The wavelet ψ is said to be progressive if its Fourier transform ψ(ξ vanishes identically for ξ 0. (In the signal processing community, a signal with this property is called analytic, following the terminology introduced by J. Ville [364].) In addition, ψ is often required to have a certain number of vanishing moments: ∞ d x x n ψ(x) = 0, n = 0, 1, . . . N . (1.12) −∞

This property improves the efficiency of ψ at detecting singularities in the signal, since it is then blind to polynomials up to order N , which constitute the smoothest part of the signal. Notice that, instead of (1.8), which defines the WT as the scalar product of the signal s with the transformed wavelet ψb,a , S(b, a) may also be seen as the convolution of s with the scaled, flipped and conjugated wavelet ψa# (x) = |a|−1/2 ψ(−x/a) : ∞ S(b, a) = (ψa# ∗ s)(b) = d x ψa# (b − x) s(x). (1.13) −∞

In other words, the CWT acts as a filter with a function of zero mean. This property is crucial, for the main virtues of the CWT follow from it, combined are as well with the support properties of ψ. Indeed, we must assume that ψ and ψ localized as possible, but respecting, of course, the Fourier uncertainty principle. This means that, up to minute corrections, the product of the lengths of the supports of ψ is bounded from below by a fixed constant, usually taken as 1/2. Equivalently, and ψ 2 is bounded from below. the product of the variances of the distributions |ψ|2 and |ψ| More precisely, one defines the centers of gravity (which may in fact be normalized to

7

1.2 The continuous wavelet transform

zero by a suitable redefinition of the coordinates): ∞ ∞ )|2 , x0 = d x x |ψ(x)|2 , ξ0 = dξ ξ |ψ(ξ −∞

(1.14)

−∞

and the corresponding variances ∞ ( x)2 = ψ−2 d x (x − x0 )2 |ψ(x)|2 ; −∞ ∞ 2 −2 )|2 . dξ (ξ − ξ0 )2 |ψ(ξ ( ξ ) = ψ

(1.15) (1.16)

−∞

Then the Fourier uncertainty theorem [Fla93] says that

x ξ

1 . 2

(1.17)

Under these assumptions, the transformed wavelets ψb,a and ψ b,a are also well localized. Therefore, the WT s → S performs a local filtering, both in time (b) and in scale (a). The transform S(b, a) is nonnegligible only when the wavelet ψb,a matches the signal, that is, the WT selects the part of the signal, if any, that lives around the time b and the scale a. has a numerical support (bandwidth) of width ξ , then ψ In addition, if ψ b,a has a numerical support of width ξ/|a|. Thus, remembering that 1/a behaves like a frequency, we conclude that the WT works at constant relative bandwidth, that is,

ξ/ξ = constant. This implies that it is very efficient at high frequency, i.e., small scales, in particular for the detection of singularities in the signal. By comparison, in the case of the Gabor transform, the support of ψ b,a keeps the same width ξ for all a, that is, the WFT works at constant bandwidth, ξ = constant. This difference in behavior is often the key factor in deciding whether one should choose the WFT or the WT in a given physical problem. Another crucial fact is that the transformation s(x) → S(b, a) may be inverted exactly, which yields a reconstruction formula (this is only the simplest one, others are possible, for instance using different wavelets for the decomposition and the reconstruction): ∞ ∞ da −1 s(x) = cψ db ψb,a (x) S(b, a), (1.18) 2 −∞ −∞ a where the normalization constant cψ is given in (1.10) (incidentally, this relation shows why the admissibility condition cψ < ∞ is required for the transformation to be invertible). This means that the WT provides a decomposition of the signal as a linear superposition of the wavelets ψb,a with coefficients S(b, a). Notice that the natural measure on the parameter space (a, b) is da db/a 2 , and it is invariant not only under time translation, but also under dilation. This fact is important, for it suggests that these geometric transformations play an essential rˆole in the CWT.

8

Warm-up: the 1-D continuous wavelet transform

One should emphasize here that the choice of the normalization factor |a|−1/2 in (1.4) or (1.8) is not essential. This choice makes the transform unitary: ψb,a 2 = ψ2 and also S2 = s2 , where · 2 denotes the L 2 norm in the appropriate variables (the squared norm is interpreted as the total energy of the signal). In practice, one often uses instead a factor a −1 , which has the advantage of giving more weight to the small scales, i.e., the high frequency part (which contains the singularities of the signal, if any). Thus, defining 1 x −b ψ(b,a) = ψ , (1.19) |a| a we obtain the so-called L 1 -normalized transform: ∞ x −b ˘ a) = ψ(b,a) |s ≡ |a|−1 S(b, s(x), dx ψ a −∞

(1.20)

which preserves the L 1 -norm of the signal, as follows immediately from the corresponding convolution formula ˘ a) = (ψa# ∗ s)(b), S(b,

(1.21)

˘ 1 = s1 , where where ψa# (x) = |a|−1 ψ(−x/a). Thus indeed ψa# 1 = ψ1 and S 1 · 1 denotes the L -norm in the corresponding variables.

1.2.1

Examples In order to fix ideas, we exhibit here two simple examples of wavelets, both in the time domain and in the frequency domain. (1) The Mexican hat wavelet This wavelet is simply the second derivative of a Gaussian: H (ξ ) = ξ 2 exp(− 1 ξ 2 ). ψH (x) = (1 − x 2 ) exp(− 12 x 2 ), ψ 2

(1.22)

(2) The Morlet wavelet This wavelet is essentially a plane wave within a Gaussian window: M (ξ ) = exp(− 1 (ξ − ξo )2 ) + c(ξ ). ψM (x) = exp(iko x) exp(− 12 x 2 ) + c(x), ψ 2 (1.23) Here the correction term c must be added in order to satisfy the admissibility condition (1.11), but in practice one will arrange that this term be numerically negligible ( 10−4 ) and thus can be omitted (it suffices to choose the basic frequency |ξo | large enough, typically |ξo | > 5.5). These two wavelets have very different properties and, naturally, they will be used in quite different situations. Typically, the Mexican hat is sensitive to singularities in the signal, and it yields a genuine time-scale analysis. On the other hand, since it is complex, the Morlet wavelet will catch the phase of the signal, hence will be sensitive to frequencies, and will lead to a time-frequency analysis, somewhat closer to a Gabor

1.2 The continuous wavelet transform

1

1.5

Log Scale

9

2

2.5

3

100

200

300

400

500

600

700

800

900

1000

Position Fig. 1.3. Wavelet analysis with a Mexican hat wavelet of the discontinuous signal bumps (shown in

the bottom panel).

analysis. In both cases, additional flexibility is obtained by adding a width parameter to the Gaussian (see (3.8) in the equivalent 2-D situation). As an illustration of the performance of the CWT as a singularity scanner, we first show in Figure 1.3 the analysis with a Mexican hat wavelet of a discontinuous signal, called bumps and consisting of three pieces, a δ function, a boxcar function and a tent function. Clearly the wavelet locates all discontinuities in the signal and in its successive derivatives well. However, if one wants to discriminate between the various types of singularities, one has to invoke the concept of vanishing moment, defined in (1.12). Let us consider the successive derivatives of a Gaussian: ψH(n) (x) = −

dn exp(− 12 x 2 ). dxn

(1.24)

For increasing n, these wavelets have more and more vanishing moments, and are thus sensitive to increasingly sharper details. As an example, we consider a continuous signal obtained by glueing together an arc of parabola (the so-called function x+2 ) and a linear piece and we analyze it successively with the first three derivatives of a Gaussian, ψH(n) (x), n = 1, 2, 3. The result is shown in Figure 1.4. In (a), the first-order wavelet ψH(1) has only one vanishing moment, hence it sees the full content of the two pieces of the signal. In (b), the second-order wavelet ψH(2) (x) ≡ ψH does not see the linear part anymore, only the singularities at the two ends, but still sees the quadratic piece on the left (in technical terms, one would say that this wavelet is blind to a linear trend in the signal). In (c), finally, the third-order wavelet correctly erases both pieces of the

10

Warm-up: the 1-D continuous wavelet transform

3.6

Log Scale

3.8 4 4.2 4.4 4.6 4.8

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

3.6

3.6

3.8

3.8

4

4

Log Scale

Log Scale

(a)

4.2

4.2

4.4

4.4

4.6

4.6

4.8

4.8

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

(b)

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

(c)

Fig. 1.4. Analysis of a composite signal (bottom panel) with successive derivatives of a Gaussian.

(a) First order; (b) second order; (c) third order.

signal, keeping only the three singularities. This example shows the advantage of the local filtering effect of the CWT. Notice that a Gabor analysis would be utterly unable to achieve such a discrimination between singularities, let alone to detect them! As a direct application of this behavior, an interesting technique has been designed by A.Arn´eodo et al. [49], which consists in analyzing the same signal with several wavelets ψH(n) , for different n. The features common to all the transforms surely belong to the signal, they are not artifacts of the analysis.

1.3

Discretization of the CWT, frames All this concerns the continuous WT (CWT). But, in practice, for numerical purposes, the transform must be discretized, by restricting the parameters a and b in (1.8) to the points of a discrete lattice = {a j , bk , j, k ∈ Z} in the (a, b)-(half)-plane. Then we

11

1.3 Discretization of the CWT, frames

say that yields a good discretization if an arbitrary signal s(x) may be represented as a discrete superposition jk (x), s(x) = ψ jk |s ψ (1.25) j,k ∈ Z

jk should be instead of the reconstruction formula (1.18). In (1.25), ψ jk ≡ ψbk ,a j and ψ explicitly constructible from ψ jk . We emphasize that (1.25) must be an exact representation, i.e., there is no loss of information as compared to a direct discretization of the continuous reconstruction (1.18). Notice that here also, as in the latter, the reconstruction formula is in general not unique, which offers an additional degree of freedom in a given situation. One may wonder whether a discrete representation of the type (1.25) is really possible. The answer lies in the reproducing property, da db −1

S(b , a ) = cψ ψb ,a |ψb,a S(b, a) , (1.26) a2 that every wavelet transform must satisfy. Indeed (1.26) implies that the information content of the wavelet transform S(b, a) is highly redundant. In fact the signal has been unfolded from one to two dimensions, and this explains the practical efficiency of the CWT for disentangling parts of the signal that live at the same time, but on different scales. This redundancy (which is the source of the nonuniqueness of the reconstruction formula) may be eliminated – this is the rationale behind the discrete wavelet transform. It may also be exploited, for instance, by observing that it must be possible to obtain the full information about the signal from a small subset of the values of the transform S(b, a). In particular, the validity of a representation (1.25) means that a discrete subset will do the job, and this is precisely what is needed for the reconstruction of a signal from its wavelet transform. The problem is to find the minimal sampling grid ensuring no loss of information. In order to formulate it in mathematical terms, one relies on the theory of discrete frames or nonorthogonal expansions, that we now sketch. See [121,Dau92] for a complete treatment. In fact, the discrete representation (1.25) means that the signal s(x) may be replaced by the set {ψ jk |s} of its wavelet coefficients. Since s ∈ L 2 , it is natural to require that the sequence of coefficients be also square integrable and that the map F : s → {ψ jk |s} be continuous from L 2 (R) to 2 , i.e., |ψ jk |s|2 Bs2 , 0 < B < ∞. (1.27) j,k ∈ Z

In addition, one wants the reconstruction of s(x) from its coefficients to be numerically stable, that is, a small error in the coefficients implies a small error in the reconstructed signal. In particular, if the left-hand side of (1.27) is small, s2 should be small also. Therefore, there must exist a constant A > 0 such that

12

Warm-up: the 1-D continuous wavelet transform

A s2

|ψ jk |s|2 B s2

(1.28)

j,k ∈ Z

(the lower bound indeed guarantees the numerical stability [Dau92]). By definition, this relation means precisely that the set {ψ jk } constitutes a (discrete) frame, with frame bounds A and B. This frame is said to be tight if A = B. Note that (1.28) is in fact a weakened form of the Parseval relation. The latter is recovered in the case of a tight frame, in particular for an orthonormal basis. For a general frame, however, (1.28) is sufficient for inverting the wavelet representation (we will discuss this in detail in the 2-D case, in Section 2.4). We will present a detailed analysis of these concepts in Chapter 2. Here we simply observe that, for all practical purposes, a good frame is almost as good as an orthonormal basis. By “good frame,” we mean that the expansion (1.25) converges sufficiently fast. The detailed analysis of [121,122] shows this to be the case if |B/A − 1| 1, thus in particular if the frame is tight. Many functions ψ satisfying the admissibility condition (1.10) will yield a good frame (of course, this must be proved for every given ψ). However, we will not get an orthonormal basis, since the functions {ψ jk , j, k ∈ Z} are in general not orthogonal to each other! Yet orthonormal bases of wavelets can be constructed, but by a totally different approach, based on the concept of multiresolution analysis. We emphasize that the discretized version of the CWT just described is totally different in spirit and method from the genuine discrete wavelet transform, that we will sketch in Section 1.5 below. The full story may be found in [Dau92], for instance. Of course the practical question is: how does one build a good frame? Clearly, the question of the existence of a discrete frame must take into account the geometry of the parameter space. In the present case, this means that the lattice must be invariant under discrete dilations and translations: r for scale, one chooses naturally a = a λ− j , j ∈ Z, for some λ > 1; j o −j r for time, one takes b ≡ b = k b a λ , j, k ∈ Z. k k, j o o Thus we get ψ jk (x) = λ j/2 ψ(ao−1 λ j x − kbo ),

j, k ∈ Z.

(1.29)

The most common choice is λ = 2 (octaves!) and ao = bo = 1, which results in ψ jk (x) = 2 j/2 ψ(2 j x − k),

j, k ∈ Z.

(1.30)

It is worth noticing that this so-called dyadic lattice {(k2− j , 2− j ), j, k ∈ Z} is exactly the same that indexes the DWT (see Section 1.5), which may create some confusion (and sometimes did so!). For a given choice of ψ, λ, one finds a range of values of bo such that {ψ jk }, as given in (1.29), is a frame. Detailed results may be found in [122,Dau92]. Here we will restrict ourselves to the following simplified version.

13

1.3 Discretization of the CWT, frames

Theorem 1.3.1 . Let ψ and ao be such that: ∞ − j ξ )|2 > 0; (i) inf |ψ(λ 1ξ λ

(ii)

j=−∞

)| C |ξ |α (1 + |ξ |)−γ , α > 0, γ > α + 1. |ψ(ξ

Then there exists boo such that {ψ jk } constitutes a frame for all choices bo < boo . Both the Mexican hat and the Morlet wavelet satisfy the conditions of the theorem for the dyadic case, λ = 2, ao = bo = 1, thus they both generate discrete frames on the dyadic lattice. Explicit values for the corresponding frame bounds A, B may be found in [Dau92]. However, the numerical implementation of such a dyadic frame is unwieldy, in particular, the reconstruction works well only if the frame is very redundant. Two variants are used in practice, and both of them amount to increase the redundancy. The first one consists in subdividing further the octaves, introducing what are called additional voices. This means that one further subdivides each octave, replacing in (2.77) the exponent j by Nν j, ν = 0, 1, . . . , N − 1 (ν is called a voice and N the number of voices). The effect is to “densify” the dyadic lattice, which improves the ratio B/A and thus speeds up convergence of the discrete approximation. By taking sufficiently many voices by octaves, one may even get frames which are numerically tight. For further details, we refer the reader to [Dau92]. However, if the speed is the determining criterion, one can do better by using the continuous wavelet packets described in Section 1.6.1. The other solution consists in replacing the dyadic lattice with a rectangular one, that is, taking the sampling rate independent of the scale. This has the advantage of making the whole analysis invariant under global discrete translations. The resulting version of the discretized CWT has been advocated by Mallat [Mal99,265] (who, curiously, calls it the dyadic WT!) and Torr´esani [Tor95], and we will meet it again repeatedly in the sequel (see Section 1.6.1). Note that similar ideas were already used in the classical Littlewood–Paley analysis [Fra91] and the famous Laplacian pyramid from vision analysis [91]. The scheme was also rediscovered by statisticians, who called it “stationary wavelet transform” [112,295]. To illustrate this technique, we show in Figure 1.5 a five-level decomposition of the signal bumps with a translation invariant frame of quadratic spline wavelets, as used in [265]. The figure shows, from bottom up, the low resolution approximation and the five levels of details with increasing resolution. As we shall see in detail in the next chapter, Section 2.5, the whole machinery of frames extends almost verbatim to the two-dimensional case. In particular, both the 2-D Mexican hat and the 2-D Morlet wavelet yield reasonably good 2-D frames. Besides their simplicity and their efficiency, this explains the widespread use of these two wavelets in image processing, as was already the case in one dimension.

14

Warm-up: the 1-D continuous wavelet transform

Signal 2 1 0 2 0 −2 1 0 −1 1 0 −1 2 0 −2 2 0 −2 10 5 0

100

200

300

400

500

600

700

800

900

1000

Fig. 1.5. Five-level decomposition of the signal bumps with a translation invariant frame of quadratic spline wavelets. The low resolution approximation is shown on the bottom panel and the five levels of details with increasing resolution a = 2− j , j = −5, −4, −3, −2, −1, in the next five panels.

1.4

Ridges and skeleton Real life signals are frequently very entangled and noisy, and their WT is difficult to interpret. However, a clever exploitation of the intrinsic redundancy of the CWT is often able to bypass the difficulty and thus to improve the efficiency and the range of applicability of wavelet analysis. The technique consists in using the skeleton of the CWT instead of its modulus. Roughly speaking, the skeleton is a collection of lines, called ridges, which are approximately lines of local maxima. The concept is easy to visualize in two extreme situations [Mal99,262,358]. Assume first that the signal s(t) consists of a singularity γα (x − xo ), of order α, at time xo , superimposed on a smooth background and some stochastic noise [262]. Here the singularity function γα is defined as follows: γα (x − xo ) =

0, x xo , α (x − xo ) , x > xo .

(1.31)

15

1.4 Ridges and skeleton

[The index α is in fact an index of homogeneity or a Lipschitz regularity exponent. For instance, a δ function has α = −1.] Thus we have d α+1 γα (x − xo ) = (α + 1) δ(x − xo ). d x α+1 Let the wavelet be the nth derivative of a smooth positive function φ, that is, ψ(x) = dn φ(x), with n α + 1 (typically, a derivative of a Gaussian, like the Mexican hat dxn and its higher order analogs (1.24)). Then the CWT of γα with respect to ψ to reads: n−α−1 φ xo − b α d . (1.32) Sγα (b, a) = (α + 1) a d x n−α−1 a Assume now that the modulus of the (n − α − 1)th derivative of φ has N maxima {φl , l = 1, . . . , N } at positions {xl , l = 1, . . . , N }. Then, for each a, the modulus |Sγα (b, a)| has N maxima localized at positions {bl = axl + xo , l = 1, . . . , N }, which converge toward xo as a → 0. Furthermore, the maxima of |Sγα (b, a)| lie on N lines, called (vertical) ridges {b = axl + xo , l = 1, . . . , N }, which converge toward the singularity xo of the signal, and the modulus of |Sγα (b, a)| along the lth ridge behaves as aα : |Sγα (b = axl + xo , a)| = (α + 1) a α φl .

(1.33)

Hence the strength α of the singularity may be read off a log–log plot: ln |Sγα (axl + xo , a)| ∼ α ln a + ln φl .

(1.34)

This technique, introduced by Mallat and Hwang [262], has been developed to a considerable extent for the analysis of fractals by Arn´eodo and his collaborators under the name of Wavelet Transform Modulus Maxima (WTMM) (see [Arn95] for a survey). The important point is that the restriction of the WT to its skeleton (the set of its ridges) characterizes the signal completely [264,265]. Thus, in practice, it is enough to compute the skeleton. We will discuss the 2-D extension of the WTMM method in Section 2.3.5 and its application to fractals in Section 5.4. To give a simple example, we take again our signal bumps, and compute the skeleton of its wavelet transform shown in Figure 1.3. The result, presented in Figure 1.6, clearly confirms the analysis above. As a second example, we consider the analysis of the behavior of a material under impact made in [358]. The physical context is that of a so-called “instrumented falling weight impact” testing. During such a test, a striker falls from a certain height on a clamped disk, so that either the striker rebounds or the disk breaks. In both cases, one records the time and the force on the striker. This type of event occurs on a very short time scale and is thus essentially transient, so that a time–frequency method is required for the analysis. Among the various methods discussed in [358], we focus here on the case of a rebound and a wavelet analysis of the force signal. The Mexican hat detects precisely three discontinuity points, namely, first contact, maximal penetration and last

Warm-up: the 1-D continuous wavelet transform

1

1.5

Log Scale

16

2

2.5

3

100

200

300

400

500

600

700

800

900

1000

Position Fig. 1.6. Skeleton of the CWT of the signal bumps, as presented in Figure 1.3.

contact. The results are shown in Figure 1.7. The signal is given in panel (a), whereas the CWT and its skeleton are presented in (b) and (c). The latter, in particular, shows the three ridges that point towards the three instants mentioned. A further analysis exploits the behavior of the modulus of the CWT along each ridge. This yields precious insight into the physics of the phenomenon, particularly in the case of the rupture of the sample, not shown here. We refer to [358] for more details. More generally, vertical ridges allow us to discriminate between genuine signal features and noise. First, noise ridges are usually much shorter, being visible mostly at small scales [21]. Then the modulus of the CWT tends to increase for increasing scale a on a noise ridge, whereas it decreases along genuine signal ridges. This fact has been exploited, for instance (in a discrete set-up), for the correction of aberrated images of the Hubble Space Telescope [85]: noise and signal have opposite behavior with increasing scale. As for the second typical situation, the idea is that many signals are well approximated by a superposition of simple spectral lines: s(x) =

N

sl (x),

sl (x) = Al (x) eiξl x ,

(1.35)

l=1

where the amplitude Al (x) varies slowly. By linearity, the WT of this signal is a sum

of terms, S(b, a) = l Sl (b, a), where, from (1.9),

1.4 Ridges and skeleton

250 (a)

Force [Volts]

200

150

100

50

500

600

700

800

900 1000 1100 1200 1300 1400 1500

Time [10−4 s]

(b) 1.86

Scale

3.59

6.90

13.29

25.58

49.25

600

700

800

900

1000 1100 1200 1300 1400 1500 −4

Time [10

s]

(c) 1.86

3.59

Scale

17

6.90

13.29

25.58

49.25

600

700

800

900 1000 1100 1200 1300 1400 1500 −4

Time [10

s]

Fig. 1.7. Analysis of the rebound signal, with a Mexican hat wavelet: (a) the signal and the points

detected by the respective ridges; (b) the modulus of the CWT; and (c) the corresponding skeleton (from [358]).

18

Warm-up: the 1-D continuous wavelet transform

Sl (b, a) =

√ a

+∞

−∞

√ = a eiξl b

l (ξ − ξl ) eiξ b )A dξ ψ(aξ

+∞

−∞

l (ξ ) eiξ b . dξ ψ(a(ξ + ξl )) A

(1.36) (1.37)

around aξl , one obtains the following expansion Inserting the Taylor expansion of ψ for Sl : l )sl (b) − iaeiξl b Sl (b, a) = ψ(aξ

d Al dψ (aξl ) (b) + higher order terms dξ db

(1.38)

Assuming the amplitude to be a smooth function (C 1 ), all terms beyond the first are easily bound using the intermediate value theorem and one can limit oneself to the lowest order, namely, l )sl (b), Sl (b, a) ψ(aξ

(1.39)

and thus S(b, a)

N

l )sl (b). ψ(aξ

(1.40)

l=1

) has a unique maximum in frequency space at ξ = ξo , like Assume that the wavelet ψ(ξ the Morlet wavelet (1.23). Then, if the values of the frequencies ξl are sufficiently far l ) allows to treat each spectral line independently. away from each other, the factor ψ(aξ In this case, the contribution of the lth spectral line to S(b, a) is localized on the scale al = ξo /ξl and, along the line of maxima a = al , called the lth (horizontal) ridge, the CWT is approximately proportional to the lth spectral line: S(b, al ) sl (b). o) ψ(ξ

(1.41)

The set of all the ridges is again called the skeleton of the CWT. Thus the restriction of the WT S(b, a) to its skeleton contains the whole information. The analysis extends to the more general case where the spectral lines have the form sl (x) = Al (x) eiφl (x) ,

(1.42)

with the amplitude Al (x) varying slowly with respect to the phase φl (x) (such signals are called asymptotic). Typical examples are spectra in NMR spectroscopy [210]. In this case, each term Sl in the CWT (1.8) in the time domain is a rapidly oscillating integral, the essential contribution to which is given by the stationary points of the phase of the integrand. These points are the solutions xs (b, a) of the equation dφl ξo (xs ) = . (1.43) dx a Then the corresponding ridge of the WT is defined as the set of points (b, a) for which xs (b, a) = b. These constitute a curve in the (b, a)-half-plane, which essentially

19

1.5 The discrete WT: orthonormal bases of wavelets

reduces to a line of local maxima. A detailed analysis [131] shows again that, on this curve, the WT S(b, a) coincides, up to a small correction, with the component sl (b) of the signal. Taking all ridges together, one obtains the skeleton of the CWT, and the analysis shows that the restriction of S(b, a) to it essentially coincides with the analytic signal Z (b) associated to s(x). It follows again that the restriction of the WT S(b, a) to its skeleton contains the whole information. In particular, the so-called frequency modulation law x −1 arg{s(x)} of s(x) is easily recovered from it. Thus, it is not necessary to compute the whole CWT, but only its skeleton [99]. This is, of course, much less costly computationally, because there are fast algorithms available. Spectacular applications of this method may be found, for instance, in spectroscopy [131], geomagnetism [4], chirp detection/estimation in gravitational waves [Mor02,277] or shape determination [21]. The last quoted paper, in particular, contains a thorough analysis of the two types of ridges, vertical and horizontal, interpreted in both cases as lines of local maxima (see Section 5.4.2).

1.5

The discrete WT: orthonormal bases of wavelets One of the successes of the WT was the discovery that it is possible to construct functions ψ for which {ψ jk , j, k ∈ Z} is indeed an orthonormal basis of L 2 (R). In addition, such a basis still has the good properties of wavelets, including space and frequency localization. Moreover, it yields fast algorithms, and this is the key to the usefulness of wavelets in many applications. The construction is based on two facts. First, almost all examples of orthonormal bases of wavelets can be derived from a multiresolution analysis, and then the whole construction may be transcripted into the language of digital filters, familiar in the signal processing literature. Notice that it is precisely at this point that arises the basic difference between the discretized continuous wavelet transform, discussed in the previous section, and the discrete wavelet transform (DWT). In the former case, the wavelet ψ is chosen a priori (with very few constraints, as we have seen above), and the question is whether one can find a lattice such that {ψ jk } is a frame with decent frame bounds A, B. In the other approach, one imposes from the beginning that the set {ψ jk } be an orthonormal basis and tries to construct a function ψ to that effect. The construction is rather indirect and the resulting function is usually very complicated (sometimes it has a fractal behavior). Definition 1.5.1 . A multiresolution analysis of L 2 (R) is an increasing sequence of closed subspaces . . . ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . , with j ∈ Z V j = {0} and j ∈ Z V j dense in L 2 (R), and such that

(1.44)

20

Warm-up: the 1-D continuous wavelet transform

(1) f (x) ∈ V j ⇔ f (2x) ∈ V j+1 (2) There exists a function φ ∈ V0 , called a scaling function, such that the family {φ(x − k), k ∈ Z} is an orthonormal basis of V0 . Combining conditions (1) and (2), one gets an orthonormal basis of V j , namely {φ jk (x) ≡ 2 j/2 φ(2 j x − k), k ∈ Z}. Note that the scaling function φ is often required only to generate a Riesz basis of V0 , that is, a frame {φ0k , k ∈ Z} of linearly independent vectors spanning V0 . However, since the Riesz basis can be orthonormalized, for instance by a Gram–Schmidt procedure, the condition (2) is equivalent, and simpler. Remark: Some authors (for instance, [Dau92] or [Mal99]) use the opposite convention for the index j, namely V j ⊂ V j−1 (both have advantages and inconveniences). With the present convention, large j means small scale (of order) 2− j or high frequency 2 j , thus high resolution 2 j . Each subspace V j can be interpreted as an approximation space. The approximation of f ∈ L 2 (R) at the resolution 2 j is defined by its projection onto V j , and the larger j, the finer the resolution obtained. Then condition (1) means that no scale is privileged. The additional details needed for increasing the resolution from 2 j to 2 j+1 are given by the projection of f onto the orthogonal complement W j of V j in V j+1 : V j ⊕ W j = V j+1 ,

(1.45)

and we have:

Wj. L 2 (R) =

(1.46)

j∈ Z

Equivalently, fixing some lowest resolution level jo , one may write ∞

L 2 (R) = V jo ⊕ Wj .

(1.47)

j= jo

The crucial theorem then asserts the existence of a function ψ, sometimes called the mother wavelet, explicitly computable from φ, such that {ψ jk (x) ≡ 2 j/2 ψ(2 j x − k), j, k ∈ Z} constitutes an orthonormal basis of L 2 (R): these are the orthonormal wavelets. The construction of ψ proceeds roughly as follows. First, the inclusion V0 ⊂ V1 yields the relation (called the scaling, or two-scale, or refinement equation): φ(x) =

∞ √ 2 h k φ(2x − k),

h k = φ1,k |φ.

(1.48)

k=−∞

Taking Fourier transforms, this gives ∞ ) = h(ξ ) φ(ξ ), with h(ξ ) = √1 h k e−ikξ . φ(2ξ 2 k=−∞

(1.49)

21

1.5 The discrete WT: orthonormal bases of wavelets

Thus h is a 2π-periodic function and it satisfies the relation |h(ξ )|2 + |h(ξ + π)|2 = 1,

a.e.

(1.50)

Iterating (1.49), one gets the scaling function as the infinite product ) = (2π)−1/2 φ(ξ

∞

h(2− j ξ ),

(1.51)

j=1

which may be proven to be convergent [Dau92]. Then one defines the function ψ ∈ W0 ⊂ V1 by the relation ) = g(ξ ) φ(ξ ), ψ(2ξ

(1.52)

where g is another 2π -periodic function. By the relation (1.45) and the orthonormality of the functions {φ jk }, the functions h, g must satisfy the identity g(ξ ) h(ξ ) + g(ξ + π) h(ξ + π) = 0,

a.e.

(1.53)

The simplest solution is to put g(ξ ) = eiξ h(ξ + π), which implies, in particular, |h(ξ )|2 + |g(ξ )|2 = 1, a.e. Then one obtains ψ(x) =

∞ √ 2 (−1)k−1 h −k−1 φ(2x − k),

(1.54)

k=−∞

and one proves that this function indeed generates an orthonormal basis with all the required properties. Another, equivalent, solution is ψ(x) =

√

2

∞

(−1)k h −k+1 φ(2x − k).

(1.55)

k=−∞

Various additional conditions may then be imposed on the basic wavelet ψ, such as arbitrary regularity, several vanishing moments (in any case, ψ has always mean zero), symmetry, fast decrease at infinity, and even compact support [Dau92]. Remark: Some authors (for instance, [Dau92] or [Tor95]) denote the functions h, g by m 0 , m 1 , respectively. The simplest example of this construction is the Haar basis, which comes from the scaling function φ(x) = 1 for 0 x < 1, and 0 otherwise √ (boxcar function). The coefficients of the corresponding filter h are h 0 = h 1 = 1/ 2, h k = 0, for k = 0, 1. Applying the recipe (1.55) then yields the Haar wavelet 1, if 0 x < 1/2, (1.56) ψHaar (x) = −1, if 1/2 x < 1, 0, otherwise.

22

Warm-up: the 1-D continuous wavelet transform

Similarly, various B-spline bases may be obtained along the same line. Other explicit examples may be found in [Chu92] or [Dau92]. In order to set up the discrete WT, the technique consists in translating the multiresolution structure into the language of digital filters, which is precisely what we have just done. Indeed, a filter is simply a multiplication operator in frequency space or a linear convolution in the time variable, and the discussion above amounts to nothing more than expanding (filter) functions in a Fourier series. For instance, h(ξ ) is a filter, with Fourier coefficients h n , g(ξ ) is another one, and {h, g} are called Quadrature Mirror Filters or QMFs whenever they satisfy the identities (1.50) and (1.53). Then the various restrictions imposed on ψ translate into suitable constraints on the filter coefficients h n . For instance, ψ has compact support if only finitely many h n differ from zero (one then speaks of a finite impulse response or FIR filter). Of course, the goal is to obtain a fast algorithm, and this relies on two aspects, namely, short filters and a pyramidal structure, already familiar in signal processing. Indeed, the rapidity of the algorithms depends crucially on the length of the filters involved, because the pyramidal structure rests on a concatenation of several filters. One major stumbling block is the dilation. This is easy to understand from the very definition of the WT. Indeed, as we have seen above, the WT (1.8) or (1.13) basically a convolution. Once discretized, these formulas become discrete convolutions of digital sequences. Then the point is that, if the sequence ψ # (n) has length N , then the dilated sequence ψ2# (n) has length 2N , and this leads to an algorithm of exponential increase, clearly not admissible. One trick for avoiding this difficulty is to replace the (natural) dilation by a so-called pseudodilation, which consists in inserting a zero between any two successive entries of ψ # (n), and then correcting for the distortion so introduced. In this way, one obtains a fast algorithm. Since the sequences resulting from successive dilations have all the same length, but are full of zeros or holes, this algorithm is known as the “algorithme a` trous” [157,222]. The interesting fact is that this procedure may be extended to very general situations, involving wavelet transforms on abelian groups, which form a kind of intermediate step (“missing link”) between the CWT and the DWT [Kou00,32]. Several other fast algorithms have been designed, mostly along the line proposed by Mallat [Mal99,339]. Altogether, there does exist a Fast Wavelet Transform, exactly as a Fast Fourier Transform. In practical applications, the (sampled) signal is taken in some V J , and then the decomposition (1.47) is replaced by the finite representation V J = V jo ⊕

J −1

Wj .

(1.57)

j= jo

Figure 1.8 shows an example of a decomposition of order 5, namely our familiar signal bumps decomposed over an orthonormal basis of Daubechies d6 wavelets [Dau92]. Thus we take J = 0 and jo = −5 in formula (1.57):

23

1.5 The discrete WT: orthonormal bases of wavelets

Signal 2 1 0 1 0 −1 1 0 −1 1 0 −1 2 0 −2 2 0 −2 2 0 −2 10 0 −10

100

200

300

400

500

600

700

800

900

1000

Fig. 1.8. Five-level decomposition of the bumps signal on an orthonormal basis of Daubechies d6

wavelets. The low resolution approximation c−5 ∈ V−5 is shown on the bottom panel and the five levels of details with increasing resolution, d j ∈ W j , j = −5, . . . , −1, in the next five panels.

V0 = V−5 ⊕ W−5 ⊕ W−4 ⊕ W−3 ⊕ W−2 ⊕ W−1 .

(1.58)

Correspondingly, the signal s ∈ V0 is decomposed as s=

c−5,k φ−5,k +

k∈Z

−1

d jk ψ jk

j=−5 k∈Z

≡ c−5 +

−1

dj.

j=−5

As in Figure 1.5, the figure shows, from bottom up, the low resolution approximation c−5 and the five levels of details with increasing resolution, successively d−5 , d−4 , d−3 , d−2 , d−1 . At this point, we should add a word of caution concerning the numerical implementation of the reconstruction formula associated to (1.57), s=

k∈Z

c jo ,k φ jo ,k +

J −1 j= jo k∈Z

d jk ψ jk .

(1.59)

24

Warm-up: the 1-D continuous wavelet transform

If s is an analog signal, i.e. a function, the approximation coefficients c jo ,k and the wavelet coefficients d j,k are calculated in the standard way, c jo ,k = φ jo ,k |s,

d j,k = ψ jk |s.

However, if the signal is only accessible through sampled values, then the coefficients must be estimated from the latter. This may be done by making some assumption about the signal (for instance, in the “algorithme a` trous” [157,222], the signal is supposed to be a spline function), which amounts to some pre-filtering. Alternatively, one takes for c jo ,k the sampled values themselves. This is a good approximation [Mal99; Section 7.2.3], but without any real theoretical justification. In any case, these procedures generate errors that have to be controlled. See [132] for a comprehensive discussion.

1.6

Generalizations As we just saw, appropriate filters generate orthonormal wavelet bases. However, this result turns out to be too rigid and various generalizations have been proposed (see also the comments in Section 1.6.2). To name a few: biorthogonal wavelet bases, wavelet packets and the Best Basis Algorithm, the lifting scheme and second generation wavelets. We shall refrain from describing these here. Instead, we will give a rather detailed treatment in Chapter 2, for the two-dimensional case. Further information may also be found in [Mey94].

1.6.1

Continuous wavelet packets Besides the full discretization described in Section 1.3, and the discrete WT just discussed, there is an intermediate procedure, which consists in discretizing the scale variable alone, on an arbitrary sequence of values (not necessarily powers of a fixed ratio), but leaving translations fully continuous. The resulting transform has the advantage of being completely covariant with respect to translations, a very desirable feature, for instance, in pattern recognition. If we use dyadic scales, the result is called a dyadic wavelet transform and was introduced in [264,265], precisely for that reason. A detailed account will be given in the 2-D case, in Section 2.4.4. An elegant way of deriving this dyadic WT from the CWT was described in [159], under the name of infinitesimal multiresolution analysis, or continuous wavelet packets. This approach leads to fast algorithms that could put the CWT on the same footing as the discrete WT in terms of speed and efficiency, by extending the advantages of the latter to cases where no exact QMF is available [Tor95,Vdg98,291,360]. While already interesting in 1-D, this method displays its full potential in 2-D, offering a very fast implementation of the so-called directional 2-D wavelets. Accordingly we shall describe it in detail in the next chapter, Section 2.6. Here we shall only sketch briefly the idea, in the version called linear formalism in [Tor95,291].

25

1.6 Generalizations

Given a wavelet ψ, one lumps together all low-frequency components in a scaling function (here we take a > 0) ∞ ∞ x da 1 x ) da , (x) = (1.60) ψ(aξ ψ = ψ(s) ds, (ξ ) = 2 a a x 0 a 1 1 and introduces the integrated wavelet 1 x da 1 2x ψ = ψ(s) ds, (x) = a a2 x x 1/2

)= (ξ

1

) da . ψ(aξ a 1/2

(1.61)

These functions satisfy two-scale relations: (x) = 2 (2x) − (x),

) = (ξ/2) ). (ξ − (ξ

(1.62)

Next, one chooses a regular grid, as opposed to the dyadic one used in the discrete case, namely, xj ≡ 2 j (2 j (· − x)),

xj ≡ 2 j (2 j (· − x))

(1.63)

[note that one uses here the L 1 -normalization (1.19)]. Although the resulting transform will be redundant, it has the great advantage over the conventional DWT of maintaining (integer) translation covariance. Then, exactly as for the DWT, one gets a discrete reconstruction formula: s(x) = xjo |s +

∞

xj |s.

(1.64)

j= jo

Truncating the summation, as usual, one gets thus a finite sum, to be compared with the decomposition (1.57). However, there still remains a major problem. Indeed, we may try and mimick the ˇ gˇ satisfying the following continuous formalism and assume there exist two functions h, relations, analogous to (1.49), (1.52): ) = h(ξ ), ˇ ) (ξ (2ξ

) = g(ξ ), ˇ ) (ξ (2ξ

a.e.

(1.65)

The difficulty now is that these functions are in general not 2π -periodic, which precludes designing any fast (pyramidal) algorithm. There is a way out, however. Since using the regular grid means sampling (x) at unit rate, we have to assume that the function ˇ gˇ always appear is essentially supported in [−π, π ]. Therefore, since the functions h, in a product with , according to the relations (1.65), it is reasonable to approximate them in a neighborhood of zero by 2π-periodic functions h a , g a . In fact [Tor95], there exists a unique pair h, g that minimizes the distance between hˇ and h a , respectively gˇ a and g . These approximate filters, called pseudo-QMFs, will be described at length in Section 2.6.4. The end result is a very fast implementation of the continuous wavelet transform, truly competitive with the discrete wavelet transform.

26

Warm-up: the 1-D continuous wavelet transform

1.6.2

Orthogonal or redundant wavelet expansions? Now we have seen the full spectrum of possible wavelet decompositions, from the minimalist, that is, nonredundant, orthonormal bases to increasingly redundant systems, frames and pseudo-QMFs. Which one to use in practice? Of course, there is no unique answer, it depends on the problem at hand. If it comes to signal compression, for instance, the most economical representation is certainly preferable, thus orthonormal bases will be the first choice. In addition, they often yield the least correlation between wavelet coefficients, and, of course, mathematicians have a long tradition in orthogonal expansions. This explains the popularity of orthonormal bases in many communities. However, this is certainly not the last word. In statistical analysis, for instance, one prefers overcomplete systems to orthonormal bases, because they have higher adaptivity properties and they allow to control the degree of redundancy. In addition, the lack of (even discrete) translation invariance is a serious drawback for all applications involving some pattern recognition. In fact, increasing redundancy has many advantages, in particular, it improves both the quality of reconstruction and the stability with respect to perturbation, e.g., by noise. We will now comment on these two aspects. As an illustration, let us compare the three analyses of the bumps signal made above, the full CWT analysis of Figure 1.3, the translation invariant frame decomposition of Figure 1.5 and the orthonormal basis decomposition of Figure 1.8. The CWT sees correctly all discontinuities, with intensities depending of their strength, measured by the singularity index α, defined in (1.31). The three parts of the signal have, respectively, α = −1 for the δ function, α = 0 for the boxcar function, and α = 1 for the tent or triangle function. Correspondingly, see (1.32), the wavelet coefficients behave as a α , which fits with Figure 1.3 for a small enough. Compare now the two finite decompositions. When a → 0, the wavelet coefficients of the δ function increase, whereas those of the triangle function decrease. As a consequence, both the orthonormal basis and the translation invariant frame barely see the discontinuities of the triangle. The high resolution (small scale) wavelet coefficients are so small that one cannot see them on the figure! Indeed, if we redo the same analysis on the triangular signal alone, using appropriate units, the high resolution coefficients are perfectly seen (Figure 1.9). The reason is that the δ function singularity is so strong that it swamps the whole picture. This is a perfect illustration of the fact that wavelet orthonormal bases are too rigid: the same basis cannot analyze correctly the three pieces of the signal, in a sense it is not local enough! For that reason, substitutes were invented for improving the local character of the analysis. One of the most popular examples is that of a local cosine basis. The idea is to divide the time axis into arbitrary intervals (which overlap slightly for continuity), depending on the signal itself, and performing into each of these an independent discrete Fourier-type decomposition (DCT) [see Figure 1.11 (d)]. We refer to [Mal99] for further information (we quote this technique here for the sake of comparison only).

27

1.6 Generalizations

Signal 1 0.5 0 x 10−3 2 0 −2 x 10−3 5 0 −5 0.01 0 − 0.01 0.02 0 − 0.02 0 0.5 0 − 0.5 10 0 −10

100

200

300

400

500

600

700

800

900

1000

Fig. 1.9. Five-level decomposition of the triangular part of the bumps signal, with the same

presentation as in Figure 1.8.

On the other hand, the frame of Figure 1.5 performs better. The reason here is spatial (time) resolution. In the orthogonal basis case, the resolution becomes so loose at larger scales that even the wavelet coefficients at j = −3 are barely visible on Figure 1.8. But, since the frame is translation invariant, the spatial resolution remains the same at all scales, and the wavelet coefficients are now visible up to j = −1. Another bonus of redundancy is an increased robustness of the representation with respect to small perturbations. Suppose indeed we are working in finite dimension N . Let s ∈ R N be a signal and denote by si , i = 1 . . . N the coefficients of s in a given orthonormal basis. Suppose we perturb each si by a random variable n i (modeling noise, for example). For simplicity, we choose the n i s independent, with zero mean and variance σ 2 . It is easy to verify that the mean square error (MSE) of the perturbed signal s˜ is N σ 2 . Now, if we decompose s in a frame composed of M > N elements, the intuition is that this result should be improved because we have diluted noise in a much larger space. Actually, if the frame is tight and denoting by r = M/N its redundancy, the MSE reduces to [200]: MSE =

Nσ2 . r

Warm-up: the 1-D continuous wavelet transform

5

0

−5 Log MSE

28

−10

−15

−20

0

2

4

6

8 Number of Bits

10

12

14

16

Orthogonal Wavelet Basis Redundant Wavelet Frame

Fig. 1.10. Mean square error of the two reconstructions of the bumps signal, using an orthonormal wavelet basis and a redundant translation invariant wavelet frame, respectively.

Redundancy thus implies stability, or robustness to perturbations of the decomposition coefficients. This is illustrated on Figure 1.10. We have decomposed the bumps signal using an orthonormal wavelet basis and a redundant translation invariant wavelet frame. We have then quantized the respective coefficients and measured in each case the MSE of the reconstructed signal. The plot shows that the redundant representation always yields a better MSE. Furthermore, the distance between the two curves is constant and equal to the redundancy of the frame used. Finally, as a visual aid to the various discrete wavelet schemes, it is instructive to characterize each of them by the tiling of the time–frequency plane it induces. We show a schematic presentation of these in Figure 1.11. On the top row, we have, from left to right, the cases of a Gabor transform and of the DWT (dyadic partition). Note, however, that the Gabor tiling is idealized, since this sharp partition can never be realized, because the time–frequency localization is intrinsically limited (the so-called Balian–Low theorem [Dau92]). On the bottom row, we show the cases of wavelet packets and that of a local cosine basis, respectively.

1.7 Applications of the 1-D CWT

Frequency

Frequency

29

Time

Time

(b)

Frequency

Frequency

(a)

Time

(c)

Time

(d)

Fig. 1.11. Tilings of the time–frequency plane corresponding to four different analysis schemes. (a) The (idealized) Gabor analysis; (b) the discrete WT; (c) wavelet packets; (d) a local cosine basis.

1.7

Applications of the 1-D CWT The CWT has found a wide variety of applications in various branches of physics and/or signal processing. Of course, our main concern in this book will be the practical applications of the 2-D wavelet transform, so we will devote two full chapters to them (Chapters 4 and 5). Nevertheless, we will list here a representative selection of onedimensional applications, in order to convey to the reader a feeling about the scope and richness of the field. Most of the early applications, and the original references, may be found in the proceedings volumes [Com89,Mey91,Mey93]. Another interesting source for applications is the recent volume of Addison [Add02]. In all cases, the CWT is primarily used for analyzing transient phenomena, detecting abrupt changes in a signal or comparing it with a given pattern. r Sound and acoustics The first applications of the CWT were in the field of acoustics. A few examples are musical synthesis, speech analysis [123] and modeling of the sonar system

30

Warm-up: the 1-D continuous wavelet transform

r

r

r

r

r

of bats and dolphins. Other examples include various problems in underwater acoustics, such as the disentangling of the different components of an underwater refracted wave (see Section 5.3 for the 2-D case) and the identification of an obstacle. Geophysics This is the origin of the method, which was designed in an empirical fashion by J. Morlet for analyzing the recordings of microseisms used in oil prospection. More recently, the CWT has been applied to the analysis of various types of geophysical data, e.g., in gravimetry (fluctuations of the local gravitational field), in seismology (arrival time of the various waves), in geomagnetism (fluctuations of the Earth magnetic field [4]) or in astronomy (fluctuations of the length of the day, variations of solar activity, measured by the sunspots, etc). Fractals, turbulence The CWT is an ideal tool for studying fractals, or more generally phenomena with particular properties under scale changes [221]. Thus it is quite natural that the CWT has found many applications in the analysis of (1-D and 2-D) fractals, artificial (diffusion limited aggregates) or natural (arborescent growth phenomena) [Arn95,43,44]. Related to these is the use of the CWT in the analysis of developed turbulence (identification of coherent structures, uncovering of hierarchical structure) [Abr97,163–165]. An interesting example of fractal or self-similar behavior is that of telecommunications network traffic, and here too the WT (although rather the DWT) has given interesting results [1]. We will come back to 2-D fractals and similar objects in Section 5.4. Atomic physics When an atom is hit by a short intense laser pulse, it emits radiation that cover a whole spectrum of harmonics of the laser frequency (experimentally, harmonics of order larger than 400 have been observed). This is a fast and complex physical process, which cannot be understood without a time-frequency analysis. This has been done, both with a Gabor analysis and with wavelets (CWT), yielding for instance the time profile of each individual harmonic [30,31] and the effect of the polarization of the laser field on harmonic generation [39,40]. Spectroscopy This was one of the earliest and most successful applications, in particular for NMR spectroscopy, where the method proves extremely efficient in subtracting unwanted spectral lines or filtering out background noise [60,131,210]. We may note that here, as in the previous application, a Gabor analysis may be fully competitive with the wavelet analysis [33,34]. Medical and biological applications The CWT has been used for analyzing or monitoring various electrical or mechanical phenomena in the brain (EEG, VEP) or the heart (ECG) [Ald96,Tho98,354]. It also yields good models for the auditory mechanism [123]. Another striking result is the

31

1.7 Applications of the 1-D CWT

characterization of long-range correlations in DNA sequences (and the solution of a long-standing puzzle) by Arn´eodo and his group [49]. r Analysis of local singularities The strong point of the CWT is to detect singularities in a signal, but it yields also a fine characterization of their strengths (Lipschitz regularity, expressed via the local H¨older exponents), using the homogeneity relations (1.32) and (1.34), in particular in the case of oscillating singularities [48,50,51]. r Shape characterization A particular case of analysis of local singularities is the determination of the shape of an object, a standard problem in image processing, for instance in robotic vision. A novel approach [Ces97,Cos01,21] consists in treating the contour of the object as a complex curve in the plane and analyzing it with the 1-D CWT. The method benefits from all the good properties of the wavelet transform, for instance its robustness to noise, and looks promising for applications. Since this is in fact a 2-D problem, we will analyze it in detail, in Section 5.4.2. r Industrial applications Here again the important aspect is monitoring, for instance in detecting anomalies in the functioning of nuclear, electrical or mechanical installations. A typical application is the analysis of the behavior of materials under impact made in [358] and discussed above for illustrating the concept of ridge and skeleton. In that paper, the same signal is analyzed with a Gabor transform, a CWT with a Morlet wavelet, a CWT with a Mexican hat, a Wigner–Ville transform, and the respective merits of each method are compared. Another, closely related, application is the determination, with a Morlet wavelet, of the vibration normal modes of a high tower excited by wind [259]. The results of this paper fully confirm the previous ones and, in particular, emphasize the role of the width parameter of the Gaussian.

2

The 2-D continuous wavelet transform

Chapter 1 has given us a brief overview of the basic facets of the CWT in the simpler one-dimensional (1-D) case, including its relationship with the various discrete approaches and a glimpse of some applications. Now it is time to enter the proper subject of the book, namely, the two-dimensional (2-D) wavelet analysis.

2.1

Derivation In 1-D, the CWT (1.8) amounts to projecting the signal onto the wavelet ψb,a , obtained by translation and dilation of the mother wavelet ψ. Thus the transform is fully determined by these elementary operations of the line. Accordingly, in order to derive the CWT in 2-D, a good starting point is to consider first the elementary operations we want to apply to our signals. Actually, as we will see later (Chapter 6), this point of view allows one to extend the CWT to much more general situations, such as wavelets in higher dimensions, wavelets on the sphere, time-dependent wavelets, etc.

2.1.1

Images and elementary operations on them By an image, we mean a two-dimensional signal of finite energy, represented by a complex-valued function defined on the real plane R2 and square integrable, i.e., a function s ∈ L 2 (R2 , d 2 x): 2 s = d 2 x |s( x )|2 < ∞ (2.1) R2

(sometimes it is useful to take s integrable as well). In practice, a black and white image will be represented by a bounded non-negative function: 0 s( x ) M, ∀ x ∈ R2 (M > 0),

(2.2)

the discrete values of s( x ) corresponding to the level of gray of each pixel. However it is useful to keep general functions s as above. In fact, one often considers also as admissible signals generalized functions (distributions), such as a delta function 32

33

2.1 Derivation

δ( x − xo ), a plane wave exp(i k · x), a fractal measure, etc. This will be justified below (see the comments after Definition 2.1.3). The Fourier transform of the signal s is defined, as usual, by 1 s(k) ≡ (Fs)(k) = d 2 x e−i k·x s( x ), (2.3) 2π R2 where k ∈ R2 is the spatial frequency and k · x = k1 x1 + k2 x2 is the Euclidean scalar 2 = k · k. Of course, the Fourier transform is unitary (Parseval product. We also write |k| relation): s ∈ L 2 (R2 , d 2 k)

and

s2 = s2 .

(2.4)

Given an image s, all the geometric operations we want to apply to it are obtained by combining three elementary transformations of the plane, namely, rigid translations in the plane of the image, dilations or scaling (global zooming in and out) and rotations. (More complicated operations are sometimes applied, such as deformations (shearing), but we will not consider them here. We will come back to this point in Chapter 7.) Explicitly, the transformations act on x ∈ R2 in the familiar way: (i) translation by b ∈ R2 : x → x = x + b; (ii) dilation by a factor a > 0 : x → x = a x; (iii) rotation by an angle θ : x → x = rθ ( x ), where rθ is the usual 2 × 2 rotation matrix: cos θ − sin θ , 0 θ < 2π. rθ ≡ sin θ cos θ It will prove convenient to combine a rotation by an angle θ and a dilation by a > 0 into a single 2 × 2 matrix, namely, a cos θ −a sin θ h = h(a, θ) = . (2.5) a sin θ a cos θ Using this form, we verify that x · x = a cos θ | x |2 , demonstrating that the angle be

tween x and x is indeed θ and that x is scaled by the amount a relative to x . Finally, combining all three operations, we get as general transformation in the plane x + b. x → x = h

(2.6)

In the present context, these transformations are represented by the following unitary operators in the space L 2 (R2 , d 2 x) of finite energy signals: b ∈ R2 ; (i) translation : (Tb s)( x ) = s( x − b), (2.7) −1 −1 (ii) dilation : (Da s)( x ) = a s(a x), a > 0; (2.8) (iii) rotation : (Rθ s)( x ) = s(r−θ ( x )), θ ∈ [0, 2π). (2.9)

34

The 2-D continuous wavelet transform

In addition, we introduce the modulation operator: (E b s)( x ) = ei b·x s( x ), b ∈ R2 .

(2.10)

Then a straightforward calculation yields the commutation rules among the operators (2.7)–(2.9), and with the Fourier operator (2.3): Tb Da = Da Tb/a F Da = D1/a F, , Tb Rθ = Rθ Tr−θ (b) , F Tb = E −b F, F Rθ = Rθ F. Rθ Da = Da Rθ ,

(2.11)

Combining now the three operators (2.7)–(2.9), we define the unitary operator a, θ ) = Tb Da Rθ , U (b,

(2.12)

which acts on a given function s as a, θ )s ( U (b, x ) ≡ sb,a,θ ( x ) = a −1 s(a −1 r−θ ( x − b)),

(2.13)

or, equivalently, in the space of Fourier transforms,

= a e−i b·k s(ar−θ (k)). s (k) b,a,θ

(2.14)

If the function s is rotation invariant, we simply omit the index θ: sb,a x ) = a −1 s(a −1 ( x − b)). (

(2.15)

We may remark that, here, contrary to the 1-D case, it is sufficient to take positive dilations, since the effect of a negative dilation a < 0 may be obtained by combining a positive one, a > 0, with a rotation by θ = π . The geometrical effect of these transformations is easily visualized assuming s and s to be well localized, for instance in an ellipse. An example is shown in Figure 2.1 (in Section 2.3.1).

2.1.2

Wavelets and continuous wavelet transform As in 1-D, a wavelet is a particular type of finite energy signal, whose properties make it a good analyzing tool. Thus we define, as in (1.10): Definition 2.1.1 . A two-dimensional wavelet is a complex-valued function ψ ∈ L 2 (R2 , d 2 x) satisfying the admissibility condition: k)| 2 |ψ( cψ ≡ (2π)2 d 2 k < ∞, (2.16) 2 |k| R2 is the Fourier transform of ψ and |k| 2 = k · k = (k1 )2 + (k2 )2 . where ψ The origin of this condition will be clarified below.

35

2.1 Derivation

If ψ is regular enough, the admissibility condition (2.16) implies the following easier one, which simply means that the wavelet has zero mean: 0) = 0 ⇐⇒ ψ( d 2 x ψ( x ) = 0. (2.17) R2

Strictly speaking, the condition (2.17) is only necessary, but in fact it is almost sufficient (see [Dau92] for a precise mathematical statement), and for all practical purposes (2.17) may be taken as admissibility condition. Intuitively, as in one dimension, it expresses the fact that a wavelet must be an oscillating function. Clearly the three unitary operators Tb , Da , Rθ preserve the admissibility condition, a, θ ). Hence any function ψb,a,θ a, θ)ψ obtained and so does therefore U (b, = U (b, from a wavelet ψ by translation, rotation or dilation is again a wavelet. Thus the given wavelet ψ generates the whole family Dψ = {ψb,a,θ }, indexed by the elements b ∈ R2 , a > 0, θ ∈ [0, 2π). In the sequel we will denote by G this four-dimensional parameter space. a, θ) ∈ G} is a Proposition 2.1.2 . The linear span of the family Dψ = {ψb,a,θ , (b, 2 2 dense subspace of L (R ). Proof . Let f ∈ L 2 (R2 , d 2 x) be orthogonal to every vector in the family Dψ , that is a, θ) ∈ G. This means ψb,a,θ | f = 0, ∀ (b, a, θ ) ∈ G, | f = a d 2 k ei b·k ψ(ar ∀ (b, ψ −θ (k)) f (k) = 0, b,a,θ R2

−θ (k)) f (k) = 0 a.e., for all a > 0, θ ∈ [0, 2π ). Now the joint which implies that ψ(ar is a “patch” action of rotations and dilations on R2 is transitive. Thus, if the support of ψ −θ (k)) will cover (for instance, a disk or an ellipse) in the k-plane, the supports of ψ(ar the whole plane when a and θ vary over their range. Therefore, this implies f (k) = 0 a.e., that is, f = 0. Note that rotations are needed to get that result. A naive generalization of the 1-D formalism would consist in combining translations with separate dilations along the x- and the y-axis. But then, if only positive dilations are used, each quadrant in the k-plane is invariant and additional conditions on the wavelet ψ would be necessary for the argument of the proof above to work – thus spoiling the result. Actually, this is precisely the technique used in the 2-D DWT, where the 2-D multiresolution is obtained by taking the tensor product of two 1-D copies, one in x, one in y (see Section 2.5.1). As we shall see later (see Section 2.5.2), this approach, while commonly used in practice, has severe shortcomings. As a consequence of Proposition 2.1.2, any vector in L 2 (R2 , d 2 x) is uniquely determined by its projections on the vectors of the family Dψ . This justifies the basic definition of the CWT.

36

The 2-D continuous wavelet transform

Definition 2.1.3 . Given an image s ∈ L 2 (R2 , d 2 x), its continuous wavelet transform (with respect to the fixed wavelet ψ), S ≡ Tψ s is the scalar product of s with the a, θ): transformed wavelet ψb,a,θ , considered as a function of (b, a, θ ) = ψb,a,θ S(b, |s s( d 2 x ψ(a −1 r−θ ( x − b)) x) = a −1 R2 s(k). d 2 k ei b·k ψ(ar =a −θ (k))

(2.18) (2.19) (2.20)

R2

The relations (2.18)–(2.20) permit us to extend the formalism beyond the Hilbert space framework. As explained above, the signal s may be taken as a singular function (a distribution), provided the wavelet ψ is sufficiently regular (most wavelets used in practice are smooth functions, see below). Before exploring in detail the mathematical properties of the CWT, it is instructive to exhibit two typical 2-D wavelets (actually the simplest ones). (1) The isotropic Mexican hat wavelet This wavelet is simply the Laplacian of a Gaussian: ψH ( x ) = (2 − | x |2 ) exp(− 12 | x |2 ), H (k) = |k| 2 exp(− 1 |k| 2 ). ψ

(2.21)

2

(2) The Morlet wavelet This wavelet is essentially a plane wave within a Gaussian window: ψM ( x ) = exp(i ko · x) exp(− 12 | x |2 ) + corr.; M (k) = exp(− 1 |k − ko |2 ) + corr. ψ

(2.22)

2

As in 1-D, a a correction term must be added in order to satisfy the admissibility condition (2.17), but in practice one will arrange that this term be numerically negligible and thus can be omitted (it suffices to choose the norm |ko | of the wave vector large enough). The first wavelet is real, the Morlet wavelet is complex. They have very different properties and, naturally, they will be used in quite different situations. Both wavelets, and many more, will be studied in detail in Chapter 3, and many examples of applications in Chapters 4 and 5.

2.2

Basic properties of the 2-D CWT The main properties of the continuous wavelet transform are conveniently expressed in terms of a linear map Wψ from the space of finite energy signals L 2 (R2 , d 2 x) into the space of transforms. We summarize them in three propositions [Mur90,13,15,283].

37

2.2 Basic properties of the 2-D CWT −1/2

Proposition 2.2.1 . Let the map Wψ : s → cψ

S be defined by

a, θ) = cψ−1/2 ψb,a,θ (Wψ s)(b, |s, s ∈ L 2 (R2 , d 2 x), where cψ is the constant given in (2.16). Then: (1) Wψ conserves the norm of the signal, thus its total energy: 2 da 2 d b 3 dθ |S(b, a, θ )| = cψ d 2 x |s(x)|2 , a G R2

(2.23)

(2.24)

i.e., it is an isometry from the space of signals into the space of transforms. The latter is a closed subspace Hψ of L 2 (G, dg), where dg ≡ a −3 d 2 b da dθ is the natural measure on G. Equivalently, the family of wavelets {ψb,a,θ }, with b ∈ R2 , a > 0, and 0 θ < 2π, generates a resolution of the identity: da −1 cψ d 2 b 3 dθ |ψb,a,θ ψb,a,θ | = I. (2.25) a G (2) Since it is an isometry, the map Wψ is invertible on its range Hψ , and the inverse transformation is the adjoint of Wψ . This means that the image s( x ) may be a, θ) by the formula: reconstructed from its wavelet transform S(b, da a, θ). s( x ) = cψ−1 d 2 b 3 dθ ψb,a,θ ( x ) S(b, (2.26) a G Proof . The relation (2.24) follows from a straightforward calculation: da a, θ)|2 = d 2 b 3 dθ |S(b, a G da 2 2 dθ = d k d k d 2 b a R2 R2 G

−θ (k)) −θ (k )) ψ(ar s(k ) s(k) × ei b·(k−k ) ψ(ar ∞ 2π da −θ (k))| 2 | 2 d 2 k dθ |ψ(ar s(k)| = (2π)2 a 0 0 R2 (the exchange of integrals is justified by Fubini’s theorem). Introducing polar coordi k) p (ρ, φ), with ρ ≡ |k|, ≡ψ we get nates: ψ( ∞ ∞ da 2π da 2π −θ (k))| p (aρ, φ − θ )|2 2= dθ |ψ(ar dθ |ψ a a 0 0 0 0 ∞ 2π dρ p (ρ , θ )|2 = dθ |ψ ρ 0 0 d 2 k 2 = |ψ(k )| . |2 R2 |k By comparison with the definition (2.16) of cψ , and using Plancherel’s theorem, this proves the statement.

38

The 2-D continuous wavelet transform

Then (2.25) is simply a reformulation of (2.24), in view of the definition (2.23) of −1/2 Wψ : s → cψ S. Finally, the reconstruction formula (2.26) follows immediately by applying both sides of (2.25) to a signal s( x ) and taking into account the definition (2.23) of Wψ .

Three remarks are in order here. First, the relation (2.25) must be taken as a weak integral, that is, both sides are equal when sandwiched between arbitrary vectors. This precisely means that the reconstruction formula (2.26) holds in the weak sense. But, in fact, much more is true, and actually necessary for obtaining good approximation schemes, namely, the relation (2.26) holds in strong L 2 convergence. This will be demonstrated in Section 2.6.1, in two different versions. Second, the measure dg ≡ a −3 d 2 b da dθ on G is precisely the unique measure (up to normalization) that is invariant under all the operations of translation, dilation, and rotation (this is why we have called it natural). Third, the reconstruction formula (2.26) may also be proven by an explicit calculation of the adjoint map of Wψ : −1/2

f |sL 2 (R2 ) = cψ

= cψ−1

−1/2

f |Wψ∗ SL 2 (R2 ) = cψ Wψ f |SL 2 (G) da a, θ). d 2 b 3 dθ f |ψb,a,θ S(b, a G

It is the possibility of having a reconstruction formula (2.26) that justifies the admissibility condition cψ < ∞ imposed on wavelets. However, (2.26) is not only a reconstruction formula, it also means that the wavelet transform, like its 1-D counterpart, provides a decomposition of the signal in terms of the analyzing wavelets ψb,a,θ , a, θ ). Under both interpretations, this formula leads in practice with coefficients S(b, to discretization problems (see Section 2.4). In the same spirit, it is interesting to see the inverse Fourier transform 1 s( x) = s(k), (2.27) d 2 k ei k·x 2π R2

as the decomposition of the signal into the improper basis {ei k·x , k ∈ R2 } of eigenvectors of the translation operators (2.7). In view of the crucial importance of dilations in the wavelet context, it is useful to write down also the polar coordinate version of the Fourier transform, which involves the (improper) eigenvectors {einϕ , n ∈ Z} of the rotation operator (2.9) and those of the dilation operator (2.8), {riν , ν ∈ R}: 2π ∞ 1 dr −iν f (ν, n) = dϕ e−inϕ (2.28) r r f (r, ϕ), 2π 0 r 0 ∞ ∞ 1 inϕ f (ν, n). (2.29) e dν r iν r f (r, ϕ) = 2π n=−∞ −∞

39

2.2 Basic properties of the 2-D CWT

Of course, we recover the well-known fact that the polar coordinate version of the Fourier transform is a combination of a Mellin transform in the radial variable r and a Fourier series in the angle ϕ. Actually, as in one dimension, the reconstruction formula (2.26) may be generalized in several ways. First, the wavelet used for the analysis, ψ, and the one used for the reconstruction, χ , need not coincide, they have only to satisfy a cross-admissibility condition [Hol95,223], namely, 0 < |cψχ | < ∞, where d 2 k 2 < ∞. ψ(k) χ (k) (2.30) cψχ = (2π) 2 R2 |k| Then one gets a more general reconstruction formula: da −1 a, θ). d 2 b 3 dθ χb,a,θ ( x ) (Wψ s)(b, s( x ) = cψχ a G

(2.31)

The proof consists in a straightforward verification, including some interchanges of integrals justified by Fubini’s theorem. As we shall see later, this is the analog, in the continuous case, of the bilinear scheme commonly used in the discrete approach, namely the construction of biorthogonal wavelet bases (see Section 2.5.2). In particular, if one takes for the reconstruction wavelet χ a delta function, one obtains the simplified reconstruction formula: ∞ da 2π −1 s( x ) = cψδ dθ S( x , a, θ), (2.32) a2 0 0 where cψδ =

R2

d 2 k ψ(k). 2 |k|

On the other hand, if ψ is rotation invariant, the wavelet transform S does not depend on θ and we obtain, instead of (2.26), a simpler reconstruction formula: ∞ da a). d 2 b ψ ( x ) S(b, (2.33) s( x ) = 2π cψ−1 a 3 b,a R2 0 Finally, combining the two preceding points, one obtains the simplified reconstruction formula originally used by Morlet in 1-D [206], in which one reconstructs the original image by summing over scales only: ∞ da −1 s( x ) = 2π cψδ S( x , a). (2.34) a2 0 Next, a characteristic feature of the CWT, in fact shared by a large class of transformations, as we shall see later (Chapter 6), is the existence of a so-called reproducing kernel, which actually is nothing but the wavelet transform of the wavelet itself, that is, the autocorrelation function of the wavelet. More precisely:

40

The 2-D continuous wavelet transform

Proposition 2.2.2 . The projection from L 2 (G, dg) onto the range Hψ of Wψ , the space a, θ ) is the of wavelet transforms, is an integral operator whose kernel K (b , a , θ |b, autocorrelation function of ψ, also called reproducing kernel: a, θ ) = cψ−1 ψb ,a ,θ |ψb,a,θ K (b , a , θ |b, .

(2.35)

Therefore, a function f ∈ L 2 (G, dg) is the wavelet transform of a certain signal iff it satisfies the reproduction property: da

a, θ) f (b, a, θ). d 2 b 3 dθ K (b , a , θ |b, (2.36) f (b , a , θ ) = a G Proof . Since Wψ is an isometry from L 2 (R2 , d 2 x) into L 2 (G, dg), i.e., Wψ∗ Wψ = I , its range Hψ is a closed subspace and the corresponding projection operator is Pψ = Wψ Wψ∗ . Thus a vector f ∈ L 2 (G, dg) belongs to Hψ iff f = Pψ f . Explicitly, this gives: f (b , a , θ ) = Wψ Wψ∗ f (b , a , θ ) da −1 a, θ ), = cψ d 2 b 3 dθ ψb ,a ,θ |ψb,a,θ f (b, a G which proves (2.35)–(2.36).

Because it may be interpreted as the autocorrelation function of the wavelet, the reproducing kernel leads to the notion of correlation length, that is, it determines the a, θ parameter space. As such, it plays region of influence of a given wavelet in the b, a role in the determination of the capabilities of a given wavelet (calibration), and in particular in the process of discretization. We will discuss these features in Chapter 3, Section 3.4. Finally, the continuous wavelet transform has the important property of covariance (improperly called invariance in the signal processing literature) under all the operations used in its definition.

Proposition 2.2.3 . The map Wψ is covariant under translations, dilations and rotations, a, θ) implies the following which means that the correspondence Wψ : s( x ) → S(b, ones: s( x − bo ) → S(b − bo , a, θ ) ao−1 a, θ) ao−1 s(ao−1 x) → S(ao−1 b,

(2.38)

a, θ − θo ). x )) → S(r−θo (b), s(rθo (

(2.39)

(2.37)

41

2.3 Implementation and interpretation of the 2-D CWT

It is worth noting that, conversely, the wavelet transform is uniquely determined by the three conditions of linearity, covariance and energy conservation, plus some continuity [Mur90]. These covariance relations, which are proved by a straightforward calculation, have a crucial importance for the applications. Translation covariance (2.37), often called improperly shift invariance, is lost in the standard formulation of the discrete WT, based on multiresolution (see Definition 1.5.1 and Section 2.5.1), and this generates many problems in practice, for instance in pattern recognition. Covariance under dilations, (2.38), is the basis for the application of the wavelet transform to the analysis of fractals (see Section 5.4). Finally, joint covariance under rotations and dilations justifies the use of the CWT for detecting rotation–dilation (inflation) properties of several classes of 2-D patterns, for instance, Penrose tilings of the plane or diffraction patterns of quasicrystals. We will discuss this recent application in Chapter 4, Section 4.5. As a final remark, we emphasize that, as in the 1-D case, the choice of the normalization factor a −1 in (2.8) or (2.12) is not essential and is made mainly for mathematical reasons. It is the only one that makes the dilation operator Da , and thus the wavelet transform, unitary: ψb,a,θ 2 = ψ2 and Wψ s2 = s2 , as stated in Proposition 2.2.1. In practice, one often uses instead a factor a −2 , so as to enhance the high-frequency part of the signal, and thus to make more conspicuous its singularities, if any. This amounts to introducing, instead of the unitary Da , a nonunitary dilation operator D a , which preserves the L 1 -norm of the signal: D a ψ1 = ψ1 and ψ(b,a,θ ) 1 = ψ1 , where a −2 −1 −1 = a ψb,a,θ = Tb D Rθ ψ = a ψ(a r−θ ( x − b)) . Correspondingly, one deψ(b,a,θ) 1 fines the L -normalized transform ˘ b, a, θ ) = ψ(b,a,θ) |s. S(

(2.40)

This transform is also useful for making contact with the so-called dyadic wavelet transform (see Section 2.4.4), in particular, for the design of fast algorithms, using the continuous wavelet packets developed in Section 2.6. We will meet it again in Chapter 9, while extending the CWT to the 2-sphere.

2.3

Implementation and interpretation of the 2-D CWT

2.3.1

Interpretation of the CWT as a singularity scanner In order to get a physical interpretation of the CWT, we notice that in signal analysis, as in classical electromagnetism, the L 2 norm is interpreted as the total energy of the a, θ)|2 as the energy signal. Therefore, the relation (2.24) suggests we interpret |S(b, density in the wavelet parameter space [284]. Assume now, as in 1-D, that the wavelet ψ is fairly well localized both in position Then so is the transformed wavelet ψb,a,θ space ( x ) and in spatial frequency space (k). ,

42

The 2-D continuous wavelet transform

rotated by θ and dilated by a. Because with effective support suitably translated by b, (2.19) is essentially a convolution with a function ψ of zero mean, the transform a, θ ) is appreciable only in those regions of parameter space (b, a, θ) where the S(b, signal is. Thus we get an appreciable value of S only where the wavelet ψb,a,θ “matches” the features of the signal s. In other words, the CWT acts on a signal as a local filter in a, θ: S(b, a, θ) sees only that portion of the signal that lives around all four variables b, a, θ and filters out the rest. Therefore, if the wavelet is well localized, the energy b, density of the transform will be concentrated on the significant parts of the signal. This is the key to all the approximation schemes that make wavelets such an efficient tool. In order to clarify the filtering effect in scale and angle variables, we rewrite the expression (2.20) of the CWT in polar coordinates k = (ρ, φ): s (k) S(b, a, θ ) = a d 2 k ei b·k ψ(ar (2.41) −θ (k)) 2 R 2π ∞ φ − θ ) s (ρ, φ). (2.42) ρ dρ dφ eibρ cos φ ψ(aρ, =a 0

0

On the last relation, we see that the CWT amounts to a convolution in the scaleangle variables ρ, φ. In order to better appreciate this, we switch throughout to polar coordinates, so the Fourier transform turns into a Mellin transform, as seen in (2.29). In position space, write (2.19) as 1 S(b, a, θ ) = d 2 x ψ(r−θ (a −1 x)) sb ( x ), a R2 In polar coordinates x = (r, ϕ), we get x ) ≡ s( x + b). where sb ( ∞ 2π dr dϕ a −1r ψ(a −1r, ϕ − θ) sb (r, ϕ), S(b, a, θ ) = 0

(2.43)

0

or, in conjugate variables [see (2.29)], ∞ ∞ 1 inθ n) S(b, a, θ ) = e dν a iν ψ(ν, sb (ν, n). 2π n=−∞ −∞

(2.44)

Performing the change of variables r = eu , a = ev , we obtain ev , θ ) = S(b,

(2.45)

∞

du −∞

0

2π

dϕ eu−v ψ(eu−v , ϕ − θ) eu sb (eu , ϕ).

(2.46)

Upon introducing the functions u, ϕ) = eu sb (eu , ϕ), G(u, ϕ) = ψ(eu , ϕ), F(b,

(2.47)

43

2.3 Implementation and interpretation of the 2-D CWT

(2.46) turns into

ev , θ ) = S(b,

∞

du −∞

2π

u, ϕ). dϕ G(u − v, ϕ − θ) F(b,

(2.48)

0

a, θ ) reduces into that of computing a convolution Thus, the problem of computing S(b, of the signal F with a function G of zero mean, hence the filtering effect – and the rapidity of the algorithm [286]. to be as Let us make more precise the support properties of ψ. Assume ψ and ψ well localized as possible (but in a way still compatible with the Fourier uncertainty property), namely, ψ has for numerical support (i.e., the region outside of which the while function is numerically negligible) a “disk” of diameter T , centered around 0, ψ has for numerical support a “disk” of diameter , centered around ko . Then, for the transformed wavelets ψb,a,θ and ψ we have, respectively: b,a,θ r supp ψ is a “disk” of diameter aT centered around b and rotated by an angle b,a,θ θ; r supp ψ is a “disk” of diameter /a, centered around rθ (ko )/a and rotated by b,a,θ θ. Notice that the product of the two diameters is constant (we know it has to be bounded below by a fixed constant, by Fourier’s theorem). These support properties are illustrated in Figure 2.1 for an elliptic shape. As a consequence, we may characterize the filter properties of the wavelet. r If a ! 1, ψ is a wide window, whereas ψ is very peaked around a small b,a,θ b,a,θ spatial frequency rθ (ko )/a: this transform will be most sensitive to low spatial frequencies. r If a 1, ψ is a narrow window and ψ is wide and centered around a high b,a,θ b,a,θ spatial frequency rθ (ko )/a: this wavelet has a good localization capability in the space domain and is mostly sensitive to high spatial frequencies. Thus wavelet analysis operates at constant relative bandwidth, k/k = const, where Therefore, the analysis is most efficient at high spatial frequencies or small k ≡ |k|. scales, and so it is particularly apt at detecting discontinuities in images, either point singularities (contours, corners) or directional features (edges, segments). In other words, the CWT is a singularity scanner. Actually, we will see later that it is also a singularity analyzer. For instance, the CWT allows one to measure fractal dimensions in images (see Chapter 5, Section 5.4). In addition to these localization properties, one often imposes on the analyzing wavelet ψ a number of additional properties, for instance, restrictions on the support Or ψ may be required, as in the 1-D case, to have a certain number of ψ and of ψ. of vanishing moments, up to order N 1 (by the admissibility condition (2.17), the moment of order 0 must always vanish):

The 2-D continuous wavelet transform

(b)

(a)

............ ..... ........ ... .... ... .. .. .. ... .. .. ... ... ... ... ... .. ... ... .. .. .. . . .. . . ... ... ... .... ... ........ ........... ....

✻

.. ..................... ............ ... .......... ........ .. ....... . ...... . . . . ... . .. .. ... ... .. . . . . . . .... ... .. ... .. ... .... ... ... .... . .. ..... .. ...... ..... ....... ...... ... . . . . . . . ..... .... .................................

✻

❅ I ❅

✻

❅

T

✟✟ b

✲

✟

✯ ✟✟

✟

✟

θ

❅ aT ❅ ❘ ❅ ✲

❄

(c)

✻

✒ ✒ /a θ✠

................................................. ........... ...... ...... .... .... .. .... .... ... o ....... .... . . . . . . ............. . ............................................

✲ k

✛

(d)

✻ .......................... .... ...... .... ... ... ... ... .... ... .... .. ...... . ........ ..................

44

✲

✲

Fig. 2.1. Support properties under the basic operations: (top) in the time domain: (a) the original

signal ψ( x ); (b) the modified signal ψb,a,θ ( x ), with b = (2.4, 1.2), a = 1.5, θ = 45◦ ; (bottom) in k); (d) the modified signal ψb,a,θ the frequency domain: (c) the original signal ψ( (k).

x ) = 0, d 2 x x α y β ψ(

x = (x, y),

0 α + β N.

(2.49)

This property improves its efficiency at detecting singularities in the signal. Indeed, the transform (2.19) is then blind to the smoothest part of the signal, that which is polynomial of degree up to N – and less interesting, in general. Only the sharper part remains, including all singularities (jumps in the signal or one of its derivatives, for instance). Equivalently, ψ detects singularities in the (N + 1)th derivatives of the signal [264]. For instance, if the first moments (N = 1) vanish, the transform will erase any linear trend in the signal, such as a linear gradient of luminosity. Conversely, if the signal is rough, a fortiori if it is a measure (as in the analysis of fractals [43,221]), it is sufficient to take a wavelet with no nontrivial vanishing moment, i.e., no condition has to be imposed beyond (2.17). Altogether, as in the 1-D case, the 2-D wavelet transform may be interpreted as a mathematical, direction selective, microscope, with optics ψ, magnification 1/a and orientation tuning parameter θ [Arn95,44]. Two features must be emphasized here. First, the magnification (zoom) 1/a is global, independently of the direction, because we have excluded distortions of the image. Then, there is the additional property of directivity, given by the rotation angle θ. This last feature opens the way to a whole new

45

2.3 Implementation and interpretation of the 2-D CWT

class of applications, in which directions play an essential role. We will detail some of them in Chapters 4 and 5.

2.3.2

The CWT as a phase space representation In order to get a better insight, it is worth recasting the basic formulas (2.18)–(2.20) into a different form. First, we notice that the CWT is in fact a phase space representation (in the usual sense of Hamiltonian mechanics). To see this in a simple way, we observe onto itself. that the correspondence k ⇔ (a −1 , θ) is a bijection from R2∗ ≡ R2 \ {0} Thus, writing κ = a −1 and p ≡ (a −1 , θ) = (κ, θ) ∈ R2∗ , we get da dθ = κ dκ dθ = d 2 p, a3

(2.50)

so that the measure on G becomes simply the volume element of R2 × R2∗ : da dθ = d 2 b d 2 p. (2.51) a3 Thus, the full four-dimensional parameter space of the 2-D WT, G, may be interpreted as phase space, with q ≡ b the position variable and the pair (a −1 , θ) ≡ (κ, θ) playing the rˆole of spatial frequency p, expressed in polar coordinates. The same result holds in the 1-D case [122,259]: a −1 defines the frequency scale, so that the full parameter space of the 1-D WT, the time-scale half plane, is in fact a time–frequency space, thus a phase space. Of course, this interpretation is borne out by mathematical analysis (see Chapter 7). The variable p follows also the common practice in image processing: a = 0 is the horizon in spatial frequency plots, corresponding to extremely high frequencies. It is amusing to note that the same interpretation is even supported by some physiological evidence, namely the so-called orientation hypercolumns of Hubel and Wiesel [DeV88,Duv91,226]. In certain species, cortical neurons are organized into columns, whose sensitivity to position, orientation, and frequency variables correspond exactly to the geometry of R2 × R2∗ just described. In order to manifest the fact that the CWT is really a phase space realization of the signal, we express it explicitly into phase space variables ( q , p). For any vector x = (x, y) = (r cos ϕ, r sin ϕ), with polar coordinates (r, ϕ), define the matrix x −y s(x ) = = r rϕ (rϕ is the 2 × 2 rotation matrix). (2.52) y x d 2 b

x )z = s(z ) x and rθ ( x ) = s( x )eθ , where eθ denotes a unit One shows immediately that s( vector in the direction θ. Then one has = s( p)−1 k = s(k) R p , ar−θ (k) (2.53) | p |2 where R denotes the reflection with respect to the x-axis. We come back now to the expression (2.20) of the CWT and rewrite it in terms of the phase space variables ( q , p):

46

The 2-D continuous wavelet transform

a, θ ) ≡ S(b, S( q , p) = | p |−1

R2

s (k). d 2 k ei q·k ψ( s( p)−1 k)

(2.54)

Alternatively, one may consider the “inverse” phase space variables v = (a, −θ ) ≡ R p/| p |2 . Clearly s( p)s( v ) = I. Although the variable v is less natural, its use some = times simplifies the computations (see Section 5.3, for instance). Since ar−θ (k) v , one gets for the CWT s(v )k = s(k) ˇ s (k). S(b, a, θ ) ≡ S(b, v) = | v| d 2 k ei b·k ψ( s(v )k) (2.55) R2

2.3.3

Visualization of the CWT: the various representations In practice, once the CWT of a given signal s( x ) has been computed, one immediately a, θ) is a function of four variables: faces a problem of visualization. Indeed, S(b, 2 two position variables b = (bx , b y ) ∈ R2 , and the pair (a, θ ) ∈ R+ ∗ × [0, 2π) R∗ −1 (equivalently, (a , θ)). Now, to compute and visualize the full CWT in all four variables is hardly possible. Therefore, in order to obtain a manageable tool, some of the variables, a, θ, bx , b y must be eliminated. There are two ways of achieving this. The first one consists in fixing the value of some of the variables. In other words, one must restrict oneself to a section of the parameter space. Of course, this makes sense only if the variables in question may take arbitrary values in a continuous range. Alternatively, one may integrate out the variables in question. Using the proper part of the natural measure a −3 d 2 b da dθ , one obtains in this way partial energy densities, a, θ )|2 over all the variables is interpreted as the total energy since the integral of |S(b, of the signal, as results from (2.24). This procedure turns out to be crucial whenever the relevant values of the variables to be eliminated (typically, the scale variable a) take only discrete values. We will see an illuminating example of the difference between the two approaches in the problem of symmetry detection in patterns, discussed in Section 4.5. Let us treat first the problem of sections, the partial energy densities will be discussed in detail in Section 2.3.4. In general, one considers two- and three-dimensional sections. While there are six possible choices of 2-D sections, the geometrical considerations made above indicate that two of them are more natural. Either (a, θ) or (bx , b y ) are fixed, and the WT is treated as a function of the two remaining variables. The corresponding representations have the following characteristics [13,19]. (1) The position representation: a and θ are fixed and the CWT is considered as a function of position b alone (this amounts to take a set of snapshots, one for each value of (a, θ ), which may then be collected together into a video sequence). The position representation is the standard one, and it is useful for the general purposes of image processing: detection of position, shape and contours of objects; pattern recognition; image filtering by resynthesis after elimination of unwanted features α), in (noise, for instance). Alternatively, one may use polar coordinates, b = (|b|,

47

2.3 Implementation and interpretation of the 2-D CWT

and aspect or perception angle which case the variables are interpreted as range |b| α, another familiar representation of images. the CWT is considered as a function of (2) The scale-angle representation: for fixed b, scale a and anisotropy angle θ, i.e., of spatial frequency. In other words, one looks and observes all scales and all at the full CWT as through a keyhole located at b, directions at once. The scale-angle representation will be particularly interesting whenever scaling behavior (as in fractals) or angular selection is important, in particular, when directional wavelets are used. Clearly, these two representations are complementary, together they provide the full information contained in the signal. Accordingly, both are needed for a full understanding of the properties of the CWT in all four variables, as demonstrated in [13]. In addition to these two familiar representations, there are four other two-dimensional α, a, θ), and analyzing the sections, obtained by fixing two of the four variables (|b|, CWT as a function of the remaining two. and anisotropy angle (3) The scale-perception angle representation: for fixed range |b| θ , one obtains an analysis at all scales a and all perception angles α. (4) The range-anisotropy angle representation: one fixes the scale a and the perception and all anisotropy angles θ. angle α. This gives an analysis at all ranges |b| (5) The scale-range representation: fixing the perception angle α and the anisotropy angle θ gives an analysis at all scales a and all ranges |b|. and the (6) The angle-angle representation: on the contrary, if one fixes the range |b| scale a, one gets an analysis at all perception angles α and all anisotropy angles θ . This case is particularly interesting, because the parameter space is now compact (it is a torus) and the discretization easy (linear) in both variables. This representation will be used in Section 3.4, for illustrating the difference in angular selectivity between two standard wavelets. For the numerical evaluation, in particular for exploiting the reconstruction formula (2.26), one has to discretize the CWT. In any of these representations, a systematic use of the FFT algorithm will lead to a numerical complexity of 3N1 N2 log2 (N1 N2 ), where N1 , N2 denote the number of sampling points in the two free variables. In the case of the position representation, where (bx , b y ) are free, the geometry is Cartesian and a square lattice will give an adequate sampling grid. In the scale-angle representation, the CWT is naturally expressed in polar coordinates, like (a, θ ) or (a −1 , θ), and the discretization must be logarithmic in the scale variable a and linear in the anisotropy angle θ. For each variable, the size of the sampling mesh may be estimated from the support properties of the reproducing kernel K , which plays the rˆole of a correlation length. We shall come back to this discussion in Section 3.4 of Chapter 3 (see also [13] and [18]). Similar considerations apply to the remaining four representations. In addition, one may also consider three-dimensional sections, for which a single variable is fixed. Two of them look promising for applications.

48

The 2-D continuous wavelet transform

(1) The position-scale representation: suppose the anisotropy angle θ is fixed, or that it is irrelevant, because the wavelet is rotation invariant. Then the transform is a function of position b and scale a. This representation is optimal for detecting the presence of coherent structures, that is, structures that survive through a whole range of scales. Examples may be found, for instance, in astrophysics (hierarchical structure of galaxy clusters and superclusters) [343] or in the analysis of turbulence in fluid dynamics [164,165]. Further information on these two topics will be found in Chapter 5. (2) The position-anisotropy representation: here the scale a is fixed, and the transform is viewed as a function of position b and anisotropy angle θ. If the latter is plotted on the vertical axis of a three-dimensional graph, this means that the plane θ = θo selects all features (targets) that live in the corresponding line of sight. Similarly, an angular sector of opening θ is represented in such a plot by a horizontal slice of thickness θ . This visualization may offer distinct advantages over the conventional ones.

2.3.4

Partial energy densities of the CWT As explained in the previous section, the visualization problem of the CWT is solved by eliminating a certain number of variables, either by fixing their values, or by integrating them out. The principal example is the scale variable a. If the signal has significant features for a discrete set of scales only, {a j , j ∈ J }, the corresponding properties will be visible only if one chooses one of these values a j . Otherwise, nothing will be seen, and the transform is useless. A typical example is the problem of dilation symmetry in patterns, discussed in Section 4.5. In such a situation, clearly one should not fix the scale variable, but integrate over all scales (exactly as in the construction of continuous wavelet packets discussed in Section 2.6). Of course, the measure to use is the dilation invariant one, that is, the scale part a −3 da of the natural measure of the parameter space G. Proceeding in this way with the squared modulus of the wavelet transform, one obtains a quantity which has the physical meaning of a partial energy density. The same reasoning applies to any combination of “ignorable” variables. Thus, one gets such a partial energy density for each of the representations described in the previous section. The most important ones, of course, are the following. (1) Position (or range and aspect) energy density ao , θo ) is In the position representation, a = ao and θo are fixed and the CWT S(b, considered as a function of position b alone, either in Cartesian coordinates bx , b y , α (range and aspect). Accordingly, if one integrates the or in polar coordinates |b|, a, θ)|2 over all scales and orientations, one obtains phase space energy density |S(b, the position energy density, ∞ da 2π a, θ )|2 , dθ |S(b, (2.56) P[s](b) = a3 0 0

49

2.3 Implementation and interpretation of the 2-D CWT

either in Cartesian coordinates (position) or in polar coordinates (range and aspect). This density has been used as the basis of a CWT-based algorithm for automatic detection and recognition of targets (ATR) in forward-looking infrared radar (FLIR) imagery [285]. This application will be discussed in Section 4.2.2. (2) Scale-angle energy density In the scale-angle representation, the CWT is looked at from a fixed position bo as a function of scale a and anisotropy angle θ , i.e., of spatial frequency. In the phase space language of Section 2.3.2, this means considering | S( qo , p)|2 as a function of p alone, for fixed qo . The corresponding partial energy density is obtained by integrating over all positions b or q: a, θ)|2 . d 2 b |S(b, (2.57) M[s](a, θ ) = R2

This energy density, called the scale-angle measure or, better, the scale-angle spectrum of the signal, may be used, for example, for discriminating objects of interest according to their size and orientation [16], in particular target classification in FLIR imagery [287]. It yields also an efficient technique for detecting symmetries, even local ones, in patterns such as quasicrystals or Penrose tilings, the rationale being that such objects have no exact translation invariance, so that any dependence on position variables must be eliminated [24]. Both of these applications will be discussed in Chapter 4. A related concept, introduced in [249], is the relative scale-angle spectrum of the signal, obtained by normalizing M[s] over all angles: Z[s](a, θ ) = 2π 0

M[s](a, θ) dθ M[s](a, θ)

.

(2.58)

Whereas M[s] gives the distribution of energy at different scales and directions, Z[s] gives the relative distribution of energy at different directions at a particular scale with respect to the total energy at that scale. It turns out that Z[s] reveals more efficiently the scale-space anisotropic behavior of the signal. Similar partial energy densities may be introduced for the other representations, corresponding to other choices of “ignorable” variables. For instance, one may write the anisotropy angle and aspect energy density as ∞ da ∞ A[s](α, θ ) = | p | d| p | |S(| p |, α, a, θ )|2 . (2.59) 3 a 0 0 Altogether, there are four one-dimensional partial energy densities, six two-dimensional and M[s](a, θ ), and four three-dimensional ones. Besides the two main ones P[s](b) only two have found an application so far. One is the angular spectrum (called angular measure in [24]): ∞ da α[s](θ ) = M[s](a, θ), (2.60) a3 0

50

The 2-D continuous wavelet transform

which is used for detecting discrete rotation invariances of patterns (see Section 4.5). The other one is the wavelet spectrum, or scale spectrum, a function of scale only, obtained by integrating the scale-angle spectrum M[s](a, θ ) over all angles, or, more often, taking the latter in a rotation-invariant situation: a)|2 . W[s](a) = d 2 b |S(b, (2.61) R2

This is the proper generalization to the wavelet setup of the familiar Fourier power spectrum. It has been used, under various names (wavelet spectrum, wavelet (auto-)power spectrum, wavelet variance, scalogram), by many authors both in 1-D [211,227,273] and in 2-D [175].

2.3.5

Ridges in the 2-D CWT As we saw in Chapter 1 for the 1-D case, the reproduction property (2.36) means that the a, θ) is highly redundant. As a consequence, we information contained in the WT S(b, might hope that no content will be lost if we restrict the WT to a subset of the parameter space. As in 1-D again, there are two ways of achieving this. The first possibility is to take as determining subset the regions where the energy of the signal is concentrated, that is, essentially the lines of local maxima or ridges, the set of all ridges being called again the skeleton of the WT. This we will do in the present section. The alternative is to choose a discrete subset of the parameter space (a lattice), and this leads to the theory of frames that we shall discuss in detail in Section 2.4. When trying to extend the notion of ridge to 2-D signals, one faces again the two extreme situations described for the 1-D case in Section 1.4. We begin with the vertical a, θ ). The square modulus ridges. Let s( x ) be a 2-D signal (an image), with CWT S(b, of the latter is to be interpreted, as we have seen already, as the energy density of the signal, that we shall denote a, θ) = |S(b, a, θ)|2 . E[s](b,

(2.62)

In the case of a rotation-invariant wavelet, the θ-dependence drops out, and we write a) = |S(b, a)|2 . This is the case we will meet in applications, in Chapter simply E[s](b, 5, for instance, in the analysis of astrophysical images (Section 5.1). a) [272]. In this situation, we define the ridges as the lines of local maxima of E[s](b, More precisely, we will define a (vertical) ridge R as a 3-D curve ( r (a), a) such that, + for each scale a ∈ R , E[s]( r (a), a) is locally maximum in space and r is a continuous function of scale. Figure 2.2 gives a concrete example. The signal (left panel) is a set of singularities in a smooth background, simulating a set of bright points on the surface of the Sun and modeled by a random distribution of Gaussians of small (but random) width. The corresponding vertical ridges of the CWT of that signal are shown on the right panel, clearly each ridge points towards a singularity.

51

2.3 Implementation and interpretation of the 2-D CWT

Fig. 2.2. An example of a 2-D ridge: (left) the signal: a field of singularities, simulating a set of bright points on the surface of the Sun; (right) the corresponding vertical ridges of the CWT of that signal.

Given such a vertical ridge, one may distinguish three characteristic features. The first one is the amplitude of the ridge, that is, the value of E[s] on the ridge when a tends to zero, AR = lim E[s]( r (a), a).

(2.63)

a→0

The second one is the slope order, or slope, of E[s] on the ridge when a is close to 0: d ln E[s]( r (a), a) . a→0 d ln a

SR = lim

(2.64)

The last feature is the ridge energy, that is, the integral of E[s] along the ridge, assuming the latter to have a finite length, corresponding to the scale interval [0, amax ]: amax da E[s]( r (a), a). (2.65) ER = a3 0 Here the measure da/a 3 follows from the L 2 normalization of the wavelet. If one uses the L 1 normalization (see Section 2.6.1), one gets instead da/a. As in 1-D, one can show that the restriction of the CWT to its skeleton characterizes the signal completely [265]. A more sophisticated definition of ridge has been introduced by Mallat and Hwang [262], extending to 2-D the WTMM representation described in Section 1.4. The idea is to consider as (directional) wavelets the partial derivatives of a smoothing function ζ ( x ), a Gaussian, for instance: ψ1 ( x ) = ∂x ζ ( x ),

ψ2 ( x ) = ∂ y ζ ( x ),

x ≡ (x, y).

52

The 2-D continuous wavelet transform

Then, given any function s ∈ L 2 (R2 ), its WT with respect to ψ1 and ψ2 can be expressed as a vector b, a) = ∇(S a)), ζ (b, S( where a) = a −2 Sζ (b,

R2

(2.66)

s( d 2 x ζ (a −1 ( x − b)) x)

(using the L 1 -normalization). At a given scale a, the WTMM are defined by the positions b, a)| is locally maximum in the direction of the gradient b where the WT modulus | S( b, a). It turns out that the WTMM lie on connected chains in the b plane. vector S( One then defines the WTMM maxima (WTMMM) as the local maxima of the modulus b, a)| along the WTMM chains. Globally, these WTMMM live along connected | S( lines across scales, that is, the (vertical) ridges. We will describe in Section 5.4.1 the application of this new concept of ridge to the analysis of fractal surfaces. Here again, one can reconstruct from the ridges a good approximation of the original image [262]. An alternate possibility is to introduce horizontal ridges, as in [21]. The algorithm applies to signals which are superpositions of terms of the type s( x ) = A( x ) eiφ(x ) ,

(2.67)

where the amplitude A( x ) varies slowly with respect to the phase φ( x ). A first approximation of the CWT of this signal with the wavelet ψ is obtained by a Taylor expansion, which yields

−θ (∇φ( + R(b, a, θ), a, θ ) = A(b) eiφ(b) ψ(ar b))) S(b,

(2.68)

where the last term is a remainder that can be estimated. Assuming that the wavelet ψ has a unique maximum at a given frequency k0 , one sees that the CWT is concentrated along a surface in the parameter space, the corresponding ridge (more precisely, the 2-D ridge is a vector field, as we shall see below). A mathematically more precise approximation may be obtained by a stationary phase argument, following [195,196]. Take again the signal (2.67) and write the wavelet as ψ( x ) = Aψ ( x ) eiφψ (x ) . Then the CWT (2.19) of the signal s reads −1 ( x) ei b,a,θ d 2 x A( x ) Aψ (a −1 r−θ ( x − b)) , S(b, a, θ ) = a

(2.69)

R2

where ( x ) = φ( x ) − φψ (a −1 r−θ ( x − b)). b,a,θ

(2.70)

53

2.3 Implementation and interpretation of the 2-D CWT

This is an oscillatory integral and, therefore, the main contribution comes from the a, θ) such that stationary points xs of the phase b,a,θ , that is, the points xs = xs (b, ( xs ) = 0, or ∇ b,a,θ xs ). ψ (a −1 r−θ ( xs ) = a −1rθ (∇φ x − b)))( ∇φ(

(2.71)

to second order then yields Expanding the phase b,a,θ C(b, a, θ)−1 + R(b, a, θ ), a, θ ) = s( xs − b)) S(b, xs ) ψ(a −1 r−θ (

(2.72)

a, θ) is a correction term depending on the phase and R(b, a, θ) is again where C(b, a remainder that can be estimated. Because of the support properties of the wavelet ψ, a, θ), if we see from (2.72) that the CWT is essentially localized around the points (b, a, θ) = b. The set of these points is, by definition, the (horizontal) any, such that xs (b, ridge. Thus, on the latter, we have = a −1rθ (∇φ ψ (0)). b) ∇φ(

(2.73)

Writing this relation as ψ (0)), x ) = ar ( x )−1rθr (x ) (∇φ ∇φ(

(2.74)

of polar coordinates we see that the ridge takes the form of a vector field kr = kr (b) −1 (ar (b) , θr (b)), to be interpreted as a local wave vector. We call again skeleton of the WT the restriction of the latter to the ridge. Moreover, a, θ ) to the ridge it may be shown [196] that the restriction of the correction term C(b, is completely determined by the ridge itself, so that the skeleton of the WT reads as ar (b), θr (b)) = Ss (b,

2π ψ(0) s(b). ar (b), θr (b)) C(b,

(2.75)

It follows that the knowledge of the ridge and the skeleton of the WT is sufficient to characterize the signal s( x ). The discussion extends to a multicomponent signal, i.e., a linear superposition of N terms of the form (2.67). Assuming that the wavelet is sufficiently well localized to prevent any overlap, it will see the ridge corresponding to each term separately, and the skeleton will simply consist of N separate ridges, from which the N components can be extracted and reconstructed individually. Explicit examples may be found in [196]. In general, however, ridges may interact, but this case is much more difficult to handle. Finally, additional information may be obtained, as in 1-D, from the length of the various ridges (short vertical ridges tend to come from noise) and the behavior of the modulus of the CWT along each ridge, as a function of scale. Techniques based on 2-D ridges have been exploited in the problem of texture determination and, in particular, in the “shape from texture” problem, that we will discuss briefly in Section 5.5.

54

The 2-D continuous wavelet transform

2.4

Discretization, frames

2.4.1

Generalities on frames As we saw in Chapter 1 for the 1-D case, the reproduction property (2.36) means that the a, θ) is highly redundant. As a consequence, we information contained in the WT S(b, might hope that no content will be lost if we restrict the WT to a subset of the parameter space, in particular, a discrete subset (for instance, a lattice). Then the integral is replaced by a sum over a discrete (but infinite) family of wavelets ψbm ,a j ,θl : s( x) =

ψbm ,a j ,θl ( x ) S(bm , a j , θl ).

(2.76)

m jl

Here too, and by the same reasoning, one is thus led to the introduction of frames. Whereas we have barely sketched this topic in Chapter 1, we will now give a fairly detailed treatment. Further information (albeit mostly in 1-D) may be found in [121, 122,Dau92]. Let us start with a precise definition. According to the terminology introduced by Duffin and Schaefer [156] in the context of nonharmonic Fourier series, one has: Definition 2.4.1 . A countable family of vectors {ψn } in a Hilbert space H is called a (discrete) frame if there are two positive constants A, B, with 0 < A B < ∞, such that A f 2

∞

|ψn | f |2 B f 2 , ∀ f ∈ H.

(2.77)

n=1

The two constants A, B are the frame bounds. If A = B > 1, the frame is said to be tight. If A = B = 1, and ψn = 1, ∀ n, the set {ψn } is simply an orthonormal basis. The properties of a frame are best discussed in terms of the frame operator F : H → , defined by 2

F : f → {ψn | f }. As discussed in Section 1.3, the upper bound in (2.77) simply means that F is a bounded operator, whereas the left inequality guarantees the numerical stability for the recovery of the signal f from its frame coefficients {ψn | f } – in other words, it gives an estimation of the inverse operator F −1 . As for the frame bounds A, B, they measure the redundancy of the representation of the signal in terms of its coefficients. For a tight frame, in particular, A = B > 1 means that the frame is redundant, and B is its index of redundancy.

55

2.4 Discretization, frames

In terms of the frame operator F, the inequalities (2.77) may be written as A I F ∗ F B I,

(2.78)

where I is the identity operator. This in turn implies that F ∗ F is invertible and B −1 I (F ∗ F)−1 A−1 I.

(2.79)

Define now, for each n ∈ N : n = (F ∗ F)−1 ψn , ψ

(2.80)

n . Then the following is true: so that ψn = F ∗ F ψ n } constitute a frame, with frame bounds B −1 , A−1 and Theorem 2.4.2 . The vectors {ψ = F(F ∗ F)−1 . In addition, the expansion frame operator F f (x) =

∞

n (x), ψn | f ψ

(2.81)

n=1

∗ F = I . converges strongly in H, that is, F ≡ {ψ n } is the frame described in the statement results from the equalities Proof. That F n | f |2 = |ψ |(F ∗ F)−1 ψn | f |2 n

n

=

|ψn |(F ∗ F)−1 f |2

n

= F(F ∗ F)−1 f 2 = f |(F ∗ F)−1 f and the inequalities (2.77). Furthermore ∗ F = (F ∗ F)−1 F ∗ F = I, F that is, (2.81) is an identity.

= ∗ F = F ∗ F In other words, the duality between the two frames may be written as F I or explicitly, n | = n ψn | = I. |ψn ψ |ψ (2.82) n

n

= {ψ n } is called the dual or reciprocal frame of F = {ψn }. Notice that the The frame F dual of a tight frame is again a tight frame. This notion is crucial for applications. In the case of wavelet expansions, it is the basis of the so-called biorthogonal scheme [Dau92], briefly discussed in Section 2.5.2. In practice, orthonormal bases are not always available for representing arbitrary functions, but one may often use instead a good frame. By this, we mean that the expansion (2.81) converges sufficiently fast. How could one estimate the speed of this

56

The 2-D continuous wavelet transform

n = (F ∗ F)−1 ψn . If B and A are close convergence? By (2.81), we need to compute ψ 1 2 to each other, F ∗ F is close to 2 (B + A)I , hence (F ∗ F)−1 is close to B+A I and thus n is close to 2 ψn . Hence we may write ψ B+A f =

2 ψn | f ψn + R f, B+A n

(2.83)

where R=I−

2 F ∗ F. B+A

(2.84)

Hence 2 (I − R)−1 B+A ∞ 2 = Rk . B + A k=0

(F ∗ F)−1 =

(2.85)

The series converges in norm, since, by (2.84), −

B−A B−A I R I, B+A B+A

(2.86)

which implies R

B/A − 1 B−A = < 1. B+A B/A + 1

Therefore the expansion (2.81) converges essentially as a power series in |B/A − 1|. Thus the frame is good if |B/A − 1| 1, in particular if it is tight. To the first order, the expansion (2.81) becomes 2 f = ψn | f ψn . (2.87) B+A n The quantity w(F ) =

B−A B+A

(2.88)

is called the width or the snugness of the frame F. It measures the lack of tightness, since w(F ) = 0 iff the frame F is tight. Notice that a frame and its dual have the same width. In practical applications, the infinite sum in (2.81) or (2.87) will be truncated and the approximate reconstruction so obtained is numerically stable. If the width of the frame is sufficiently small, a few terms will suffice. More details on frames may be found in [121,122,Dau92,220,Mal99]. Let us now come back to the notion of redundancy of a frame, which may be characterized in terms of the frame operator F. Let Ran F ⊂ 2 denote the range of F, that is, the set of sequences F f = (ψn | f ), f ∈ H. First, we remark that the inclusion is

57

2.4 Discretization, frames

strict if the frame vectors {ψn } are linearly dependent. In that case, indeed, there exists

a nonzero vector y = (yn ) ∈ 2 such that n yn ψn = 0. But then, for any f ∈ H, one

has n yn ψn | f = y|F f = 0, that is, y ∈ (Ran F)⊥ = {0}, where (Ran F)⊥ is the orthogonal complement of Ran F in 2 . Moreover, the more redundant the frame, the larger the orthogonal complement (Ran F)⊥ . Since F is injective by the left inequality in (2.77), it may be inverted on its image, but the inverse operator is not uniquely defined, because it remains arbitrary on the complement (Ran F)⊥ . Among all possible inverses, the pseudo-inverse F˘ −1 is defined as the inverse operator that vanishes on (Ran F)⊥ . It is easy to show [Mal99] that it is also the inverse with the smallest norm, and it is given by F˘ −1 = (F ∗ F)−1 F ∗ (thus ∗ . Thus, it is uniquely defined). From the discussion above, it is clear that F˘ −1 ≡ F the pseudo-inverse of the frame operator is always bounded. This explains the rˆole of the lower bound in (2.77) as guaranteeing numerical stability in the computation of the inverse operator F −1 . In addition, we see that the dual frame is built from the pseudo-inverse. One should also notice that there exist alternative inversion techniques, for instance, the conjugate gradient method, that sometimes converge faster than the pure wavelet reconstruction formulas. For more information, see for instance [167]. Another useful notion (already met in the Definition 1.5.1 of a multiresolution analysis, in Section 1.5) is that of a Riesz basis, namely, a frame F of linearly independent vectors. Of course, this definition implies that the corresponding frame operator F maps is also a Riesz basis, and in H onto 2 , Ran F = 2 . It follows that the dual frame F fact, it is biorthogonal to F. To see this, apply (2.82) to a basis vector ψk : n |ψk . ψk = ψn ψ n

n |ψk = δn,k , i.e., Since the vectors {ψn } are linearly independent, this implies that ψ the two bases are biorthogonal. Next, if the vectors {ψn } were not linearly independent,

n = 0. But then, taking the there would exist nonzero numbers {λn } such that n λn ψ inner product with any ψk gives n |ψk = λk , 0= λn ψ n

a contradiction. Further information about Riesz bases and their numerical implementation may be found in [Mal99].

2.4.2

Two-dimensional wavelet frames Thus there remains the question: given a specific wavelet, does it generate a frame? The first problem is how to choose the sampling grid # in an optimal fashion. As in 1-D, one should take into account the geometry of the parameter space, that is, the lattice # must be invariant under discrete sets of dilations, rotations and translations.

58

The 2-D continuous wavelet transform

Note, however, that in practice the sampling points are quite often fixed empirically. For the (a, θ) variables, in particular, they are mostly chosen on the basis of biological considerations or symmetry requirements [126,260,264]. Proceeding as in 1-D, one thus obtains the following natural discretization scheme. r For the dilations, a logarithmic scale a = a λ− j , j ∈ Z, for some λ > 1; here again j 0 we will put a0 = 1. r For the rotations, one subdivides the interval [0, 2π ) uniformly into L pieces, for 0 some natural number L 0 ∈ N, that is, θl = lθ0 , θ0 = 2π , l ∈ Z = {0, ..., L 0 − 1}. L 0 L0 r For the translations, one takes into account the two previous discretizations, putting u m 0 m 1 ), bm ≡ b jlm 0 m 1 = λ− j rlθ0 ( with um 0 m 1 ≡ (m 0 β0 , m 1 β1 ), m 0 , m 1 ∈ Z, β0 , β1 0. Thus the discretization grid reads:

−j

# = #(λ, L 0 , β0 , β1 ) = (λ

2π 2 , l , b jlm 0 m 1 ), ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z . L0 (2.89)

The resulting discretized wavelet transform, which is a map from L 2 (R2 , d 2 x) to l 2 (Z × Z L 0 × Z2 ), reads now: S jlm 0 m 1 ≡ S(bm , λ− j , lθ0 ) = ψ jlm 0 m 1 |s = λj d 2 x ψ(λ j r−lθ0 ( x ) − um 0 ,m 1 ) f ( x) R2 −j d 2 k ei bm0 m1 ·k ψ(λ r−lθ0 (k)) f (k), = λ− j

(2.90) (2.91)

R2

with wavelet coefficients

S jlm 0 m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z2 .

Our task now is to find conditions on the grid #(λ, L 0 , β0 , β1 ), that is, on the parameters λ, L 0 , β0 , β1 , such that the family of wavelets {ψ jlm 0 m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z2 } is a frame. As in 1-D, the answer is that the 2-D wavelet transform obeys a sampling theorem, that gives a lower bound on the density of sampling points, like the standard Shannon theorem of signal analysis. The following result, first proven in [Mur90], is the exact counterpart of [122; Theorem 2.7], and the proof follows closely [Dau92; Section 3.3.2].

59

2.4 Discretization, frames

Theorem 2.4.3 . Assume the wavelet ψ satisfies the following conditions: L0 ∞

(i) s(λ, L 0 , ψ) = ess inf R2 k∈

=

− j r−lθ0 (k))| 2 |ψ(λ

(2.92)

j=−∞ l=0

ess inf

L0 ∞

)∈[0,λ)×[0,2π) (|k|,θ

p (λ− j |k|, ϕ − lθ 0 )|2 > 0, |ψ

j=−∞ l=0

(2.93) p is the Fourier transform of ψ in polar coordi where k = |k|(cos ϕ, sin ϕ) and ψ nates. (ii) S(λ, L 0 , ψ) = sup

R2 k∈

=

L0 ∞

− j r−lθ0 (k))| 2 |ψ(λ

(2.94)

j=−∞ l=0

sup

L0 ∞

)∈[0,λ)×[0,2π) j=−∞ l=0 (|k|,θ

p (λ− j |k|, ϕ − lθ0 )|2 < ∞. |ψ (2.95)

u |)1+$ α( u ) < ∞, (iii) sup (1 + |

(2.96)

u∈R2

where $ > 0 and α( u ) = sup

L0 ∞

R2 j=−∞ l=0 k∈

− j r−lθ0 (k) − j r−lθ0 (k))|. + u)||ψ(λ |ψ(λ

(2.97)

Then there exist constants β0 c , β1 c > 0 such that: (1) ∀ β0 ∈ (0, β0 c ) , β1 ∈ (0, β1 c ), the family {ψl jm 0 m 1 } associated to (λ, L 0 , β0 , β1 ) is a frame of L 2 (R2 , d 2 x); (2) ∀ δ > 0, there exist β0 ∈ (β0 c , β0 c + δ) , β1 ∈ (β1 c , β1 c + δ) , such that the family {ψl jm 0 m 1 } associated to (λ, L 0 , β0 , β1 ) is not a frame of L 2 (R2 , d 2 x).

Proof . We want to find the conditions on λ, L 0 , β0 , β1 for which there exists 0 < A, B < ∞, such that:

A f 2

L0

∞

l=0 j,m 0 ,m 1 =−∞

|ψ jlm 0 m 1 | f |2 B f 2 .

(2.98)

60

The 2-D continuous wavelet transform

The central term in these inequalities reads K =

|ψ jlm 0 m 1 | f |2

j,l m 0 ,m 1

=

λ

−2 j

2

R2

j,l m 0 ,m 1

d 2 k ei b jlm0 m1 ·(k−k )

d k

R2

− j r−lθ0 (k)) − j r−lθ0 (k )) ψ(λ f (k ) × ψ(λ f (k) =

λ

−2 j

2

d 2 k ei um0 m1 ·λ

−j

d k R2

j,l m 0 ,m 1

k ) r−lθ0 (k−

R2

− j r−lθ0 (k)) − j r−lθ0 (k )) ψ(λ f (k ) × ψ(λ f (k) =

λ

−2 j

2

d 2 k ei um0 m1 ·(k−k )

d k R2

j,l m 0 ,m 1

R2

k) k ) ψ( f (λ j rlθ0 (k )) × ψ( f (λ− j rlθ0 (k)). Using the Poisson formula, ∞

ei um0 m1 ·(k−k ) =

m 0 ,m 1 =−∞

4π 2 β0 β1

∞

δ(k − k − u m 0 m 1 ),

(2.99)

m 0 ,m 1 =−∞

where 2π 2π u m 0 m 1 = (m 0 , m 1 ), β0 β1

(2.100)

we obtain K =

4π 2 k) k − ψ( d 2 k ψ( u m 0 m 1 ) β0 β1 j,l m 0 ,m 1 R2 f (λ j rlθ0 (k − u m 0 m 1 )) × f (λ j rlθ0 (k))

4π 2 − j r−lθ0 (k) − j r−lθ0 (k)) ψ(λ − d 2 k ψ(λ u m 0 m 1 ) = β0 β1 j,l m 0 ,m 1 R2 f (k − λ j rlθ0 ( u m 0 m 1 )). × f (k)

61

2.4 Discretization, frames

Let us split the double sum as K = |P| + Q, where |P| denotes the term with (m 0 , m 1 ) = (0, 0) and Q the rest: K = |P| + Q 4π 2 − j r−lθ0 (k))| 2 | 2 = d 2 k |ψ(λ f (k)| β0 β1 j,l R2 4π 2 − j r−lθ0 (k)) − j r−lθ0 (k) ψ(λ − + d 2 k ψ(λ u m 0 m 1 ) β0 β1 j,l m 0 ,m 1 ∈Z∗ R2 f (k) f (k − λ j rlθ0 ( u m 0 m 1 )). × Then we obtain the following estimates. (1) For the first term, we get immediately: 4π 2 4π 2 2 2 s(λ, L 0 , ψ) f |P| S(λ, L 0 , ψ) f , β0 β1 β0 β1 where s(λ, L 0 , ψ) and S(λ, L 0 , ψ) are defined in (2.92) and (2.94), respectively. (2) For the second term, we obtain, by the Cauchy–Schwarz inequality: 4π 2 − j r−lθ0 (k))| |Q| d 2 k |ψ(λ β0 β1 j,l m 0 ,m 1 ∈Z∗ R2 − j r−lθ0 (k) − | u m 0 m 1 )| | f (k)| f (k − λ j rlθ0 ( u m 0 m 1 ))| × |ψ(λ 4π 2 β0 β1 j,l m 0 ,m 1 ∈Z∗ !1/2 − j r−lθ0 (k) − j r−lθ0 (k))| − 2 d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ | f (k)| R2 − j r−lθ0 (k) − j r−lθ0 (k))| − × d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ R2

| f (k − λ j rlθ0 ( u m 0 m 1 ))|2

1/2

4π 2 β0 β1 j,l m 0 ,m 1 ∈Z∗ !1/2 − j r−lθ0 (k) − j r−lθ0 (k))| − | 2 d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ f (k)| ×

R2

− j r−lθ0 (k))| − j r−lθ0 (k) |ψ(λ + 2 d k |ψ(λ u m 0 m 1 )| | f (k)| 2

R2

!1/2 .

62

The 2-D continuous wavelet transform

Applying Cauchy–Schwarz a second time, to the summation over j, l, then gives: |Q|

4π 2 β0 β1

" " ×

d 2 k

R2

m 0 ,m 1 ∈Z∗

#1/2

− j r−lθ0 (k) − j r−lθ0 (k))| − 2 |ψ(λ u m 0 m 1 )| |ψ(λ | f (k)|

j,l 2

d k R2

|ψ(λ

−j

|ψ(λ r−lθ0 (k))|

−j

#1/2

+ r−lθ0 (k) u m 0 m 1 )| | f (k)|

2

.

j,l

The terms in braces in the integrals are majorized using (2.97), and we get finally 4π 2 2 α( u m 0 m 1 ) α(− um0m1 ) f . (2.101) |Q| β0 β1 m 0 ,m 1 ∈Z∗ Define the quantity E(λ, L 0 , β0 , β1 , ψ) =

α( u m 0 m 1 ) α(− u m 0 m 1 ) =

m 0 ,m 1 ∈Z∗

%2 α( u m 0 m 1 ) .

$

(2.102)

m 0 ,m 1 ∈Z∗

In virtue of the decay condition (2.96) in assumption (iii), the sum over (m 0 , m 1 ) converges. Then, using the inequality |P| − |Q| K ≡ |P| + Q |P| + |Q|, we obtain a lower bound for the left-hand side: ' 4π 2 & 2 f |P| − |Q| (2.103) s(λ, L 0 , ψ) − E(λ, L 0 , β0 , β1 , ψ) β0 β1 and an upper bound for the right-hand side: ' 4π 2 & 2 f . S(λ, L 0 , ψ) + E(λ, L 0 , β0 , β1 , ψ) |P| + |Q| β0 β1

(2.104)

By condition (2.92) in assumption (i), s(λ, L 0 , ψ) is strictly positive, and the decay condition (2.96) implies that lim

(β0 ,β1 )→(0,0)

E(λ, L 0 , β0 , β1 , ψ) = 0.

(2.105)

Therefore, there exists critical values β0 c , β1 c > 0 such that the first term in (2.103) is dominant for all β0 < β0 c , β1 < β1 c . Thus the left-hand side of (2.103) is strictly positive and yields a lower frame bound. This proves the statement. Notice that, as in [Dau92; Section 3.3.2], we had to use in (2.92) the essential infimum, that is, infimum except on a set of measure zero, instead of the usual infimum, which 0) = 0. This distinction is not necessary in (2.94), because in all is 0 here, since ψ( is continuous. Here also, the infimum in (2.93) has to be practical cases, the function ψ can be brought to this range taken only over a ball of radius λ, since other values of |k| j by dilation with a suitable power λ , except for k = 0, which is a set of measure zero.

63

2.4 Discretization, frames

From the inequalities (2.103) and (2.104), we obtain immediately estimates for the frame bounds: Corollary 2.4.4 . Whenever the conditions of Theorem 2.4.3 are satisfied, and β0 , β1 are such that s(λ, L 0 , ψ) > E(λ, L 0 , β0 , β1 , ψ),

(2.106)

then the bounds of the frame can be estimated by: 4π 2 {s(λ, L 0 , ψ) − E(λ, L 0 , β0 , β1 , ψ)}, β0 β1 4π 2 {S(λ, L 0 , ψ) + E(λ, L 0 , β0 , β1 , ψ)}. B β0 β1 A

(2.107) (2.108)

It is clear from this discussion that we have obtained essentially the same results as in the 1-D case – and they extend to three or more dimensions in the same way [283,Mur90]. As argued in [Dau92; Section 3.3.2], the moral is that, whenever the wavelet ψ is admissible and decays reasonably at infinity, it will yield a frame. This is true, for instance, for all types of Mexican hat or Morlet wavelets. However, the estimates of Corollary 2.4.4 imply, as in 1-D [103], the following inequalities as well, for all k = 0, A

L0 ∞ 4π 2 − j r−lθ0 (k))| 2 B. |ψ(λ β0 β1 j=−∞ l=0

(2.109)

This puts a rather strong limitation on the wavelet, because the frame tends to become loose when λ is not very small, in particular for λ = 2, the preferred value. In 1-D, this defect may be corrected by introducing additional voices. This means that one further subdivides each octave, replacing in (2.109) the exponent j by ν j, ν = 0, 1, . . . , N − 1 (N is called the number of voices). The effect is to “densify” N the lattice #, which improves the ration B/A and thus speeds up convergence of the discrete approximation. For further details, we refer the reader to [Dau92]. The same procedure may be applied in 2-D as well. However, if the speed is the determining criterion, one can do better by using the pseudo-QMF algorithm described in Section 2.6. In addition, other discretization schemes may be considered. For instance, the discretization step in the angular variable θ may be made scale dependent. The idea is that one needs more directions at small scales (high frequencies) than at large scales (low frequencies). This is in the spirit of certain applications of discrete wavelet bases in solid state physics (see [30]), in which additional, smaller, scales are considered in the vicinity of the atomic crystal nodes. We will come back to this point in the next section.

64

The 2-D continuous wavelet transform

2.4.3

Implementation of a 2-D wavelet frame In general, implementing the inverse frame operator is a nontrivial task which requires using the expansion formula (2.84). In the case of a tight (or approximately tight) frame, however, the situation becomes much easier since the reconstruction is provided by the simple sum (2.87). Let us particularize this for the concrete case of the 2-D Morlet wavelet. Using Lee’s parametrization [253], the Fourier transform of this wavelet is simply: κ2 κ2 2 2 2 √ − 2 ((k1 −k0 )2 +4k22 ) κ 2 (k1 +4k2 +k0 ) 2k 2k 1 , k2 ) = ψ(k 8π e 0 −e 0 . (2.110) k0 √ Chosing κ = 3 2 log 2 yields a spatial frequency bandwidth of approximately one octave, which is compatible with the characteristics of the receptive fields of human simple cells [372]. Using the general framework described before, one is able to compute frame bounds for this wavelet. Psychovisual experiments suggesting a density of about 16 to 20 orientations, one can obtain a tight frame by carefully choosing a discretization strategy. Figure 2.3 shows a reconstructed image with a tight frame of Morlet wavelets corresponding to the discretization grid 2π # = 2− j/n , l , b j,l,m 0 ,m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z16 × Z2 , n = 2 and 4, 16 that is, 2 and 4 voices, respectively, per octave and a uniform translation step β0 = β1 = 0.8.

2.4.4

The dyadic wavelet transform As we saw already in 1-D (Sections 1.3 and 1.6.1), a slightly different construction consists in computing a hybrid wavelet transform in which we sample the scale variable, but leave translations untouched. If we use dyadic scales, the resulting set of coefficients is called the Dyadic Wavelet Transform and offers the advantage of being completely covariant with respect to translations, a very desirable feature [264,265]. In pattern recognition, for instance, translating the object as a whole should affect all wavelet coefficients in the same way, so as not to distort the transform. This is precisely the practical meaning of the so-called “shift invariance” (properly, shift covariance), and the condition is not satisfied in the usual discrete WT. In addition, when the translation parameter is properly discretized, such dyadic wavelet transforms may lead to genuine tight frames. We denote, as usual, the scaled and translated wavelets by , ψb,2 x ) = 2− j ψ (2− j ( x − b)) j (

(2.111)

assume for simplicity that cψ = 1 and use the L 2 -normalization. Then the dyadic WT of f ∈ L 2 is given by

65

2.4 Discretization, frames

(a)

(b)

(c)

Fig. 2.3. Reconstruction of the lena image with a tight frame of Morlet wavelets: (a) the original

image; (b) and (c) the reconstructed image, with 2 and 4 voices, respectively, per octave.

2 j ) = ψ j | f = ψ j# % f , W (b, 2 b,2

(2.112)

x ) = ψa (− x ) and ψa ≡ ψ0,a where ψa# ( . Note that we use a different notation, W(., .) instead of W (., .), in order to emphasize the fact that we are dealing with the dyadic WT instead of the CWT. The superscript in equations (2.111), (2.112) refers to different wavelets, but the practical situation will be that of rotated versions of a single generating function. The following result shows that the dyadic WT is a stable and complete representation of images provided the set of scaled wavelets appropriately pave the frequency plane.

66

The 2-D continuous wavelet transform

Proposition 2.4.1 If there exists two strictly positive constants A and B such that A

L

(2 j k)| 2 B, |ψ

∀ k,

(2.113)

=1 j∈Z

then, A f 2

L

2−2 j W (., 2 j )2 B f 2 .

(2.114)

=1 j∈Z

Moreover, if we define dual or reconstruction wavelets χ through the relation L

(2 j k) χ = 1, (2 j k) ψ

, ∀ k ∈ R2 \ {0}

(2.115)

=1 j∈Z

then, the following reconstruction formula holds strongly in L 2 : f ( x) =

L =1 j∈Z

2−2 j W (., 2 j ) % χ2j ( x ).

(2.116)

2 j ) with respect to the variable Proof . Let d j stand for the Fourier transform of W (b, b: (2 j k) = 2j ψ d j (k) f (k). Condition (2.113) yields: 2 A| f (k)|

L

2 B| 2. 2−2 j |d j (k)| f (k)|

=1 j∈Z

Integrating over k and using the Plancherel formula gives (2.114). Now taking the Fourier transform on both sides of (2.116) gives = f (k)

L

(2 j k) ψ χ f (k) (2 j k).

=1 j∈Z

Finally, applying condition (2.115) gives the final result.

The norm equivalence represented by (2.113) shows that this family of wavelets behaves exactly like a frame, with frame operator 2 j , ) = 2− j W (b, 2 j ). (F f )(b, In particular, if the sum standing in the middle of (2.113) is a constant, we can take A = B and we get the analogue of a tight frame.

67

2.4 Discretization, frames

Among all possible inverses of the frame operator F, the particular choice = ψ (k) χ (k) j 2 j∈Z |ψ (2 k)| leads to an attractive solution. In this case, indeed, the inverse frame operator given by (2.116) is the pseudo-inverse of F as shown by the following result: Proposition 2.4.2 Let F be the normalized frame operator u , 2 j , ) = 2− j W ( u , 2 j ). [F f ] ( Then the following left inverse is the pseudo-inverse of F: ( FW u, 2 j ) =

L =1 j∈Z

u ). 2− j W (., 2 j ) % χ2j (

(2.117)

Proof . Let 2 (L 2 ) be the Hilbert space of square integrable sequences of functions d, j =1...L , j∈Z , d, j ∈ L 2 and d, j 2 < +∞.

j∈Z

We are going to show that, ∀ g ∈ L 2 and ∀ d ∈ Ran{F}⊥ , d = 0. g, F

(2.118)

Let d = Fg and d ∈ Ran{F}⊥ . Using (2.117), we compute ( j (k) d = d ψ (2 k) g (k) d 2 k . g, F , j ( j 2 2 j∈Z R j ∈Z |ψ (2 k)| Now, writing d, j = Fg, we also have ( (2 j k) ψ = d g (k) , j (k) , and using equation (2.113) we find 1 g 1 d, d 2 (L 2 ) g, F 2 (L 2 ) . d, d B A Now, since d ∈ Ran{F} and d ∈ Ran{F}⊥ , (2.119) yields the result.

(2.119)

Another interesting case arises when we define a new set of wavelets through the relation (k) ψ =) ϕ (k) ,

j 2 j |ψ (2 k)|

(2.120)

68

The 2-D continuous wavelet transform

and use the same set of functions to implement the reconstruction: f ( x) =

L =1 j∈Z

2−2 j W (., 2 j ) % ϕ2 j ( x ).

(2.121)

The reader can easily check that the wavelet transform operator is an isometry between L 2 and 2 (L 2 ). The range of this operator is a closed subspace V of 2 (L 2 ) characterized by a reproducing kernel, exactly as for the continuous transform. Indeed, any d, j ∈ V satisfies j, j d, j ( x) = d , j % K, ( x) , , j

where the reproducing kernel is given by

j, j x ). x) = ϕ2 j % ϕ2 j ( K, (

(2.122)

This is easily checked by writing explicitly % $ u ). u ) = F F −1 d , j ( d, j ( Taking the Fourier transform with respect to u on both sides and using (2.120) yields the result. This property is a sign of the redundancy of the representation, exactly as in the continuous case. Section 2.6 gives an efficient and automatic way for building such dyadic wavelet transforms starting from the continuous wavelet transform. This offers more flexibility in the design of directional dyadic wavelets.

2.5

Comparison with the 2-D discrete wavelet transform Before analyzing the recent approach to fast algorithms (Section 2.6), we have to sketch briefly the 2-D discrete wavelet transform, in its various forms and generalizations. As mentioned in Chapter 1, a key step in the success of the 1-D discrete WT was the discovery that almost all examples of orthonormal bases of wavelets may be derived from a multiresolution analysis, and furthermore that the whole construction may be translated into the language of digital filters. In the 2-D case, the situation is exactly the same, as we shall see in this section. Further information may be found in [Dau92] or [Mey94].

2.5.1

Multiresolution analysis in 2-D and the 2-D DWT The simplest approach consists in building a 2-D multiresolution analysis simply by taking the direct (tensor) product of two such structures in 1-D, one for the x direction, one for the y direction. If {V j , j ∈ Z} is a multiresolution analysis of L 2 (R), (2) then { V j = V j ⊗ V j , j ∈ Z} is a multiresolution analysis of L 2 (R2 ). Writing again

69

2.5 Comparison with the 2-D discrete wavelet transform (2)

(2)

(2)

V j ⊕ W j = V j+1 , it is easy to see that this 2-D analysis requires one scaling function: (x, y) = φ(x) φ(y), but three wavelets: h (x, y) = φ(x) ψ(y) v (x, y) = ψ(x) φ(y)

(2.123)

(x, y) = ψ(x) ψ(y). d

As the notation suggests, h detects preferentially horizontal edges, that is, discontinuities in the vertical direction, whereas v and d detect vertical and oblique edges, respectively. Indeed, for j = 1, the relation V1 = V0 ⊕ W0 yields: (2)

(y)

V1 = V1(x) ⊗ V1

(y)

(y)

= (V0(x) ⊕ W0(x) ) ⊗ (V0 ⊕ W0 ) (y)

(y)

(y)

(y)

= (V0(x) ⊗ V0 ) ⊕ (V0(x) ⊗ W0 ) ⊕ (W0(x) ⊗ V0 ) ⊕ (W0(x) ⊗ W0 ) (2)

(2)

= V0 ⊕ W0 , (y)

where V0 = V0(x) ⊗ V0 & φ(x) φ(y) and W0 is the direct sum of the three other products, generated by the three wavelets given in (2.123), respectively. (2) From these three wavelets, one gets an orthonormal basis of V j by defining (2) j { kl (x, y) = φ j,k (x) φ j,l (y), k, l ∈ Z}, and one for W j in the same way, namely α, j {kl (x, y), α = h, v, d and k, l ∈ Z}. Clearly this construction enforces a Cartesian geometry, with the horizontal and the vertical directions playing a preferential role. This is natural for certain types of images, such as in television, but is poorly adapted for detecting edges in arbitrary directions. Other solutions are possible, however (see below). As in the 1-D case, the implementation of this construction rests on a pyramidal algorithm introduced by Mallat [259,260]. The technique consists in translating the multiresolution structure into the language of QMFs, and putting suitable constraints on the filter coefficients. For instance, ψ has compact support if only finitely many coefficients differ from zero. In the 2-D case, obviously one gets a low-pass filter h and three high-pass filters g α , α = h, v, d. Thus, a signal f ∈ L 2 (R2 ) is represented at

resolution 2 j by the function f j = a j + α d αj , α = h, v, d, where (2) j cj = c j,kl kl ∈ V j (2)

(2)

k,l∈Z

d αj

=

α, j

α

d αj,kl kl ∈ W j , α = h, v, d. (2)

k,l∈Z

In this scheme, the decomposition of an image into a low resolution approximation, plus three types of details (h, v, d), at successive finer scales, takes the familiar form of nested boxes, with the low resolution part in the upper left corner. Figure 2.4 presents a schematic three-level decomposition of an image into a low resolution approximation, with coefficients c−3 , plus increasingly finer details, of the three types, with coefficients

70

The 2-D continuous wavelet transform

c3

d3vv

d3h

d3d

d2vv d1vv

h

d

d2

d2

d1h

d1d

Fig. 2.4. Schematic three-level decomposition of an image into a low resolution approximation,

plus increasingly finer details, of the three types (h, v, d).

d αj , j = 1, 2, 3, α = h, v, d. Then we show a (standard) real example in Figure 2.5, which is a three-level orthonormal basis decomposition, with a Daubechies wavelet (compact support, three vanishing moments).

2.5.2

Generalizations As in one dimension, the scheme based on orthonormal wavelet bases is too rigid for most applications and various generalizations have been proposed. We discuss some of them here, for two reasons. First, for the sake of completeness. But also in order to demonstrate that some features which are natural in the continuous transform, such as covariance, are not easy to enforce in the discrete case, however desirable they may be.

2.5.2.1

More isotropic 2-D wavelets The tensor product scheme privileges the horizontal and the vertical directions; more isotropic wavelets may be obtained, either by superposition of wavelets with specific orientation tuning [Mar82], as we did above with the CWT, or by choosing a different way of dilating, using a nondiagonal 2-D dilation matrix, which amounts to dilating by a noninteger factor [Dau92]. Consider, for instance, the following dilation matrices: 2 0 1 1 1 1 D0 = , D1 = , D2 = . (2.124) 0 2 1 −1 −1 1 The matrix D0 corresponds to the usual dilation scheme by powers of 2, whereas D1 and D2 lead to the so-called “quincunx” scheme [Fea90]. In the standard scheme, a unit square is dilated, in the transition j → j + 1, to another square, twice bigger, with the same orientation. This means that three kinds of additional details have to be supplied,

71

2.5 Comparison with the 2-D discrete wavelet transform

Fig. 2.5. Typical three-level decomposition of an image into a low resolution approximation, plus

increasingly finer details, of the three types (h, v, d). The basic wavelet is a Daubechies wavelet with compact support and three vanishing moments.

horizontal, vertical and oblique (see Figure 2.6, left). By contrast, the same operation in the “quincunx” scheme leads to a square circumscribed to the original one, that is, √ rotated by 45◦ and larger by a factor 2, so that only one kind of additional detail is necessary (Figure 2.6, right). Indeed only one wavelet is needed in this scheme, instead of three. This is consistent with a result of Meyer, according to which the number of independent wavelets needed in a given multiresolution scheme equals (| det D| − 1), where D is the dilation matrix used.

2.5.2.2

Biorthogonal wavelet bases In the case of the continuous transform, the wavelet used for reconstruction need not be the same as that used for decomposition, they have only to satisfy a cross-compatibility

72

The 2-D continuous wavelet transform

(v)

❅

(d)

❅

j+2

❅

❅

scale j

(h)

scale j

❅

❅

j+1❅

❅

❅

j+1

❅

❅ ❅

Fig. 2.6. Unit cell at successive resolutions: (left) for the “Cartesian” scheme; (right) for the

“quincunx” scheme.

condition, as shown in (2.30), and this has led us to a host of different, more flexible reconstruction formulas, such as (2.31)–(2.34). The same idea in the discrete case leads to biorthogonal bases [108], i.e., one has two hierarchies of approximation spaces, {V j } j }, with cross-orthogonality relations. and {V In 1-D, the construction goes as follows, and the extension to 2-D proceeds as above. Start with a scale of closed subspaces {V j }, assuming only the existence of a scaling function φ ∈ V0 such that its integer translates {φk (x) ≡ φ(x − k), k ∈ Z} is a Riesz (or unconditional) basis of V0 (see Section 1.5). Then, instead of orthogonalizing this basis, which would lead to the construction of an orthonormal wavelet basis, one takes 0 k }, that is, the vectors defined by the relation φk |φ l = δkl . Let V the dual basis {φ denote the closed subspace generated by {φk , k ∈ Z}. Then the same construction is repeated for each j, using the dilation invariance of the scale {V j }. The outcome is j }, with exactly the same properties. Next, for each a second multiresolution scale {V j , and j ∈ Z, one defines a subspace W j by the two conditions W j ⊂ V j+1 and W j ⊥ V j ⊂ V j+1 and W j ⊥ V j . In this way one obtains two sequences of subspaces similarly W j,k , j, k ∈ Z}, respectively, which are {W j } and {W j }, with bases {ψ j,k , j, k ∈ Z}, {ψ mutually orthogonal: j ,k = δ j j δkk . ψ j,k |ψ

(2.125)

In terms of these bases, one gets two types of expansion formulas, for any f ∈ L 2 (R): j,k | f ψ j,k f = ψ j,k∈ Z

=

j,k ψ j,k | f ψ

(2.126)

j,k∈ Z

in the case of a frame and its dual. Translating the whole structure into the language of digital filters, a 1-D biorthogonal scheme corresponds to four filters, (h, h, g, g ), where the first two are low-pass and the last two are high-pass. A corresponding characterization may be given in the 2-D case.

73

2.5 Comparison with the 2-D discrete wavelet transform

The resulting scheme is much more flexible and is probably the most efficient one in practical applications, both in 1-D and in 2-D (and it is widely used). For instance, it gives a better control on the regularity or decrease properties of the wavelets [108].

2.5.2.3

Wavelet packets and the Best Basis Algorithm Electrical engineers are familiar with the notion of subband coding scheme. In a few words, this means that a discrete signal {cn0 , n ∈ Z} is first subdivided into two subsignals {cn1 , dn1 , n ∈ Z}, obtained by convolving the original signal with two filters, one low-pass and one high-pass filter, respectively. Next, each subsignal is subsampled by a factor of 2, that is, one keeps only the even or odd components, respectively (thus the total number of coefficients is unchanged). Then the operation is iterated a number of times. A reconstruction of the original signal is obtained, more or less exactly, by inverting all the operations. In the construction of 1-D orthonormal wavelet bases, each approximation space V j gets further decomposed into V j−1 and W j−1 , whereas the detail space W j is left unmodified. Expanding the signals in the respective bases {φ j−1,k , ψ j−1,k } shows that the construction is in fact a subband coding scheme, but a very special one, rather asymmetrical. In order to get more flexibility, more general subband schemes have been considered, called wavelet packets, where both subspaces V j and W j are decomposed at each step [Mey94,Wic94,110,111]. Such a scheme provides rich collections (“libraries”) of orthonormal bases, each one corresponding to a given decomposition of L 2 (R) into a sequence of mutually orthogonal subspaces, chosen at successive scales j ∈ Z (an example is given below). Enumerating all possible such decompositions or orthonormal bases is a nontrivial combinatorial problem, whose solution stems from graph theory [Wic94,111]. Facing such a plethora of orthonormal bases, one needs a strategy for determining the best basis to use in a given situation. An efficient solution, based on entropic criteria, has been proposed by Coifman et al. [Wic94,111], under the name of the Best Basis Algorithm, and it has become a standard tool in signal analysis. Of course, all this applies verbatim in two dimensions, although the labeling of basis vectors becomes even more intricate. This holds, in particular, for the comments made at the end of Section 1.5 concerning the numerical implementation of finite reconstruction formulas. In order to fix ideas, we show in Figure 2.7 the subband subdivision scheme of the standard wavelet 1-D multiresolution analysis, in the case of a three-level decomposition. This corresponds to the following decomposition into orthogonal subspaces: V0 = W−1 ⊕ W−2 ⊕ W−3 ⊕ V−3 , or, in the notation of Section 1.5 [compare (1.58)], the representation of a signal s ≡ s0 ∈ V0 by its wavelet coefficients s0 = (d−1 , d−2 , d−3 , c−3 ). By comparison, Figure 2.8 shows the modified subdivision scheme used in the wavelet packets formalism, together

74

The 2-D continuous wavelet transform

V0 V−1 V−2 V−3

W−3

W−1 W−2

W−1

W−2

W−1

Fig. 2.7. The 1-D wavelet subband scheme, with a three-level decomposition.

V0 V−1 V−2 V−3

00 W−3

W−1 0 W−2

01 W−3

02 W−3

1 W−2 11 W−3

12 W−3

2 W−2 21 W−3

22 W−3

Fig. 2.8. The 1-D wavelet packet subband scheme, with a particular choice of three-level

decomposition.

with a particular choice of three-level decomposition, namely [compare Figure 1.11 (b) and (c)]: 2 11 12 V0 = V−1 ⊕ W−2 ⊕ W−3 ⊕ W−3 .

In the corresponding orthonormal basis, the signal is represented as s ≡ s0 = (c−1 , dd−2 , ccd−3 , dcd−3 ). The notation proceeds as follows. For j = −1, the coefficients are (c−1 , d−1 ). Then at each step, the coefficient x j is replaced by the pair (cx j−1 , d x j−1 ). Thus, for j = −2, one has, from left to right in Figure 2.7, (cc−2 , dc−2 , cd−2 , dd−2 ); for j = −3, (ccc−3 , dcc−3 , cdc−3 , ddc−3 , ccd−3 , dcd−3 , cdd−3 , ddd−3 ); and so on. For more details on the basis labeling, and wavelet packets in general, we refer to [Wic94] or [110,111].

2.5.2.4

The lifting scheme: second-generation wavelets One can go one step beyond, and abandon the regular dyadic scheme and the Fourier transform altogether. Using the “lifting scheme” leads to the so-called secondgeneration wavelets [349,350], which are essentially custom-designed for any given

75

2.5 Comparison with the 2-D discrete wavelet transform

problem. The starting point is that, in a biorthogonal scheme, one scale {V j } does not j } uniquely, but the freedom left in the generating wavelet determine its counterpart {V is known explicitly, and reduces to an arbitrary trigonometric polynomial. Thus the idea is to start from a given biorthogonal scheme with filters (h, h, g, g ), then tranform (1) (1) (1) (1) it using that freedom into a new one (h , h , g , g ), and so on, by a succession of “lifting steps.” But one must generalize first the very notion of biorthogonal scheme, in order to get a more flexible scheme. To that effect, one weakens Definition 1.5.1 of a multiresolution analysis {V j , j ∈ Z} by replacing the two conditions (1) and (2) (which enforce the scale invariance and thus the dyadic scheme) by the single condition (3) for each j ∈ Z, V j has a Riesz basis {ϕ j,k , k ∈ K( j)}, the elements of which are called scaling functions. Here K( j) is a general index set, which allows irregular sampling (no translation invariance!). One assumes only that K( j) ⊂ K( j + 1), without the dilation relation j }, with dual given by condition (1). In the same way, one considers a dual scale {V scaling functions { ϕ j,k , k ∈ K( j)}, biorthogonal to the previous ones: ϕ j,k | ϕ j,k = δkk , k, k ∈ K( j).

(2.127)

Then the filter h ≡ h j,k,l enters through a refinement equation ϕ j,k = h j,k,l ϕ j+1,l ,

(2.128)

l∈K( j+1)

h are and similarly for h ≡ h j,k,l (all filters are assumed to be finite). The two filters h, then biorthogonal (see (2.131) below). Next one defines wavelets in the usual way. A family of functions {ψ j,m , m ∈ M( j)}, where M( j) = K( j + 1) \ K( j), is a set of wavelet functions if: (i) the space W j = j ; (ii) the set span {ψ j,m , m ∈ M( j)} is a complement of V j in V j+1 , and W j ⊥ V 2 {ψ j,m /ψ j,m , j ∈ Z, m ∈ M( j), } is a Riesz basis for L (R). Dual wavelets are vectors of the biorthogonal basis, j ,m = δ j j δmm . ψ j,m |ψ

(2.129)

j , which complement V j in V j+1 , and W j ⊥ V j . By construction, They span spaces W the wavelets satisfy refinement relations: ψ j,m = g j,m,l ϕ j+1,l , (2.130) l∈K( j+1)

which thus define the filter g ≡ {g j,m,l }, and similarly for the dual filter g. Altogether, the four filters h, h, g, g satisfy biorthogonality relations: g j,m ,l = δmm , g j,m,l = 0 g j,m,l h j,k,l l

l

l

h j,k ,l = δkk , h j,k,l

l

h j,k,l = 0. g j,m,l

(2.131)

76

The 2-D continuous wavelet transform

Similar relations may be written for scaling and wavelet functions. Now, of course, since the index sets K( j) are general, some extra care is required to guarantee the convergence of all the expansions (hence the finiteness condition on the filters), and also the rapidity of the algorithm. We refer to [350] for technical details. This scheme becomes simpler if one introduces an operator notation (familiar in the signal processing literature), as follows. 2 2 r The filter h j,k,l is embodied in the operator H j : (K( j + 1)) → (K( j)), defined 2 2 by b = H j a, where a ≡ (al ) ∈ (K( j + 1)), b ≡ (bk ) ∈ (K( j)), and bk =

h j,k,l al .

l∈K( j+1)

r

The filter g j,m,l is embodied in the operator G j : 2 (K( j + 1)) → 2 (M( j)), and j , G j. similarly for the operators H In this notation, the conditions for exact reconstruction can be written in matrix form as: j j H 1 0 H = and = 1. H j∗ G ∗j H j∗ G ∗j j j 0 1 G G (2.132) Now we may describe the lifting scheme. As already mentioned, the idea is to exploit j biorthogonal to a given one H j , G j . The j , G the freedom in designing a set of filters H freedom is that of an arbitrary operator S j : 2 (M( j)) → 2 (K( j)). In components, this operator is represented by a set of coefficients, S j ≡ s j,k,m . In the usual, first generation wavelet scheme, this in turn would be given by a trigonometric polynomial s(ω). The technique proceeds in two steps. (i) First a lifting step, which consist in passing from a given biorthogonal filter set j , G j , G j } to a new one, {H j , H (1) , G (1) , G j }, where {H j , H j j (1) = H j, j + S j G H j

∗ G (1) j = G j − Sj Hj,

j being unchanged. Hence, the original scaling functions ϕ j,k the two filters H j , G j,m are modified. do not change, but all the other functions ϕ j,k , ψ j,m , ψ (1) , G (1) , G j , } to the set (ii) A dual lifting step, leading from the set {H j , H j j (1) (1) (1) (1) {H j , H j , G j , G j , }, where H j(1) = H j + S j G (1) j ,

(1) . (1) = G j − G S ∗j H j j

(1) , G (1) and S j : 2 (M( j)) → 2 (K( j)) is another operator. Here the two filters H j j remain unchanged.

77

2.5 Comparison with the 2-D discrete wavelet transform

Each of the two steps preserves the biorthogonality of the filter sets, as can be checked easily on the conditions (2.132). Of course, one has to verify along the way that the new scaling and wavelet functions belong correctly to the appropriate spaces (in particular, that they are square integrable). Then it may be shown that any biorthogonal filter set may be obtained by this procedure after a finite number of steps, starting from a trivial set, called the Lazy wavelet, because it does nothing but split the sequences into two subsets of components. j = E, G j = More precisely, the Lazy wavelet corresponds to the filter set H j = H j = D, where E : 2 (K( j + 1)) → 2 (K( j)) and D : 2 (K( j + 1)) → 2 (M( j)) are G simply the restriction or subsampling operators. [In the engineering literature, the lazy wavelet is also called a polyphase filter (of size 2) [Vet95].] The resulting scheme is fast, and independent from translation invariance and from the Fourier transform. Thus it applies to wavelets on intervals or on curves, and in higher dimensions as well, for instance, to wavelets on two-dimensional manifolds [Mal99,349,350]. Here the idea is to start from a succession of finer and finer grids. Consider a 2-D manifold (such as the 2-sphere) and choose a certain grid G( j) on it. Then refine the latter to a grid G( j + 1), of the form G( j + 1) = G( j) ∪ C( j), where C( j) denotes the complement. A typical example is to start from a triangulation of the manifold and refine it by bisecting each side, as illustrated in Figure 2.9 in the case of the sphere. Then the multiresolution spaces are defined as V j+1 = 2 (G( j + 1)), V j = 2 (G( j)), W j = 2 (C( j)), with appropriate bases {ϕ j,k } ∈ V j , {ψ j,m } ∈ W j , and the whole machinery is put into operation. The resulting tool has proven to be extremely versatile and efficient. For instance, Schr¨oder and Sweldens [336] have applied it to the design of wavelets on the sphere, with a very convincing application to the reproduction of coastlines on a terrestrial globe (we will see in Chapter 9, Section 9.2, another approach to the same problem, this one directly based on the CWT). As a final remark, we may point out that the lifting scheme opens the door to nonlinear multiresolution decompositions, such as the median transform of Bijaoui [Sta98] or the

Fig. 2.9. Typical grid refining for applying the 2-D lifting scheme: the geodesic sphere construction, starting with the icosahedron on the left (subdivision level 0) and the next two subdivision levels (from [336]).

78

The 2-D continuous wavelet transform

morphological wavelets introduced by Goutsias and Heijmans [198,219]. The latter are explicitly based on the lifting scheme and make the connection with the standard field of mathematical morphology. We note finally that the lifting scheme by itself offers a very pedagogical entrty into the wavelet world, as examplified in the little volume of Jensen and la Cour-Harbo [Jen01].

2.5.2.5

Integer wavelet transforms In their standard numerical implementation, the classical (discrete) WT converts floating point numbers into floating point numbers. However, in many applications (data transmision from satellites, multimedia), the input data consists of integer values only and one cannot afford to lose information: only lossless compression schemes are allowed. Recent developments, based on the lifting scheme, have produced new methods that allow one to perform all calculations in integer arithmetic [2,92]. In addition, such methods also improve the performances of lossy compression techniques [324].

2.6

Bridging the gap: continuous wavelet packets and fast algorithms

2.6.1

Custom design of dyadic frames Besides the full discretization described in Section 1.3, and the discrete WT just discussed, there is an intermediate procedure, introduced in [159], under the name of infinitesimal multiresolution analysis. It consists in discretizing the scale variable alone, on an arbitrary sequence of values (not necessarily powers of a fixed ratio). This leads to fast algorithms that could put the CWT on the same footing as the DWT in terms of speed and efficiency, by extending the advantages of the latter to cases where no exact QMF is available. We describe the method in 2-D, the 1-D case (already sketched in Section 1.6.1) being easily derived on this basis. Interested readers should refer to [Tor95] for further details. Instead of the standard L 2 -normalization used in (2.13), it is more convenient to Note that, for simchoose the L 1 -normalization and use ψ(b,a) = a −2 ψ(a −1 ( x − b)). plicity, we consider here only isotropic wavelets, but the extension to the general case is straightforward (see Section 2.6.3). Let us start with the L 1 -reconstruction formula associated to the CWT in two dimensions: ∞ da ˘ b, a) ψ(b,a) d 2 b S( x ). (2.133) f ( x) = ( 2 a R 0 The basic idea behind the proposed construction is now to segment the integral over scales in (2.133) and replace it by a sum over dyadic intervals. This is done first by rewriting the reconstruction formula as ∞ da f ( x) = x ), da ( a 0

79

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

where we have defined the infinitesimal detail ˘ b, a) ψ(b,a) da ( x) = d 2 b S( x ). ( R2

By virtue of Young’s convolution inequality, da ∈ L 2 and, taking its Fourier transform, we obtain k)| . = |ψ(a 2 da (k) f (k)

(2.134)

These equations show that da represents the amount of information captured by the wavelet between scales a and a + da, hence the name “infinitesimal details.” Summing all these details, that is, integrating over the scale variable, reproduces the original signal. In the same vein, we can synthesize a low resolution approximation of f by integrating up to a given resolution, say a0 : ∞ da da ( x) = x) . f a0 ( a a0 Taking Fourier transforms on both sides suggests to introduce the following Fourier multiplier: ∞ da 2 2 |φ(k)| = |ψ(a k)| . (2.135) a 1 It is then shown in [159] that the approximation f a can be written f a ( x) = d 2 b φ(b,a) x ), | f φ(b,a) (

(2.136)

R2

and that the following limit holds in the strong sense in L 2 : lim f a = f.

a→0

(2.137)

Thus, following [Tor95], we speak of the bilinear formalism. Remark also that (2.135) implies k)| 2=0 lim |φ(

|k|→∞

and thus defines a scaling function. Now, starting from an approximation of f at scale a0 = 2− jo , we can refine up to an arbitrary resolution by adding up details. For this purpose, we introduce slices of details 2− j da D j ( da ( x) = x) . 2−( j+1) a Taking Fourier transforms on both sides, we have 2− j da 2 ( = |ψ(a k)| f (k) D j (k) −( j+1) a 2

80

The 2-D continuous wavelet transform

and this leads us to define the integrated wavelet packets: 2= | (k)|

1

1/2

da 2 |ψ(a k)| . a

(2.138)

This function satisfies a two-scale relation, the analog of (1.62): 1 k)| k)| 2 − |φ( 2. 2 = |φ( | (k)| 2

(2.139)

Finally, putting equations (2.136) and (2.138) together, we obtain the following dyadic decomposition: f =

R2

2

d b φ(b,2 − jo ) | f φ(b,2 − jo ) +

∞ j= jo

R2

d 2 b (b,2 − j ) | f (b,2 − j ),

(2.140)

which holds in L 2 norm. In order to simplify our notations, we will write the wavelet 2− j ) = (b,2 coefficients of f as W f (b, − j ) | f and introduce the approximation coef−j ficients S f (b, 2 ) = φ(b,2 − j ) | f . Equation (2.140) now reads x) + f ( x ) = S f (·, 2− jo )|φ(·,2− jo ) (

∞

W f (·, 2− j )|(·,2− j ) ( x)

(2.141)

j= jo

(scalar product over b in each term). We have thus built a dyadic wavelet transform starting from the CWT. It is important to realize that the scaling function φ and the integrated wavelet packet inherit the localization and smoothness properties of ψ. In view of the considerable freedom we have in the choice of ψ, we are now able to easily design custom, translation invariant dyadic frames. Remark: More flexibility is obtained if one subdivides the scale interval [1/2, 1] into n subbands, by a0 = 1/2 < a1 < . . . < an = 1. In that case, one ends with one scaling function ( x ) and n integrated wavelets i ( x ), i = 0, . . . n − 1, corresponding to integration from ai−1 to ai . This more efficient version allows one to compute explicitly the characteristics of the wavelet packet, such as its central frequency, its standard deviation, etc., following (1.14)–(1.15), and these in turn may be expressed in terms of the corresponding quantities of the mother wavelet ψ. An application to sound analysis is given in [233]. A simpler decomposition formula, called the linear scheme in [Tor95], arises when one starts from the so-called Morlet reconstruction formula (2.34) (this is the scheme we have sketched in the 1-D case in Section 1.5). The same reasoning as before leads us to introduce, as in (1.61), a scaling function ∞ da k) = φ( ψ(a k) , (2.142) a 1

81

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

and integrated wavelet packets 1 da = ψ(a k) (k) 1/2 a k) 1 k) − φ( . = φ( 2

(2.143) (2.144)

One notices that these wavelets are simply expressed as a difference of smoothing functions, as in 1-D, (1.62). In this case, the reconstruction formula is much simpler, since it involves a straight sum of approximation and wavelet coefficients: f ( x ) = S f ( x , 2− jo ) +

∞

W f ( x , 2− j ) .

(2.145)

j= jo

One of the main advantages of this construction is that it allows to build wavelets and scaling functions that have fast decay both in the spatial and frequency domains. This is very useful in applications where one wants to use wavelets that have sharp prescribed localization properties in the Fourier domain and are also of fast decay in the spatial domain, as it is the case with Gabor functions. This is very difficult to achieve in practice. For example, if one wants to use spline-based wavelet frames, it appears that, although the spatial localization is very good, splines are not sharply localized in Fourier variables (they have an algebraic decay, see [14] for a review) and can even show disturbing sidelobes. A concrete example is given by texture analysis where the latter are distinguished on the basis of the statistics of frequency subbands as measured using Gabor wavelets [234]. If one wants to use a dyadic frame, special care has to be given to the frequency localization of the wavelets and this can be easily done using the technique described above.

2.6.2

Example of a typical design We will now apply the previous formalism to a concrete example. Our aim is to construct isotropic dyadic wavelets and scaling functions belonging to the class of C ∞ functions with fast decay. Our starting point is a family of wavelets associated to n/2 pseudo-differential operators defined by multiplication by k n = k12 + k22 in the Fourier domain: k) = k n e−k 2 . ψ( The associated scaling functions are computed using (2.135): ∞ da n −a 2 a e , (n 2) φn (k) = a k and it satisfies the following recurrence relation: n (k) = 1 k n−2 e−k 2 + (n − 2) φ n−2 (k) , φ 2 2

(2.146)

82

The 2-D continuous wavelet transform

with 2 (k) = 1 e−k 2 , φ 2

3 (k) = 1 ke−k 2 + φ 2

*

π erfc(k). 4

Here the error function is defined by ∞ 4 2 dk e−k . erfc(k) = √ π k Normalizing this family of functions by

R

d 2 x φn ( x ) = 1 leads us to define

n (k) = αn−1 φ(k), φ where the constant αn is defined by n−2

2 (0) αn = φ

2 n − 2i

i=1

2

, n even,

n > 2,

n−3

3 (0) αn = φ

2 2i + 1 , n odd, 2 i=1

n > 3.

2 (0) and α3 = φ 3 (0). Scaling functions of the lowest The recursion starts with α2 = φ orders are listed in Table 2.1. Using (2.138) or (2.144), we obtain the desired family of isotropic wavelets. An example of such a scaling function of order 4 is given in Figure 2.12, in the next section. Note that the parameter n also controls the number of vanishing moments of the associated wavelet. As an example of this technique, Figure 2.10 shows a three-level decomposition of the lena image into isotropic wavelet packets.

2.6.3

Designing directional dyadic frames The previous construction applied only to isotropic wavelets and yielded frames of isotropic elements. In the following we will add to this scheme sensitivity to local orientation. A simple and straightforward way to achieve this is to start from an isotropic integrated wavelet and segment it into directional ones (a precise definition of this Table 2.1. Scaling functions of lowest orders Order

Scaling function

n=2 n=3 n=4

2 (k) = e−k 2 φ √ 3 (k) = 1 ke−k 2 + π erfc(k) φ 2 4 2 4 (k) = e−k (1 + k 2 ) φ 5 (k) = α5−1 1 k 3 e−k 2 + 3 ke−k 2 + φ 2 4

n=5

√ 3 π 8

erfc(k)

83

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

Fig. 2.10. Three levels of decomposition of the lena image using an isotropic wavelet packet frame. The lower right image is the low resolution approximation.

notion will be given in Section 3.3). This can be done by introducing an angular window η(ϕ), ϕ ∈ [0, 2π ), in the Fourier domain and then defining a new wavelet ϕ) = (k, (k)η(ϕ) . Note that this construction amounts to work with wavelets that are separable in polar coordinates. The choice of the angular window is restricted by the need for an exact, linear or bilinear, reconstruction formula. More precisely, if one makes use of (2.141), η has to satisfy + L−1 ++ + 2 +η ϕ − 2π + = 1, + L + =0

(2.147)

84

The 2-D continuous wavelet transform

while the simpler formula (2.145) requires L−1 2π η ϕ− = 1, L =0

(2.148)

where we have assumed L orientations. An additional requirement, further explored in Chapter 3, is that the support of η be strictly less than π in order to have some directional sensitivity. In order to preserve the frequency localization of , it is also important that η be regular enough. The optimal choice is thus to build a partition of the circle using a suitable compactly supported C ∞ function. Altogether, we introduce the L angular windows η (ϕ) ≡ η ϕ − 2π ,= L 0, 1, . . . , L − 1, and the corresponding directional wavelets (k, ϕ) = (k) η (ϕ), = 0, 1, . . . , L − 1 .

(2.149)

An example of a dyadic directional wavelet built using this technique is depicted on Figures 2.11 and 2.12.

2.6.4

Implementation using approximate QMFs One of the main drawbacks of these oriented frames is that they are not designed to be implemented using a fast pyramidal algorithm. Nevertheless we will now show that one can design special QMF pairs that allow for a very good approximation of the discrete WT and provide a substantial gain in computational speed. This technique is mainly an extension to 2-D of the original work of Muschietti and Torr´esani [291]. −3.14

−3.14

0

0

3.14 −3.14

0

(a)

3.14

3.14 −3.14

0

3.14

(b)

Fig. 2.11. Fourier transform of a directional dyadic wavelet of order 4 and angular resolution of π/5 for two values of the rotation parameter: (a) = 0; and (b) = 1.

85

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

1 0.025

0.6

0.02 0.015

0.8

0.01

0.4 0.005

0.2

0 −0.005

0.0

−0.01

−0.2 1

−0.015 1

1

0.5

1

0.5

0.5

0 −0.5

0

−0.5

−0.5 −1

0.5

0

0 −1

−0.5 −1

(a)

−1

(b)

0.025 0.02 0.015

0.02

0.01

0.015

0.005 0

0.01

−0.005 −0.01

0.005

−0.015

0 1

−0.02 1 1

0.5 0.5

0

1

0.5

0

−0.5

0

−0.5

−0.5 −1

0.5

0 −0.5 −1

−1

−1

(d)

(c)

Fig. 2.12. (a) Scaling function of order 4; (b–d) the associated wavelet with angular resolution of π/5: (b) real part; (c) imaginary part; and (d) modulus.

Since we work with translation invariant frames, we will now sample the position parameter of the DWT over a regular grid, that is, we will consider wavelets and scaling functions indexed by the integer grid (we use here the L 1 -normalization): j;m,n (x, y) = 22 j (2 j (x − m, y − n)) ,

j, m, n ∈ Z .

We then obtain a discrete dyadic wavelet transform (compare Section 2.4.4) by just restricting the DWT to this particular grid: W j f (m, n) = j;m,n | f ,

S j f (m, n) = φ j;m,n | f .

(2.150)

86

The 2-D continuous wavelet transform

Let us assume there exists 2-D discrete filters h and g , = 0 . . . L − 1, such that one can compute these quantities using a pyramidal algorithm [260], that is, S j f (m, n) = h p,q S j+1 f (m − 2 j+1 p, n − 2 j+1 q) (2.151) p,q∈Z

W j f (m, n) =

g p,q S j+1 f (m − 2 j+1 p, n − 2 j+1 q) .

(2.152)

p,q∈Z

This is equivalent to asking that the related wavelets and scaling function satisfy a two-scale equation of the form: k) k), = h(k) φ( φ(2 (2k) (k), = g (k)

(2.153) (2.154)

≡ g (k x , k y ), = 0 . . . L − 1, are the Fourier series ≡ h(k x , k y ) and g (k) where h(k) of the filters h and g respectively. As already stressed in Chapter 1, we know that the same filters can be used to reconstruct the signal provided they satisfy a QMF relation: 2+ |h(k)|

L−1

2 = 1. |g (k)|

=0

This is the exact discrete equivalent of (2.141). Similarly, one can design filters that implement the weaker formula (2.145), provided they satisfy a simpler constraint: + h(k)

L−1

= 1. g (k)

=0

The main problem at this point is that the integrated directional wavelets we have designed in Section 2.6.1 do not in general satisfy any two-scale equation. Another way to formulate the problem is to remark that, although there usually exists regular multipliers hˇ and gˇ satisfying k) k) ˇ k) = h( φ( , φ(2

k) (2k) = gˇ (k) φ( ,

these are in general not 2π × 2π periodic and hence cannot be used to compute successive approximations and details as in (2.151) and (2.152). Nevertheless it is crucial to notice that these multipliers, exactly as the filters h and g , are always multiplied by Now, if the latter is very well localized in [−π, π] × [−π, π], the lack of periodicity φ. and it seems then reasonable to of hˇ and gˇ is compensated by the localization of φ ˇ gˇ using periodic filters. For this purpose, let us look for good approximations of h, 2 assume that our signal lives in L (R2 ) and introduce the subspace of L 2 (R2 ) spanned by the integer translates of the scaling function defined in (2.135): V0 = { f ∈ L 2 | f = cm,n φ(x − m, y − n) , cm,n ∈ 2 } m,n∈Z

87

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

and suppose that the family {φ(x − m, y − n) , m, n ∈ Z} forms a Riesz basis of V0 . Finding a best approximant for hˇ can be formulated as finding an element of V0 whose distance to 14 φ(x/2, y/2) is minimal in L 2 . Similarly, finding best approximants for the gˇ is equivalent to finding those elements of V0 that minimize the distance to 1 (x/2, y/2). In other words, the problem is to minimize the L 2 distances 4 !1/2 a 2 ˇ a 2 ˇ d k |h(k) − h (k)| |φ(k)| , ν(h, h ) =

ν(gˇ , g

,a

)=

R2

k)| − g ,a (k)| |φ( 2 d k |gˇ (k) 2

R2

!1/2 .

Now since V0 is a vector subspace of a Hilbert space, the projection theorem applies and guarantees the existence and uniqueness of such solutions. Suppose that h a ∈ V0 ˇ It can be expanded as is the solution for h. ha = h am,n φm,n , m,n∈Z

where the h am,n are the Fourier coefficients of the approximate filter ≡ h a (k x , k y ) = h a (k)

1 a −i(mkx +nk y ) h e . 4π 2 m,n∈Z m,n

The projection theorem gives also the following characterization of these coefficients:

# x y 1 d x d y φ(x + m, y + n) h ap,q φ(x + p, y + q) − φ( , ) = 0 , 4 2 2 R2 p,q∈Z "

∀ m, n ∈ Z .

(2.155)

,a Similarly, for the directional wavelets, one obtains L approximate filters gm,n characterized by " # 1 x y ,a d x d y φ(x + m, y + n) g p,q φ(x + p, y + q) − ( , ) = 0 , 4 2 2 R2 p,q∈Z

∀ m, n ∈ Z .

(2.156)

Finally, for j −1, the approximate low resolution and detail coefficients read S aj f (m, n) = h ap,q S aj+1 f (m − 2 j+1 p, n − 2 j+1 q) (2.157) p,q∈Z

and W ,a j f (m, n) =

p,q∈Z

a j+1 g ,a p, n − 2 j+1 q) . p,q S j+1 f (m − 2

(2.158)

88

The 2-D continuous wavelet transform

The following theorem, due to Gobbers and Vandergheynst [360], extends to 2-D the result obtained in 1-D by Muschietti and B. Torr´esani [291]. It gives explicit formulas for the best approximants h a and g ,a , as well as an estimation of the error with respect to the original coefficients S j f (m, n) and W j f (m, n): Theorem 2.6.1 (i) The optimal filters h a and g ,a , solutions of (2.155) and (2.156), are given by

p,q

h (k x , k y ) = a

2(k x + 2π p), 2(k y + 2πq) φ(k x + 2π p, k y + 2πq) φ ,

2 p,q |φ(k x + 2π p, k y + 2πq)| (2.159)

g

,a

(k x , k y ) =

p,q

2(k x + 2π p), 2(k y + 2πq) φ(k x + 2π p, k y + 2πq) .

2 p,q |φ(k x + 2π p, k y + 2πq)| (2.160)

and C = 2 ess sup g ,a (k), then, for all j −1, (ii) If we write† C0 = 2 ess sup h a (k) R2 k∈

R2 k∈

| j|

ˇ h a ) 1 − C 0 f 2 , S aj f − S j f ∞ 2 ν(h, 1 − C0 and

W ,a j f − W j f ∞

| j|−1

ˇ h a ) 1 − C0 2 ν(gˇ , g ,a ) + C ν(h, 1 − C0

f 2 . (2.161)

Proof . We will essentially follow the proof given in [291]. The first part is obtained by taking the Fourier transform of (2.155) and (2.156) and using the periodicity of h a and g ,a . We then have 2π 2π + +2 ik x m+ik y n + dk x dk y e φ(k x + 2π k, k y + 2πl)+ 2 h a (k x , k y ) 0

0

k,l∈Z

−

, x + 2πk, k y + 2πl) φ 2(k x + 2π k), 2(k y + 2πl) = 0 , 2 φ(k

k,l∈Z

(2.162) which gives the result for h a and similarly for g ,a . As for the second part of the theorem, the inequality f ∞ f 1 /4π 2 allows us to work directly in the Fourier domain. For the purpose of the calculation, let us introduce the intermediate quantities †

Since this quantity is orientation independent, we drop the corresponding superscript for ease of notation.

89

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

Sj (m, n) =

h ak,l S j+1 m − 2 j+1 k, n − 2 j+1l ,

k,l∈Z

(m, n) = W j

j+1 g k,a p, n − 2 j+1 q . p,q S j+1 m − 2

p,q∈Z

We have the inequality -(a -(a - + -S − S - , S − S − S S - j - j - j jjj1

1

(2.163)

1

- ,a and similarly for -W j − W j 1 . Let us now compute the second term in the right-hand side of (2.163): - dk x dk y 2 j -S j − S j - = 1 S 1 ×S 1 + + + 2 j (k x + 2π k), 2 j (k y + 2πl) f (k x + 2πk, k y + 2πl)φ + + k,l∈Z f (k x + 2πk, k y + 2πl) h a 2 j−1 k x , 2 j−1 k y − k,l∈Z

+ + j−1 2 (k x + 2π k), 2 j−1 (k y + 2πl) ++ . ×φ +

Using the Cauchy–Schwarz inequality and the periodicity of h a , we find . + + - k) k) φ( − φ(2 +2 d 2 k + h a (k) -S j − Sj - 8π 2 f 2 1

R2

ˇ h a ). 8π f 2 ν(h, 2

For the first term on the right-hand side of (2.163), we find + + + + a -(a + d 2 k +h a 2 j+1 k + +S -S j − S j - 2 j+1 (k) − S j+1 (k) 1 R2 - a 2 ess sup |h a | -S − S - . j+1 j+1 R2 k∈

1

Combining these estimations, we have -(a a ˇ h a ) + C0 − S -S - . -S j − S j - 8π 2 f 2 ν(h, j+1 j+1 1

1

(2.164)

By iteratively bounding the last term of (2.164) in the same way, we finally obtain -(a ˇ h a ) 1 + C0 + C02 + . . . + C | j|−1 -S j − S j - 8π 2 f 2 ν(h, 0 1

| j|

ˇ ha ) = 8π 2 f 2 ν(h,

1 − C0 . 1 − C0

An equivalent processing of the wavelet coefficients concludes the proof.

(2.165)

90

The 2-D continuous wavelet transform

There remain several open questions related to the design of these approximate QMFs. In particular, in contradiction to the orthogonal case discussed in the previous chapter, the convergence of the cascade algorithm is not ensured. It is thus impossible to tell precisely if these filters really correspond to a frame and what would be the properties of such a frame. This scheme is thus really a handy numerical shortcut that allows one to efficiently compute an approximation of the original frame expansion.

2.6.5

Some implementation issues When implementing (2.157) and (2.158), the first fact to consider is that, generally, we possess only a finite number M × N of samples f (m, n) (with (m, n) ∈ [0 . . . M − 1, 0 . . . N − 1] ) of the signal to be analyzed. A decision has thus to be made about the nature of the signal outside this range. Furthermore, the filters will practically be computed on a finite grid. That is, we will only use a finite number P × Q of filter coefficients h ap,q and g k,a p,q , with ( p, q) ∈ [−P/2 . . . P/2 − 1, −Q/2 . . . Q/2 − 1]. In a first approach, the signal is considered to be zero outside this range. Unfortunately, this decision leads to impractical and inefficient algorithms, as one has to compute (M + 2| j|−1 (P − 1)) × (N + 2| j|−1 (Q − 1)) coefficients for each j < 0 to avoid side effects. A second approach is to consider the signal as being periodic of period M × N . With this approach, fast circular convolution algorithms can be used and we are led to the following algorithm structure: (i) compute h a and g ,a and their associated impulse responses using Theorem 2.6.1; (ii) compute a first approximation of the analyzed signal f using (2.150); this step is traditionally skipped and the signal is considered as its first approximation; (iii) for each j < 0, iteratively compute details and approximations using (2.157) and (2.158). The cost of this algorithm is: C(M, N ) = c · M · N where the constant c depends on the size of the impulse responses of the filters. This dependency strongly limits the power of this algorithm as experiments show that, even for small sizes, this algorithm is always slower than the FFT-based algorithm used to compute the CWT. As those sizes may not be arbitrarily chosen, another algorithm should be used in order to get valuable results with small computation times. The obvious way to handle this problem is to replace convolutions in direct space by products of Fourier transforms in frequency space. The main advantage of this technique is that we no longer need to restrict ourselves to use small filters, as we may now use impulse responses that are the same size as the signal, thus giving much

91

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

more precise results than with the above algorithm. For the sake of simplicity, we will develop this new algorithm in the 1-D case, its extension to the 2-D directional case being straightforward. For the first step, we have to compute two periodic convolution products: c−1,k =

N −1

c0,k−n h n

n=0

d−1,k =

N −1

c0,k−n gn .

n=0

By introducing {C j,n } = F F TN {c j,k } , {D j,n } = F F TN {d j,k } , {Hn } = F F TN ({h k }) and {G n } = F F TN ({gk }) where the notation F F TN means the FFT algorithm ap√ √ plied to a sequence of length N , we get: C−1,n = N C0,n Hn and D−1,n = N C0,n G n . We may then get the details back in the real space with {d j,k } = I F F TN {D j,n } . For the second step, we have to compute: c−2,k =

N −1

c−1,k−2n h n

n=0

d−2,k =

N −1

c−1,k−2n gn .

n=0

Given the periodicity of the signal, we may rewrite this as:

0 c−2,k ≡ c−2,2k =

N /2−1

c−1,2(k−n) (h n + h n+N /2 )

n=0 1 c−2,k

≡ c−2,2k+1 =

N /2−1

c−1,2(k−n)+1 (h n + h n+N /2 )

n=0 0 d−2,k ≡ d−2,2k =

N /2−1

c−1,2(k−n) (gn + gn+N /2 )

n=0 1 ≡ d−2,2k+1 = d−2,k

N /2−1

c−1,2(k−n)+1 (gn + gn+N /2 )

n=0

for k = 0, . . . , N /2 − 1. Let us now define h 1n = h n + h n+N /2 for n = 0, . . . , N /2 − 1, we get: 0 c−2,k =

N /2−1

0 1 c−1,k−n h 1n c−2,k =

n=0 0 d−2,k =

N /2−1 n=0

N /2−1

1 c−1,k−n h 1n

(2.166)

1 c−1,k−n gn1

(2.167)

n=0 0 1 c−1,k−n gn1 d−2,k =

N /2−1 n=0

92

The 2-D continuous wavelet transform 0 1 0 1 where c−1,k = c−1,2k , c−1,k = c−1,2k+1 , d−1,k = d−1,2k and d−1,k = d−1,2k+1 . In the Fourier space, (2.166) and (2.167) may be rewritten as: / / 0 0 1 1 C−2,n = N /2C−1,n Hn1 C−2,n = N /2C−1,n Hn1 / / 0 0 1 1 D−2,n = N /2C−1,n G 1n D−2,n = N /2C−1,n G 1n

0 1 0 1 0 where {C−2,n } = F F TN /2 {c−2,k } , {C−2,n } = F F TN /2 {c−2,k } , {D−2,n }= 0 1 1 F F TN /2 {d−2,k } and {D−2,n } = F F TN /2 {d−2,k } . A straightforward calculation gives: √ Hn1 = 2H2n for n = 0, . . . , N /2 − 1. Furthermore: 1 0 C−1,n = √ (C−1,n + C−1,n+N /2 ) 2 2iπ Nn e 1 C−1,n = √ (C−1,n − C−1,n+N /2 ) 2 for n = 0, . . . , N /2 − 1. The last equation is particularly interesting as it introduces the same twiddle factors as those already present in the traditional FFT implementations. Extending the above results to the following steps, one comes to the conclusion that the complexity of this new algorithm is also of order N log2 (N ), but with a hidden constant that is exactly half the one encountered in the FFT-based algorithm, this constant being associated with the pyramidal structure. The same ideas apply to the 2-D case, thus giving an algorithm of complexity C(M, N ) = M · N · log2 (M.N ), with a hidden constant also exactly half that of the FFT-based algorithm. As a matter of comparison, Figure 2.13 shows timings of this algorithm for different image sizes. Timings of the usual pyramidal algorithm and standard implementation in the Fourier domain are also displayed. Putting it all together, we have a new fast algorithm, perfectly suited to compute the 2-D CWT, faster than the traditional “pseudo-pyramidal” algorithm, and sharper. Furthermore, it is essential to note that the whole construction is equivalent to that leading to the FFT algorithm in the Fourier transform theory. It should be noted that a Fourier implementation of the pyramidal algorithm is quite natural when one addresses the problem of designing maximally regular wavelets. That is why the algorithm described above shares common features with the Fourier implementation of the Meyer wavelet decomposition [Kol97]. We refer the interested reader to the work of Rioul and Duhamel [327] for more general considerations on implementing pyramidal algorithms in the frequency domain.

93

2.7 Steerable filters

6

5 pyr (32)

4

3

2 pyr (16)

std_F

1

pyr (8) pyr_F 0 128

256

512

1024

Fig. 2.13. Computation time for different implementations of the 2-D pyramidal algorithm as a function of the image width (square images assumed): pyr(n) is the standard pyramidal algorithm with n by n filters, pyr F is the modified Fourier pyramidal algorithm explained in the text. The graph is normalized with respect to the standard algorithm with convolution in the Fourier domain (std F).

2.7

Steerable filters While looking for a flexible tool for processing oriented data, Freeman and Adelson [170] introduced some time ago the concept of steerable filters. These were further developed by Perona [311] and Simoncelli et al. [342]. Here again one obtains a multiscale pyramid decomposition, which is quite efficient in a number of problems, mostly related to machine vision. Similar techniques have been used with the Gabor transform [234]. We will briefly describe this scheme and compare it to the directional wavelet packets of Section 2.6. The basic idea is quite simple, and best illustrated on the example of a Gaussian kernel 2 2 2 2 G(x, y). From the partial derivatives G x (x, y) = −2x e x +y , G y (x, y) = −2y e x +y , one computes the derivative in the direction θ: G θ (x, y) = cos θ G x (x, y) + sin θ G y (x, y). Since convolution is a linear operation, one may use G θ for filtering an image f in the direction θ by superposing the filterings in directions x and y: = (G θ % f )(b) = cos θ Fx (b) + sin θ Fy (b). Fθ ( f )(b)

94

The 2-D continuous wavelet transform

This is the property of orientability. More generally, a filter f is orientable or steerable if any oriented version of it may be obtained from a finite number of basic orientations: x )) = f (rθ (

M

km (θ) f (rθm ( x )).

(2.168)

m=1

The weights {km (θ), m = 1, . . . , M} are called interpolation functions. (The notion of orientability may be extended to other transformations, such as scaling [311], but we will not consider these generalizations here.) When the filter f admits a finite Fourier series (i.e., it is a real trigonometric polynomial), f (r, ϕ) =

N

an (r ) einϕ ,

x ≡ (r, ϕ),

(2.169)

n=−N

Freeman and Adelson have shown that the interpolation functions must satisfy the relation 1 ... 1 1 eiθ eiθ1 . . . eiθ M k1 (θ) . (2.170) .. = .. .. .. . . . k M (θ) ei N θ e i N θ1 . . . e i N θ M and that the minimal number of interpolation functions is always larger than the number of nonzero coefficients in the angular Fourier series (2.169) of the filter f . These steerable filters obviously generalize the interpolation properties associated to the partial derivatives G x , G y , which are thus prototypes of such filters. Moreover, the steerability property (2.168) is independent of the radial part of the filter, so that one may use functions that generate a dyadic pyramid. The main virtue of steerable filters is to allow the computation of filtering in any direction from the interpolation functions and the basic filters. This explains their intensive use in vision for studying oriented features in images [311,342]. From the algorithmic point of view, the complexity of steerable filters is that of the associated dyadic pyramid, thus comparable to that of directional wavelet packets. Note, however, that separable filters may sometimes be obtained, which is an additional bonus. So the natural question arises, can directional wavelets be made steerable? Unfortunately, the answer is no. Indeed, the Fourier transform of a steerable filter is also steerable. On the other hand, the angular support of a directional wavelet (2.149) necessarily has a width smaller than π , so that the relation (2.169) can never be satisfied. More intuitively, the steerability condition (2.168) requires that the basic filters overlap in a fixed way, in order to ensure the existence of interpolation functions, whereas the angular overlap of directional wavelet packets may be taken to be arbitrarily small. What about angular resolution? That of steerable filters may be made as good as one wishes, simply by taking more basic filters. However, because of the substantial overlap

95

2.7 Steerable filters

(d)

(b)

(c)

(a)

Fig. 2.14. Four reconstructions of the image barbara: (a) 2-D wavelet orthonormal basis; (b) redundant frame (3 bits/coefficient); (c) and (d) the same with 2 bits/coefficient.

between the latter, the steerable scheme will always require more basic filters than the directional wavelet packets in order to achieve a high angular resolution, so that, in the end, the computing cost may become prohibitive. In conclusion, steerable filters and directional wavelet packets are two comparable, yet incompatible, tools for decomposing an image into an oriented pyramid. In the former case, however, a large number of basic filters, with fixed overlap, is required for achieving a high angular resolution. This precludes their use in applications that require a maximal decorrelation of orientations, such as watermarking of images or texture segmentation. Here directional wavelet packets are the best choice. We will explore these applications in the next chapters (Sections 4.7.2 and 5.5, respectively).

96

The 2-D continuous wavelet transform

2.8

Redundancy: plus and minus Exactly as in 1-D, redundancy has many advantages, that more than compensate the higher computational cost it implies. In a nutshell, redundant decompositions lead to better quality reconstructions and are more robust to noise. Actually, the whole discussion of Section 1.6.2 could be repeated here almost verbatim, in particular concerning the robustness issue. We highlight the reconstruction aspect with one striking example. We consider the standard barbara image and decompose it in two ways, first with an orthonormal wavelet basis (using 2-D Daubechies DB4 wavelets), then with a redundant frame of directional wavelets. The images reconstructed by the two methods are presented in Figure 2.14. Panels (a) and (b) show the reconstruction using 3 bits per coefficient, while (c) and (d) show the result obtained with 2 bits per coefficient. In either case, the resulting image is visually better when the redundant frame is used, (b) or (d). The orthonormal basis gives more artifacts and distortions. Of course, the effect is more marked in the 2 bit case. Although the two results are poor, we show them for emphasizing the point.

3

Some 2-D wavelets and their performance

3.1

Which wavelets? The next step is to choose an analyzing wavelet ψ. At this point, there are two possibilities, depending whether one is interested or not in detecting oriented features in an image, i.e., regions where the amplitude is regular along one direction and has a sharp variation along the perpendicular direction. (i) Isotropic wavelets If one wants to perform a pointwise analysis, i.e., when no oriented features are present or relevant in the signal, one may choose an analyzing wavelet ψ which is invariant under rotation. Then the θ dependence drops out, for instance, in the reconstruction formula (2.26). The most familiar example is the isotropic 2-D Mexican hat wavelet (2.21). (ii) Anisotropic wavelets When the aim is to detect oriented features in an image (for instance, in the classical problem of edge detection or in directional filtering), one has to use a wavelet which is not rotation invariant. The best angular selectivity will be in obtained if ψ is directional, which means that the (essential) support of ψ spatial frequency space is contained in a convex cone with apex at the origin (by which we mean that the wavelet is numerically negligible outside the cone). Typical directional wavelets are the 2-D Morlet wavelet (2.22) or the conical wavelets. There are many ways of designing wavelets of either kind, but in fact almost all of those available on the market may be obtained by a general procedure, outlined in the proposition below. The starting point is a scaling function, that is, a function φ( x ), whose integral over the plane does not vanish. Then wavelets can be built by taking derivatives of the scaling function or the difference of two scaling functions. The most obvious example of a scaling function is a Gaussian, which is very easy to use and essentially localized in a disk. Another method, that we shall describe in Chapter 6, is to impose the saturation of the uncertainty product of two or more operators corresponding to infinitesimal generators

97

98

Some 2-D wavelets and their performance

of elementary operators, namely the dilation, orientation and translation operators, as given in Section 2.1 above. Proposition 3.1.1 Let φ be a (sufficiently smooth) scaling function, that is, a function satisfying: d 2 x φ( x ) = 0. (3.1) R2

Then the functions ψ1 and ψ2 defined below are wavelets: N M

∂n ∂m φ( x ), where x ≡ (x, y), N , M 1, ∂ x n ∂ ym n=1 m=1 1 1 x) = x) − x ), U (b1 , a1 , θ1 )φ ( U (b2 , a2 , θ2 )φ ( ψ2 ( a1 a2 (b1 , a1 , θ1 ) = (b2 , a2 , θ2 ), ψ1 ( x) =

cnm

(3.2)

(3.3)

a, θ ) is the unitary operator already defined in (2.13): where U (b, a, θ )s ( x − b)). U (b, x ) = a −1 s(a −1 r−θ ( The proof is straightforward. We shall now examine in detail several examples of wavelets of each kind. As will be seen in the sequel, almost all of them fall into the general types described in the proposition, possibly combining the two operations of derivation and difference. It turns out also that in many cases, the basic scaling function is the Gaussian exp(−| x |2 ). This is not a coincidence. In order to understand this, assume the wavelet ψ to be, as in (3.2), some derivative of a scaling function φ, ψ( x) =

∂n ∂m φ( x ). ∂ x n ∂ ym

Then we may rewrite the basic formula (2.19) as a genuine convolution: ∂n ∂m # a, θ ) = φa,θ ∗ n m s (b) S(b, ∂x ∂ y ∂ n ∂ m # = φa,θ ∗ s (b), ∂bx ∂b y

(3.4) (3.5)

where # φa,θ ( x ) = a −1 φ(−r−θ ( x )/a).

a, θ) is simply the derivative of the signal rotated and Equation (3.5) shows that S(b, blurred at resolution (scale) a. Thus we expect large wavelet coefficients to occur at locations of sharp variation of s. Since the Gaussian is a very common kernel for blurring, it is not a surprise that many wavelets will be related to derivatives of a Gaussian. As for the populatity of the latter, it is due to several factors: it is very well

99

3.2 Isotropic wavelets

localized, both in position and in Fourier space, it is indeed optimal for the joint space– frequency localization, as argued by Gabor [180], and finally it is rotation invariant, thus it does not privilege any particular direction.

3.2

Isotropic wavelets

3.2.1

The 2-D Mexican hat and its generalizations In its isotropic version, this is simply the Laplacian of a Gaussian: x ) = − exp(− 12 | x |2 ) = (2 − | x |2 ) exp(− 12 | x |2 ). ψH (

(3.6)

This is a real, rotation invariant wavelet, with vanishing moments of order up to 1, also known in the literature as the LOG wavelet. It is shown, in position domain and in spatial frequency space, in Figures 3.1 and 3.2, respectively. It was originally introduced by Marr and Hildreth [Mar82,266], in their pioneering work on vision, precisely because it

Fig. 3.1. The 2-D Mexican hat wavelet in position domain, seen in 3-D perspective (left) and gray

levels (right): (a) the isotropic wavelet; (b) the anisotropic wavelet with $ = 2.

100

Some 2-D wavelets and their performance

20 0.8

40

0.6 60

0.4 0.2

80 5

0 −5

100 0

0

120

5 −5

(a)

20

40

60

80

100

120

20

40

60

80

100

120

20 2

40

1.5 60

1 0.5 0 −15

80 5 −10

100 −5

0

0 5

10

120

15 −5

(b) Fig. 3.2. The same wavelets as in Figure 3.1, seen in spatial frequency space.

is obtained by applying an isotropic differential operator of second order to the Gaussian (this was in fact the original motivation of [266]). The Mexican hat will be efficient for a fine pointwise analysis, but not for detecting directions. This will be confirmed by quantitative calibration tests in Section 3.4 below. In addition, and for the same reasons, one may also use higher order Laplacians of the Gaussian, ψH(n) ( x ) = (− )n exp(− 12 | x |2 ).

(3.7)

For increasing n, these wavelets have more and more vanishing moments, and are thus sensitive to increasingly sharper details. An interesting technique, pioneered in 1-D by A. Arn´eodo [49], is to analyze the same signal with several wavelets ψH(n) , for different n. The features common to all the transforms surely belong to the signal, they are not artifacts of the analysis. In several applications, it is useful to introduce an additional parameter, namely, the width σ of the Gaussian (in k-space). Although σ is redundant, since the Gaussian

101

3.2 Isotropic wavelets

can be dilated to an arbitrary width, nevertheless it helps to fix explicitly the central frequency. Thus one uses, instead of (3.6), the wavelet x ) = − exp(− ψH (

σ 2 | x |2 ), 2

2 2 H (k) = |k| exp(− |k| ) . ψ 2 σ 2σ 2

(3.8)

An approximate version of the Mexican hat has been introduced by Arn´eodo et al. [Arn95,44,171], under the name of Bessel filter, namely: B (k) = 1 , for R/γ |k| R, ψ = 0 , otherwise,

(3.9)

where R and R/γ are the radii of the external and internal disks, respectively. The inverse Fourier transform ψB ( x ) is a Bessel function (hence the name): x) ψB (

1 J1 (r/γ ) J1 (r ) − 2 , r γ r/γ

r = | x |.

(3.10)

This filter was designed and used systematically for the so-called optical wavelet transform, which consists in a hardware (optical) realization of the CWT. We will come back to this application in Section 5.4.1 [Arn95,46]. Another isotropic wavelet, very similar to the previous one, has been introduced in [175] under the rather funny name of Pet hat. It is defined in Fourier space as k) = cos2 π log2 |k| , for π < |k| < 4π, ψ( 2 2π < π and |k| > 4π. = 0 , for |k|

(3.11)

This wavelet has a better resolving power in scale than the Mexican hat, hence it is more efficient in sorting objects in astrophysical images according to their characteristic scale, which is precisely the aim of the authors of [175]. Yet another one, which has the advantage of being both isotropic and continuous, is the Halo wavelet [117], defined as 2 −|ko |2 ) k) = c e−(|k| ψ( .

(3.12)

|ko |. This is a real wavelet, that selects precisely the annular region |k|

3.2.2

Difference wavelets Among the many wavelets (or filters) proposed in the literature, an interesting class consists of wavelets obtained as the difference of two positive functions, according to (3.3) in Proposition 3.1.1. In order to get an isotropic wavelet in this way, the only possibility is to take the difference between a single isotropic function φ and a contracted version of the latter, that is, the particular case where only the scale factors differ in (3.3). Indeed, if φ is a smooth non-negative function, integrable and square integrable,

102

Some 2-D wavelets and their performance

with all moments of order one vanishing at the origin, then the function ψ given by the relations ψ( x ) = α −2 φ(α −1 x) − φ( x ),

k) k) k) = φ(α − φ( ψ(

(0 < α < 1)

(3.13)

is easily seen to be a wavelet satisfying the admissibility condition (2.17). Since φ is typically a smoothing function, the wavelet ψ is called the “Difference-of-Smoothings” or DOS wavelet [Duv91]. A typical example is the “Difference-of-Gaussians” or DOG wavelet, obtained by taking for φ an isotropic Gaussian: ψD ( x) =

1 2α 2

exp(− 2α1 2 | x |2 ) − exp(− 12 | x |2 ) (0 < α < 1).

(3.14)

The DOG filter is a good substitute for the isotropic Mexican hat (for α −1 = 1.6, their shapes are extremely similar), frequently used in psychophysics works [DeV88,Duv91,124]. It was also considered by Grossmann [209] for signal analysis, together with more general linear combinations of Gaussians. An immediate application is the construction of wavelets on the 2-sphere, simply by lifting a DOG wavelet in the tangent plane to the sphere by inverse stereographic projection (Section 9.2). Another example is the “Difference-of-Mesas” or DOM filter, corresponding to a function φ which is a smoothed version of the characteristic function of a disk (a ‘mesa’ function) [369,370]. The resulting annular wavelet has been used, together with the Halo wavelet (3.12), in the detection of Einstein gravitational arcs in cosmological pictures [82] (see Section 5.1). The principle behind this application is again the filtering property: the wavelet detects preferentially objects that resemble it. More generally, the concept of the difference wavelet is useful for reducing noise in images. Take an image, consisting of an object to be identified, embedded in noise or clutter. Let φ( x ) be an averaged version of the image. Then the corresponding difference function ψ( x ) given by (3.13) is a wavelet ideally suited for the analysis of the object in question [16]. Indeed the difference operation substantially reduces the background noise, and ψ incorporates a maximal amount of resemblance with the object (a priori information). A similar subtraction technique, known in astronomy as “unsharp masking,” is commonly used in the treatment of astrophysical images for enhancing the relevant information (such as galaxies or stars), while reducing the noise. Here one computes first the high-frequency content of the image as the difference Ih = Io − Is between the original image Io and a smoothed version Is of it, and adds it to the original image, with a multiplicative factor λ. Then, in the corrected image Icorr = Io + λIh = Io + λ(Io − Is ), the high-frequency details are enhanced over the background (but severe distortions may result if λ is chosen too big) [258]. An example of unsharp masking may be found at the address http://www.chapman.edu/oca/gallery2/demo.htm. A similar procedure has been introduced for the problem of automatic target recognition (ATR). We will discuss the corresponding algorithm in Section 4.2.1 and the general problem of image denoising in Section 4.6.

103

3.3 Directional wavelets

An additional advantage of these difference wavelets is that they lead to interesting and fast algorithms, for instance, in the formalism of continuous wavelet packets [159]. Indeed, in 1-D, we have seen in (1.62), that the integrated wavelet (x) associated to the scaling function (x) is precisely (x) = 2 (2x) − (t x), and this property extends to 2-D, equation (2.144) (the L 1 normalization is used here, contrary to (3.13)). Notice, finally, that φ, and thus also ψ, need not be isotropic. The directional wavelet packets constructed in Section 2.6 are a striking example.

3.3

Directional wavelets

3.3.1

Oriented wavelets and edge detection Detecting oriented features (segments, edges, vector field, . . . ) poses a major challenge in computer vision, and many techniques have been proposed in the literature to meet it. In the context of wavelet analysis, one needs a wavelet which is directionally selective. A natural way of designing such a wavelet is to modify an isotropic one, such as the Mexican hat, simply by stretching it. Mathematically, this amounts to replacing in (3.6) x by A x, where A = diag[$ −1/2 , 1], $ 1, is a 2 × 2 anisotropy matrix. However, such a wavelet still acts as a second-order operator and detects singularities in all directions and it is of little use in practice. Indeed, on one hand, the calibration tests that we will discuss in Section 3.4 below show that it performs poorly [13]. On the other hand, there is a theorem due to Daugman [125], according to which no real wavelet rendered anisotropic by a mere stretching in one direction can have a good directional selectivity, no matter how large the anisotropy $ is taken.

3.3.1.1

Some precursors Stretching an isotropic wavelet being inefficient for inducing directional selectivity, the next step is to take a directional derivative. As we have seen in Section 3.2.1, Marr and Hildreth [266] apply the Laplacian to the Gaussian, thus getting the Mexican hat wavelet, precisely because it is an isotropic differential operator of second order. As a consequence, the Mexican hat is inefficient at detecting directions. In order to get a good edge detector, Canny [98] designed a filter which is optimal for several criteria (detection, localization, uniqueness of answer). However, this filter is numerically heavy to implement and it is advantageously replaced, to a very good approximation, by the first derivative of a Gaussian. The wavelet d −|x |2 /σ 2 ψ (1) ( e x) = (3.15) dx detects edges oriented in the y-direction, and it suffices to rotate it to get an edge detector that is sensitive to an arbitrary direction. In addition, Canny considered many different values of the width σ of the Gaussian G(x), which amounts to varying the

104

Some 2-D wavelets and their performance

scale. Furthermore, given an image s, his technique consists in locating the maxima of ψ (1) % s, which are the zero-crossings of ∂x2 G % s, i.e., precisely the technique of Mallat [261,264]. In other words, Canny was very close to a primitive version of wavelet analysis! However, for computational reasons, one is forced to sample the orientation angles, keeping only a few values. But, as remarked by Perona in 1992 [311], . . . this practice has the strong drawback of introducing anisotropies and algorithmic difficulties in the computational implementations. It would be preferable to keep thinking in terms of a continuum, of angles for example, and to be able to localize the orientation of an edge with the maximum accuracy allowed by the filter one has chosen.

Perona’s answer to this objection is to advocate the use of steerable filters, described in Section 2.7. However, taking together Canny’s approach and Perona’s remark leads directly to the continuous wavelet transform, which thus appears as a very natural tool for edge detection, and more generally in computer vision. Coming back to Canny’s work, it is instructive to compare the first, second and third derivatives of the Gaussian (see [Bha99]; Chapter 4) and test their performance, exactly as in 1-D [49]. The Canny edge detector was later improved by Deriche [135,136] and Bourennane et al. [83], still keeping the same philosophy. The former gets for the optimal filter the function f (x) = −c e−α|x| sin ωx −c x e−α|x|

for

α/ω ! 1.

(3.16)

An alternative is to consider a mixed derivative, such as ∂x ∂ y exp(−| x |2 ). This wavelet has good capabilities for detecting corners in a contour, but we will describe in Section 3.3.3 another one that performs even better, the so-called end-stopped wavelet of [Bha99,76].

3.3.1.2

The concept of directional wavelets Although the directional derivative wavelets just described do have some capabilities of directional filtering, they are by far not sufficient, because their angular selectivity is rather poor. In order to go beyond, we introduce the concept of directional wavelets [18,19,24]: Definition 3.3.1 . A wavelet ψ is said to be directional if the effective support of its is contained in a convex cone in spatial frequency space {k}, with Fourier transform ψ apex at the origin, or a finite union of disjoint such cones (in that case, one will usually call ψ multidirectional). Since it may sound counter-intuitive, this definition requires a word of justification. According to (2.20), the wavelet acts as a filter in k-space (multiplication by the function Suppose the signal s( ψ). x ) is strongly oriented, for instance a long segment along the

105

3.3 Directional wavelets

is a long segment along the k y -axis. In order to s(k) x-axis. Then its Fourier transform detect such a signal, with a good directional selectivity, one needs a wavelet ψ supported k) is essentially aligned in a narrow cone in k-space. Then the WT is negligible unless ψ( not ψ. The onto s(k): directional selectivity demands restriction of the support of ψ, corresponding standard practice in signal processing is to design an adequate filter in the frequency domain (high pass, band pass, . . . ). In addition, there are cases (magnetic resonance imaging, for instance) where data are acquired in k-space (then called the measurement space) and the image space is obtained after a Fourier transform: here again directional filtering takes place in k-space. According to this definition, the anisotropic Mexican hat is not directional, since the H is centered at the origin, no matter how big its anisotropy is. Indeed, the support of ψ detailed tests described in [13] confirm its poor performances in selecting directions. We will come back to this point, with quantitative results, in Section 3.4. A review of directional wavelets and their use may be found in [19].

3.3.2

The 2-D Morlet wavelet This is the prototype of a directional wavelet: ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) − exp(− 12 |A−1 ko |2 ) exp(− 12 |A x|2 ), √ M (k) 2 )). = $ (exp(− 1 |A−1 (k − ko )|2 ) − exp(− 1 |A−1 ko |2 ) exp(− 1 |A−1 k| ψ 2 2 2

(3.17) (3.18)

The parameter ko is the wave vector, and A = diag[$ −1/2 , 1], $ 1, is a 2 × 2 anisotropy matrix. The correction term in (3.17) and (3.18) enforces the admissibility M (0) = 0. However, since it is numerically negligible for |ko | 5.6, one condition ψ usually drops it altogether (but not always, see Section 3.4). In that case, putting $ = 1, we obtain the function ψG ( x ) = exp(i ko · x) exp(− 12 | x |2 ).

(3.19)

This function is well-known in the image processing literature under the name of Gabor function [126]. One reason of its popularity is its computational simplicity. Another one, in particular in the modeling of human vision, is that a large fraction of cells in the primary visual cortex of primates (including man, presumably) have a receptive field that resembles a Gabor function [DeV88,369,370,372]. An example is shown, for $ = 2, ko = (0, 6), in Figures 3.3 and 3.4. The Gabor function (3.19) has the qualitative behavior expected from a wavelet. It is well localized, both in position space, around the origin, and in spatial frequency space, around k = ko = 0, but, strictly speaking, it is not admissible. On the other hand, the full Morlet function given in (3.17)–(3.18) is always admissible but, for small |ko |, it is M consists then essentially of two disjoint pieces. Thus the useless as a wavelet, since ψ Morlet function is interesting for practical wavelet analysis only for |ko | large enough.

106

Some 2-D wavelets and their performance

0.08 0.06 0.04 0.02 0 −0.02 −0.04 −0.06 −0.08 0 10 20 30 40

0

5

10

15

20

25

30

35

40

(a) 1 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 0 10 20 30 40

0

5

10

15

20

25

30

35

40

(b) Fig. 3.3. The 2-D Morlet wavelet ψM ( x ) (or rather ψG ( x )), with $ = 2, ko = (0, 6) in 3-D

perspective, in position domain: (a) real part; (b) imaginary part.

The Morlet wavelet is complex. The modulus of the truncated wavelet ψG is a Gaussian, elongated in the x direction if $ > 1, and its phase is constant along the direction orthogonal to ko and linear in x, mod(2π/|ko |), along the direction of ko . Thus, plotting the phase of ψG ( x ) as a function of x, we get a succession of straight lines, perpendicular to ko , and with intensity varying periodically and linearly from 0 to 2π . As compared to the 1-D case, the additional feature here is the inherent directivity of the wavelet ψG , entirely contained in its phase. This turns out to be a crucial ingredient in the study of directional features of objects (Chapters 4 and 5). Indeed, from the fact that the WT is a convolution of the signal with the dilated wavelet, we see that the wavelet ψG smoothes the signal in all directions, but detects the sharp transitions in the direction

107

3.3 Directional wavelets

100

100

200

200

300

300

400

400

500

500 100

200

300

400

500

100

200

(a)

300

400

500

400

500

(b)

100

100

200

200

300

300

400

400 500

500 100

200

300

400

500

100

200

(c)

300

(d)

Fig. 3.4. The 2-D Morlet wavelet ψM ( x ) with $ = 2, k0 = 6 in gray levels, in position domain: (a) real part; (b) imaginary part; (c) phase; and (d) modulus.

perpendicular to ko . In Fourier space, the effective support (“footprint”) of the function M is an ellipse centered at ko and elongated in the k y direction, thus contained in a ψ √ convex cone. Since the ratio of the axes is equal to $, the cone becomes narrower as $ increases. Clearly this wavelet will detect preferentially singularities (edges) in the x direction, and its angular selectivity increases with |ko | and with the anisotropy $. The best selectivity will be obtained by taking ko parallel to the long axis of the ellipse in k-space, that is, ko = (0, ko ). The Morlet wavelet ψM (or rather ψG ) then becomes (see Figures 3.3 and 3.4): ψG ( x ) = exp(iko y) exp[− 12 ( 1$ x 2 + y 2 )],

x = (x, y).

(3.20)

Many variants of the basic wavelets may be designed for specific problems. For instance, we know that the Mexican hat is very good at detecting discontinuities (e.g., edges) in an image, but it is not directional. On the other hand, the Morlet wavelet is directional, but mostly selective in spatial frequency (these statements will be proved in Section 3.4). Both properties may be combined in a single wavelet, the Gabor (or modulated) Mexican hat wavelet, defined as follows:

108

Some 2-D wavelets and their performance

ψGM ( x ) = −($

∂2 ∂2 2 1 1 2 + ) exp(ik y) exp[− ( x + y )] , $ 1, o 2 $ ∂x2 ∂ y2

= (2 − 1$ x 2 − (y − iko )2 ) exp(iko y) exp[− 12 ( 1$ x 2 + y 2 )], GM (k) = ψ

√

$($k x2 + k 2y ) exp[− 12 ($k x2 + (k y − ko )2 )].

(3.21) (3.22) (3.23)

Notice that no correction term is needed here, the function ψGM is admissible as it stands. This wavelet, introduced in [289], is extremely efficient in detecting edges, even in the presence of heavy noise. One of its possible applications may be character recognition (see Section 4.1). Going back to the Gabor or (truncated) Morlet wavelet (3.20), we notice that it has the general form ψD ( x ) = φ(x) ψ(y),

x = (x, y),

(3.24)

where φ is a 1-D scaling function (a bump, typically a Gaussian) and ψ is a 1-D wavelet. Functions of this type provide an easy way to design a separable, yet directional wavelet; in the present case, a horizontal one. This technique is due to Bournay Bouchereau [Bou97] (see Section 5.2.2). A related example is the Gabor-like wavelet of Unser [357], obtained by replacing the Gaussian in (3.20) by another window function, typically a B-spline.

3.3.3

End-stopped wavelets A basic problem in the characterization of an image, for instance, in the comparison of two images, is the identification of specific features. In human vision, this is achieved by the so-called saccadic movement of the eyes: the eyes scan the scene freely, momentarily focus on some point of interest, and quickly move on to the next target point in the scene [Yar67]. The visual jump from one target to another is called a saccade, the target points of consecutive saccades being points of interest which stand out against the general background of the scene. This process has been analyzed in detail and modeled with wavelets by Bhattacharjee [Bha99], that we now quote. . . . There is evidence that such features of interest are identified by the lower levels of the visual system, and are not a result of conscious reasoning. Psychovisual studies on several mammals show the presence of certain cells, called hypercomplex cells or end-stopped cells, in the primary visual cortex. End-stopping behavior is related to oriented linear stimuli, that is, end-stopped cells are activated, under certain conditions, by linear stimuli having a particular orientation. Two kinds of end-stopping behavior have been identified. The single end-stopped cells respond strongly if a line in a particular orientation ends within the receptive field of the cell. For real-world scenes, these cells respond strongly to corners, or points of high curvature in general. Some other cells respond strongly only to short, oriented, linear stimuli. Such cells are called double or complete end-stopped cells. For cells of this kind, the response is strong as long as the stimulus has a specific orientation (the characteristic

3.3 Directional wavelets

orientation varies from cell to cell), and both ends of the stimulus lie within the receptive field of the cell. Thus, double end-stopped cells respond to short linear segments in images.

In order to model faithfully this physiological behavior, Bhattacharjee [76,Bha99] introduces two specific wavelets, called end-stopped wavelets, that we now describe. According to the discussion above, the responses of end-stopped cells are related to end points of linear structures lying in a specific orientation. Thus, end-stopping can be simulated by isolating linear structures in the image that have a particular orientation, and then processing these structures further to locate their end points, or to determine their lengths. The first stage, to detect lines having a specific orientation, can be achieved with a Morlet (more properly, a Gabor) wavelet. Then, the end-points of a line can be detected by applying the first derivative of a Gaussian filter along the line. Combining the two operations yields the end-stopped wavelet ψE1 , namely ψE1 ( x) =

1 4

x exp[− 14 {(x 2 + y 2 ) + ko (ko − 2i y)}],

(3.25)

in position space, and E1 (k) = exp[− 1 (k x2 + (k y − ko )2 )] −ik x exp[− 1 (k x2 + k 2y )] ψ 2 2

(3.26)

E1 is the product of in spatial frequency space. Equation (3.26) clearly shows that ψ two components. The first factor is a Morlet wavelet, with wave vector ko = (0, ko ) oriented along the k y -axis. The second factor is a first derivative of a Gaussian oriented along the frequency axis k x , that is, in the direction perpendicular to the orientation of the Morlet wavelet. On the other hand, if one regroups the two Gaussian factors in (3.26), the result √ is then simply the derivative in the k x -direction of a Morlet wavelet of width σ = 1/ 2 and wave vector 12 ko (up to a multiplicative constant). An example is shown, in spatial frequency space, in Figure 3.5.

−3 −2 −1

ky

109

0

Ky

1

2

3

−3

−2

−1

0

1

2

3

Kx

kx

(a)

(b)

E1 (k) in spatial frequency space: (a) in gray levels; (b) in 3-D Fig. 3.5. The end-stopped wavelet ψ perspective.

110

Some 2-D wavelets and their performance

As we will see in Section 4.1, the end-stopped wavelet ψE1 performs extremely well for detecting corners in an image. This can be understood intuitively as follows. AccordE1 may be interpreted as the difference between two Morlet wavelets. ing to the figure, ψ Thus an infinite rod oriented along the k y -axis will not be seen, but if we take a finite segment, the end-points will appear. On the other hand, if the rod or the segment are slightly misaligned, they will be seen. Thus the end-stopped wavelet ψE1 detects direction by a zero-crossing, hence it has a better resolution compared to the plain Morlet wavelet. If one wants to detect, in addition, the size l of a linear structure, oriented in a certain direction θ , one may use the double end-stopped wavelet, ψE2 . This wavelet indeed produces a strong response at the center of any linear structure that is oriented at θ and has a length close to l. Compared with ψE1 , the new wavelet is obtained by replacing the first derivative of a Gaussian by a second derivative, that is, a Mexican hat. The latter is used to determine the range of lengths to which the wavelet should respond. Its efficiency is easy to understand. Quoting [Bha99] again, . . . The Mexican hat is characterized by an excitatory central region surrounded by an inhibitory region. In 1-D, consider a linear stimulus of length l = l1 , and a Mexican hat filter which has an excitatory region of length l > l1 . The filter will produce the strongest response when it is exactly centered over the stimulus. Moreover, as l increases, the maximum response of the filter will initially increase. This response will increase as long as the length l l . After l exceeds l , a part of the stimulus will fall in the inhibitory region of the filter, thus damping the response of the filter. Thus, given several linear stimuli of different lengths, the Mexican hat filter responds most strongly to the stimulus which is closest to (but not greater than) the width of the excitatory portion of the filter.

Thus we obtain the required wavelet by taking, as before, a Morlet wavelet of unit width, with wave-vector ko = [0, ko ] oriented along the k y -axis, multiplied by a Mexican hat of width σ , oriented along the k x -axis, i.e., perpendicular to the orientation of the Morlet factor. This yields for the double end-stopped wavelet ψE2 the function 2σ 2 (x 2 + σ 2 + 1) σ 2 (x 2 + y 2 ) + ko (ko − 2iσ 2 y) ψE2 ( , (3.27) x) = exp − (σ 2 + 1)3 2(σ 2 + 1) E2 (k) = − 1 (k x2 − σ 2 ) exp[− 1 2 (k x2 + k 2y )] exp[− 1 (k x2 + (k y − ko )2 )]. ψ 2 2σ 8σ4 As for the pure Mexican hat, the width parameter σ allows us to control the resolving power of the wavelet. The two end-stopped wavelets, ψE1 and ψE2 , have been used successfully in [Bha99] for the detection of characteristic features in images, in the general context of image retrieval. We will discuss this application in more details in Section 4.3.

3.3.4

Conical wavelets In order to achieve a genuinely oriented wavelet, it suffices to consider a smooth function (C) (k) with support in a strictly convex cone C in spatial frequency space and behaving ψ inside C as P(k1 , . . . , kn )e−ζ ·k , with ζ ∈ C and P(.) denotes a polynomial in n variables.

111

3.3 Directional wavelets

Alternatively one may replace the exponential by a Gaussian, which gives a better localization in spatial frequency. In both cases, the resulting wavelets will be called conical. We begin with the former case, thus obtaining the class of Cauchy wavelets [18,19,24]. For simplicity, we consider a strictly convex cone, symmetric with respect to the positive k x -axis, namely C ≡ C(−α, α) = {k ∈ R2 | − α arg k α, α < π/2}, that is, the convex cone determined by the unit vectors e−α , eα . The dual cone, with sides perpendicular to those of the first one, is also convex and reads: C = C(−α, ˜ α) ˜ = {k ∈ R2 | k · k > 0, ∀ k ∈ C(−α, α)}, where α˜ = −α + π/2. Therefore e−α · eα˜ = eα · e−α˜ = 0. Given the fixed vector η = (η, 0), η > 0, we first define the (symmetric) Cauchy wavelet in spatial frequency variables: η m m −k· , k ∈ C(−α, α) m(α) (k) = (k · eα˜ ) (k · e−α˜ ) e (3.28) ψ 0, otherwise. m(C) (k) is strictly supported in the cone C(−α, α) and the parameter The Cauchy wavelet ψ ∗ on the edges of the cone, m ∈ N , m 1, gives the number of vanishing moments of ψ and thus controls the regularity of the wavelet. An interesting aspect of this wavelet is that its inverse Fourier transform may be calculated exactly [24], with the result: ψm(α) ( x) =

(−1)m+1 (sin 2α)2m+1 (m!)2 $ %m+1 , 2π z · σ (α)z

(3.29)

where we have introduced the complex variable z = x + i η ∈ R2 + i C and the 2 × 2 matrix cos2 α 0 σ (α) = . 0 − sin2 α Indeed, from the definition (3.28), we get: 1 (α) x) = d 2 k ei k·x (k · eα˜ )m (k · e−α˜ )m e−k·η ψm ( 2π C(−α,α) (−1)m x ]m [e−α˜ · ∇ x ]m = d 2 k e−k·(η−i x) . [eα˜ · ∇ 2π C(−α,α) The integral on the right-hand side is convergent, since k · η > 0. Write ξ = η − i x = −i( x + i η) and let A be the matrix that maps the unit vectors e1 , e2 onto e−α , eα , respectively: (eν )i = Aν j (e j )i , ν = ±α (where we use the usual summation convention), so that k j = Aν j k ν (contravariant coordinates) and ξν = Aν j ξ j (covariant coordinates). Explicitly, we have:

112

Some 2-D wavelets and their performance

A=

cos α cos α − sin α sin α

,

so that

det A = sin 2α.

In the new (nonorthogonal) coordinates, the cone becomes C(−α, α) = {k ∈ R2 : k ν 0, ν = ±α}, and the integral may be evaluated immediately: ξ 2 −k· d ke = dk 1 dk 2 exp(−Aν j k ν ξ j ) C(−α,α) C(−α,α) ∞ ∞ dk α dk −α exp(−k ν ξν ), = det A 0

0

det A = ξα ξ−α sin 2α = (eα · ξ ) (e−α · ξ ) − sin 2α . = [( x + i η) · eα ] [( x + i η) · e−α ] Then the result follows by differentiation, if one remembers that eα˜ · e−α = e−α˜ · eα = 0, eα˜ · eα = e−α˜ · e−α = sin 2α. Indeed: x ) (eα˜ · ∇

eα˜ · e−α 1 = = 0, [( x + i η) · e−α ] [( x + i η) · e−α ]2

x )m (eα˜ · ∇

(−1)m m! (eα˜ · eα )m 1 = [( x + i η) · eα ] [( x + i η) · eα ]m+1 =

(−1)m m! (sin 2α)m , [( x + i η) · eα ]m+1

and similarly for the other factor. Thus one obtains as the final result the function ψm(α) given in (3.29). Clearly, this function is square integrable, and admissible in the sense of (6.15), in other words, it is a wavelet. We show in Figures 3.6 and 3.7 various aspects of the Cauchy wavelet ψ4(10) . This is manifestly a highly directional filter, strictly supported in the cone C = C(−10◦ , 10◦ ). Notice the slow decay in x-space of the conical wavelet ψm(α) , independent of α. This is the price to pay for forcing the wavelet to be strictly supported in a cone, and is to be expected in the light of standard results on the localization properties of wavelets (theorems of Balian–Low and Battle) [Fei98,36]. However, as we have already stressed in Section 3.3.1.2, this is irrelevant for the analysis, only the behavior in k-space counts. The construction generalizes in a straightforward way [24] to an arbitrary convex ˜ = {k ∈ R2 , k · cone C ≡ C(α, β) = {k ∈ R2 | α arg k β}, with dual C ≡ C(α, ˜ β)

k > 0, ∀ k ∈ C(α, β)}, where β˜ = α + π/2, α˜ = β − π/2, and arbitrary vanishing

113

3.3 Directional wavelets

(10)

Fig. 3.6. The Cauchy wavelet ψ4 : (a) real part; (b) imaginary part; (c) modulus, all in position

domain; and (d) modulus in frequency domain.

the resulting wavelet reads moments l, m > 0 on the cone edges. For any fixed η ∈ C, in spatial frequency space as (k · eα˜ )l (k · eβ˜ )m e−k·η , k ∈ C(α, β); (C) (3.30) ψlm (k) = 0, otherwise, and in position space,

114

Some 2-D wavelets and their performance

(10)

Fig. 3.7. The Cauchy wavelet ψ4 , in gray levels, in position domain: (a) real part; (b) imaginary

part; (c) phase; and (d) modulus. (C) ψlm (x) = const. (z · eα )−l−1 (z · eβ )−m−1 ,

(3.31)

where again z denotes the complex variable z = x + i η ∈ R2 + i C. Actually the origin of the name “Cauchy” is the following example. For α = 0, β = π/2, η = eπ/4 and m = 1, one gets: ψ1(C) (x) =

1 (1 − i x)−2 (1 − i y)−2 , 2π

(3.32)

i.e., the product of two 1-D Cauchy wavelets [Hol95]; that is, derivatives of the Cauchy kernel (z − t)−1 . Of course, this example is of little use in practice. Indeed, the main interest of Cauchy wavelets is their good angular selectivity, which requires a narrow

115

3.3 Directional wavelets

20

20

15

15

10

10

5

5

0

0

−5

−5

−10

−10

−15

−15

−20 −20

−15

−10

−5

0

5

10

15

20

−20 −20

−15

−10

−5

0

5

10

15

20

(b)

(a)

M Fig. 3.8. Two directional 2-D wavelets, in k space, seen in level curves: (a) the Morlet wavelet ψ ◦

4(10 ) with supporting cone C10 = C(−10◦ , 10◦ ), (ko = (0, 6), $ = 5); (b) the Cauchy wavelet ψ ◦ rotated by 90 for the sake of comparison.

cone. For applications, it turns out that the wavelet ψ4(10) , with support in the cone C10 = C(−10◦ , 10◦ ) has properties very similar to those of the Morlet wavelet (3.17) with |ko | = 5.6, except that here the opening angle of the cone is totally controllable. For a Morlet wavelet, on the contrary, the cone gets narrower for increasing |ko |, but then the amplitude decreases as exp(−|ko |2 ). In that sense, Cauchy wavelets are better adapted. M with ko = (0, 6), $ = 5 (left) We show side by side in Figure 3.8 the Morlet wavelet ψ (10) ◦ (k), rotated by 90 for the sake of comparison (right). and the Cauchy wavelet ψ 4 Quantitative comparisons will be made in Section 3.4. In 1-D, a wavelet ψ is called progressive or a Hardy function [Hol95,205], if ψ(ω) = 0 for ω < 0. This in turn may be expressed in terms of the Hilbert transform, defined by H f (ω) = −i sign ω f (ω), namely ψ = (1 + i H )φ, φ ∈ L 2 (R, dt) (that is, ψ is the analytic signal associated to φ). Equivalently, ψ belongs to the Hardy space H+2 (R) of square integrable functions which extend analytically into the upper half-plane. We claim that the conical wavelets are the 2-D analogs of this concept; that is, the genuine 2-D progressive wavelets. In order to prove that statement, we first notice that the convex cone C(α, β) may ν = α, β: also be expressed in terms of the covariant coordinates kν˜ = (eν˜ · k), C(α, β) = {k ∈ R2 : kα˜ 0, kβ˜ 0}. Consider the directional Hilbert transforms:

(3.33)

116

Some 2-D wavelets and their performance

H ν˜ f (k) = −i sign kν˜ f (k).

(3.34)

Given φ ∈ L 2 (R2 , d 2 x), define the function ψ = (1 + i Hα˜ + i Hβ˜ − Hα˜ Hβ˜ )φ = (1 + i Hα˜ )(1 + i Hβ˜ )φ.

(3.35)

k) vanishes outside the cone C(α, β), and Then it is easy to see, as in [366], that ψ( indeed: k) = 4φ(k), k ∈ C(α, β), ψ( (3.36) 0, otherwise. Therefore the inverse Fourier transform ψ( x ) is the boundary value of a function ψ(z ) 2 holomorphic in the tube R + i C; i.e., a 2-D Hardy function. For a fixed convex cone 2 C(α, β), the set of all such functions constitutes a Hilbert space, naturally denoted H(α,β) , which is unitary equivalent, via the complex Fourier transform, to the space β), d 2 k) [Ste71; Theorem VI.3.1]. In that sense, conical wavelets are a genuine L 2 (C(α, multidimensional generalization of the 1-D Hardy functions, much more so than the so-called 2-D Hardy functions defined by Dallard and Spedding (in particular, their “Arc” wavelet) [117]. Among them Cauchy wavelets are particularly simple (they occupy a special niche, as we will see in Section 8.2). In conclusion, notice that we are 2 talking here of a Hardy space H(α,β) , but similar considerations may be made for Hardy 1 spaces H(α,β) , in terms of the Riesz operators, which are a natural multidimensional generalization of the Hilbert transform [366]. Cauchy wavelets have a good angular selectivity, provided one chooses a narrow cone. However their radial selectivity is not terribly good, because the exponential decays → ∞. In order to obtain a better radial localization, one may replace rather slowly as |k| the exponential by a Gaussian in√k x [Vdg98]. This has the effect to concentrate the wavelet on its central frequency ( 2m, 0). The resulting wavelet is called the Gaussian conical wavelet. The angular selectivity of ψ is specified by the angular aperture of the cone and is well controlled by the parameter α. However, the radial selectivity is only roughly fixed by the moment number m. The resulting wavelet is very similar to the Cauchy wavelet, except that it is more concentrated in spatial frequency space, since it is also localized in scale, around the central scale ao . However, although the pure Gaussian is well peaked, the addition of a large number of vanishing moments tends to spread it. Thus, one can achieve an even better scale localization by using an appropriate width σ > 0 for the Gaussian. We may still improve on this by adding another parameter χ(σ ) > 0, whose sole rˆole is to control the radial support of ψ [26,27]. This leads to the following formula for our conical wavelet: σ 2 (k · e−α˜ )m (k · eα˜ )m e− 2 (kx −χ(σ )) , k ∈ C(−α, α), ψC (k) = (3.37) 0, otherwise,

117

3.3 Directional wavelets

Fig. 3.9. The Gaussian conical wavelet (3.37), in frequency space, for m = 4, α = 10◦ , σ = 1.

√ −1 where χ(σ ) = 2m σ√ is called the center correction term. Notice that the central σ frequency is the point ( 2m, 0) for any σ , and for σ = 1, one recovers the pure Gaussian conical wavelet. This is the wavelet we will mostly use in the sequel. It is shown in frequency domain, in Figure 3.9. It is clear on this picture why this wavelet is sometimes called the shark wavelet! Another alternative is the conical Mexican hat, introduced by Murenzi et al. [289]. and l, m ∈ N ∗ , as The wavelet with support in C = C(β) is defined, for any η ∈ C (C) (k) ψ lm

=

1 2 2 (k · e−β˜ )m (k · eβ˜ )l ($k x2 + k 2y )e− 2 ($kx +k y ) , 0, otherwise.

k ∈ C(β)

(3.38)

Besides these directional wavelets, there exist two other tools especially designed for the detection of lines or curves, called ridgelets and curvelets, respectively. These actually define new transforms, that we will discuss later, in Section 11.1.

3.3.5

Multidirectional wavelets Given a directional wavelet ψ, as above, it is easy to build a multidirectional one, with n-fold symmetry simply by superposing n suitably rotated copies of ψ: x) = ψn (

n−1 1 2π ψ(r−θk ( x )), θk = k , k = 0, 1, . . . , n − 1. n k=0 n

(3.39)

118

Some 2-D wavelets and their performance

Taking, for instance, n = 4 and for ψ a Gabor (truncated Morlet) wavelet, we get the following real wavelet with four-fold symmetry: ψ4M (x, y) =

1 1 2 2 (cos ko x + cos ko y) e− 2 (x +y ) . 2

(3.40)

This wavelet filters out all features which are not primarily horizontal or vertical. In the same way, one gets wavelets with symmetry 6 or 10, which may find applications, respectively, in biological problems or the analysis of quasicrystals. In general, multidirectional wavelets should be useful for pattern recognition. Notice that a similar construction was proposed by Watson [369,370]. His fan filters are obtained by taking first the difference between two “mesa” functions, which yields an annular wavelet, and then repeatedly bisecting the spatial frequency space and taking only one side (i.e., the associated analytic signal). The allowed directions θ are thus restricted to a fan-shaped region: 0 2θ

2π (n = 2, 3, . . . ). 2n−1

(3.41)

This construction may then be generalized to arbitrary angles [306]. Actually, a comparison with Section 2.6 immediately shows that the construction of directional wavelet packets is based on the very same idea. These fan filters have all the properties of directional wavelets, including admissibility in the form (2.17). Applying on these filters discrete rotations and scaling, Watson builds a pyramid of oriented filters as a tool for data compression and signal reconstruction after coding, in a model of human vision. This is in fact a discretized version (in polar geometry) of the CWT. Another example, very similar to the previous one, is that of the steerable filters, described in Section 2.7.

3.4

Wavelet calibration: evaluating the performances of the CWT Given a wavelet, what is its angular and scale selectivity (resolving power)? What is the minimal sampling grid for the reconstruction formula (2.26) that guarantees that no information is lost? The answer to both questions resides in a quantitative knowledge of the properties of the wavelet, that is, the tool must be calibrated. To that effect, one takes the WT of particular, standard signals. Three such tests are useful, and in each case the outcome may be viewed either at fixed (a, θ) (position representation) or at fixed b (scale-angle representation). r Point signal: for a snapshot at the wavelet itself, one takes as the signal a delta function, i.e., one evaluates the impulse response of the filter: ψa,θ,b |δ = a −1 ψ(a −1r−θ (−b)).

(3.42)

119

3.4.1

3.4 Wavelet calibration: evaluating the performances of the CWT r

Reproducing kernel: taking as the signal the wavelet ψ itself, one obtains the reproducing kernel K , which measures the correlation length in each variable a, θ, b : −1 ψ( cψ K (a, θ, b|1, 0, 0) = ψa,θ,b |ψ = a ψ(a −1r−θ ( x − b)) x ) d 2 x. (3.43)

r

Benchmark signals: for testing particular properties of the wavelet, such as its ability to detect a discontinuity or its angular selectivity in detecting a particular direction, one may use appropriate “benchmark” signals.

The scale and angle resolving power Suppose the wavelet ψ has its effective support in spatial frequency in a vertical cone in the x and y directions of aperture ϕ, corresponding to ko = (0, ko ). The width of ψ is given by 2wx and 2w y , respectively: !1/2 !1/2 1 1 k)| k)| 2 2 d 2 k k x2 |ψ( d 2 k (k y − ko )2 |ψ( , wy = . wx = ψ ψ (3.44) is concentrated in an ellipse of semi-axes wx , w y , and its radial Then the wavelet ψ ko + w y . Thus the scale width or scale resolving power support is ko − w y |k| (SRP) of ψ is defined as: S R P(ψ) =

ko + w y . ko − w y

(3.45)

In the same way, one defines the angular width or angular resolving power (ARP) by considering the tangents to that ellipse. Then a straightforward calculation yields: ) ko2 − w 2y −1 A R P(ψ) = 2 cot ϕ. (3.46) wx For instance, if ψ is the (truncated) Morlet wavelet (3.17), one obtains: √ ) ko 2 + 1 , A R P(ψM ) = 2 cot−1 $(ko2 − 1), S R P(ψM ) = √ ko 2 − 1

(3.47)

and, for ko ! 1: √ A R P(ψM ) = 2 cot−1 (ko $).

(3.48)

This last expression coincides with the empirical result of [13]: the angular sensitivity √ of ψM depends only on the product ko $. Notice also that the SRP is independent of the anisotropy factor $. If ψ is the Cauchy wavelet (3.28) with support in the cone C(−α, α), the ARP is simply the opening angle 2α of the supporting cone.

120

Some 2-D wavelets and their performance

Fig. 3.10. Filter bank obtained with the Morlet wavelet (5 scales, 16 orientations).

3.4.2

The reproducing kernel and the resolving power of the wavelet A natural way of testing the correlation length of the wavelet is to analyze systematically its reproducing kernel. Let the effective support of the wavelet ψ in spatial frequency be, in polar coordinates, ρ and ϕ. Then an easy calculation [18] shows that the effective support of K is given by amin = ( ρ)−1 a amax = ρ for the scale variable, and − ϕ θ ϕ for the angular variable. Thus we may define the wavelet parameters (or resolving power) ρ, ϕ in terms of the parameters a, θ of K , as: √ r scale resolving power (SRP): ρ = a = √a max /amin ; r angular resolving power (ARP):

ϕ = 12 θ. a j ,θ (k)}, which yields a complete In this way, one may design a wavelet filter bank {ψ tiling of the spatial frequency plane, in polar coordinates [15,18]. An example is given in Figure 3.10, for the case of the Morlet wavelet. Clearly this analysis is only possible within the scale-angle representation. Thus it requires the use of the CWT, and it is outside of the scope of the DWT, which is essentially limited to a Cartesian geometry (see Section 2.6).

3.4.3

Calibration of a wavelet with benchmark signals The capacity of the wavelet at detecting a discontinuity may be measured on the (benchmark) signal consisting of an infinite rod (see [13] for the full discussion). The result is that both the Mexican hat and the Morlet wavelet are efficient in this respect. For testing the angular selectivity of a wavelet, one computes the WT of a semi-infinite rod, sitting along the positive x-axis, and modeled as usual with a delta function: s( x ) = ϑ(x) δ(y),

(3.49)

121

3.4 Wavelet calibration: evaluating the performances of the CWT

1 θ = 0° 0.9

0.8

θ = 5°

0.7

0.6

0.5

0.4

0.3

0.2

θ = 10°

0.1 θ = 90° 0 −5

0

θ = 20° 5

10

Fig. 3.11. Angular selectivity of the Morlet wavelet for different values of the orientation angle: θ = 0◦ , 5◦ , 10◦ , 20◦ , 45◦ and 90◦ . The graph shows the modulus |S((bx , 0), 1, θ)| as a function of bx .

where ϑ(x) is the step function. Plugging this expression into the definition of the transform yields (we take a = 1 for simplicity): +∞ 1, θ ) = d x ψ r−θ (x − bx , −b y ) . (3.50) S(b, 0

Let us take first a Morlet wavelet with $ = 5, oriented at an angle θ, and compute the CWT of s as a function of bx . As illustrated by Figure 3.11, the result is that ψM detects the orientation of the rod with a precision of the order of 5◦ . Indeed, for θ < 5◦ , the WT is a “wall,” increasing smoothly from 0, for x −5, to its asymptotic value (normalized to 1) for x 5. Then, for increasing misorientation θ, the wall gradually collapses, and essentially disappears for θ > 10◦ . Only the tip of the rod remains visible, and for large θ (θ > 45◦ ), it gives a sharp peak. Essentially the same result is obtained with a Cauchy wavelet supported in the cone C(−10◦ , 10◦ ), of opening angle A R P = 20◦ , as shown in Figure 3.12(a). Conversely, one sees in panel (b) that, for a fixed misorientation angle θ = 20◦ , the Cauchy wavelet yields the same selectivity for A R P 20◦ . On the contrary, as observed on Figure 3.13, the same test performed with an anisotropic Mexican hat gives a result almost independent of θ. Even varying the anisotropy factor $ doesn’t really change the result: the discontinuity is detected by a sharper variation, but the sensitivity to its orientation is not greatly improved. The conclusion is that the Morlet and the Cauchy wavelets are highly sensitive to orientation, but the anisotropic Mexican hat is not.

122

Some 2-D wavelets and their performance

| S(b ,1,20°) |

| S(bx ,1,θ) | 1

x

misorientation angle

θ = 20

°

1 Φ = 30 °

θ = 0°

0.9

0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

Φ = 25 °

0.2

0.2

θ = 20 °

0

Φ = 10 °

0.1

0.1

−5

θ = 90 0

°

Φ = 20 °

θ = 45 ° 5

bx

(a)

10

0

−5

0

5

(b)

bx

10

Fig. 3.12. Angular selectivity of the Cauchy wavelet ψ1,1 : (a) for a cone of fixed width = 2α = 20◦ and for different values of the orientation angle: θ = 0◦ , 5◦ , 10◦ , 20◦ , 45◦ and 90◦ ; (b) for a fixed value of the misorientation angle θ = 20◦ and various values of the ARP = 2α. The graph shows the modulus |S((bx , 0), 1, θ )|, respectively |S((bx , 0), 1, 20◦ )|, as a function of bx . 1

0.9

1 ε=5

θ = 0°

θ = 90°

0.9

ε = 2.5

θ = 45°

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0 −5

0

5

(a)

ε=1

10

0 −5

0

5

10

(b)

Fig. 3.13. Angular selectivity of the anisotropic Mexican hat wavelet with: (a) anisotropy $ = 5 and

for different values of the orientation angle (θ = 0◦ , 45◦ and 90◦ ) and (b) for a fixed orientation (θ = 0◦ ) but various anisotropy factors ($ = 1, 2.5 and 5).

Let now the signal be a segment. If one uses a Morlet or a Cauchy wavelet as above, the WT reproduces the segment if the misorientation φ between the signal and the wavelet is smaller than 5◦ , but the segment becomes essentially invisible for

φ > 15◦ , except for the tips (these are point singularities). In the end, the image of the segment reduces to two peaks corresponding to the two endpoints (see Figure 3.14).

123

3.4 Wavelet calibration: evaluating the performances of the CWT

1 θ = 0° 0.9

0.8 θ = 90°

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 −5

−4

−3

−2

−1

0

1

2

3

4

5

Fig. 3.14. Modulus of the wavelet transform of a segment positioned at θ = 0◦ , using a Morlet wavelet for several values of the orientation angle θ.

This is exactly the property used crucially in the measurement of the velocity field of a turbulent fluid (see Section 5.3.3 below). One may remark that the precision mentioned here is obtained with the modulus of the WT. In fact, if the wavelet is complex (like ψM ), one may also exploit the phase of the WT, and it gives a higher precision yet [13]. But this is practical only on academic signals, real data are in general too noisy and only the modulus is useful. Another way of comparing the angular selectivity of the two wavelets is to analyze a directional signal in the angle–angle representation (α, θ ) described above. The result confirms the previous one [17]. In order to illustrate the difference in angular selectivity between the anisotropic Mexican hat and the Morlet wavelet, we analyze a directional signal with both of them and view the transform in the angle–angle representation described in Section 2.3.3. The result is shown in Figure 3.15. The signal is a rectangular slab of size 3 × 2, positioned radially at π/2 and it is analyzed with an anisotropic Mexican hat with $ = 2 (left) and a Morlet wavelet (right). The figures show the modulus of the CWT, at range = 3 and scale a = 1, in the angle–angle representation (α, θ ). For the Mexican hat, |b| the transform exhibits a maximum located at α = π/2, θ = π/2, as it should, and the graph is periodic both in α and in θ. For the Morlet wavelet, several maxima can be distinguished, all located around α = π/2. A careful inspection shows that the strongest peaks at θ = π/2 and 3π/2 correspond to the two longest edges of the rectangle while the smaller peaks at θ = 0, π, 2π correspond to the smaller edges. Notice however that, for the Mexican hat, the peak is sharp in α and quite broad in θ , as expected, since this wavelet has a very good resolution in position (α), but a rather poor one in directional

124

Some 2-D wavelets and their performance

6

0.4

5 0.3 4

3

0.2

2 0.1 1

0

6

0

6

6 6

4

4

4

4

0

(a)

0

(b) 6

5

5

4

4

α

α

6

3

3

2

2

1

1

0

θ

2

2

θ

0

0

α

α

2

2

0

1

2

3

θ (c)

4

5

6

0

0

1

2

3

θ (d)

4

5

6

Fig. 3.15. CWT of a slab of size 3 × 2 positioned radially at 90◦ , analyzed with an anisotropic

Mexican hat (left) with $ = 2 and a Morlet wavelet (right), both at scale a = 1. The figures in the top row give a 3-D perspective view of the transform, in the angle–angle representation (α, θ), at = 3. The same result is shown in level curves in the figures of the bottom row. range |b|

selectivity (θ). In addition, the contour of the slab can be precisely detected by tracking the zero crossings of this representation. On the opposite, the Morlet wavelet gives a sharp peak in both variables, since it has good selectivity in both [13], but the contour is less visible. This example illustrates Daugman’s theorem [126], and at the same time the usefulness of the angle–angle representation.

4

Applications of the 2-D CWT. I: image processing

The 2-D CWT has been used by a number of authors, in a wide variety of problems [Com89,Mey91,Mey93]. In all cases, its main use is for the analysis of images, since image synthesis or compression problems are rather treated with the DWT. In particular, the CWT can be used for the detection or determination of specific features, such as a hierarchical structure, edges, filaments, contours, boundaries between areas of different luminosity, etc. Of course, the type of wavelet chosen depends on the precise aim. An isotropic wavelet (e.g. a Mexican hat) often suffices for pointwise analysis, but a directional wavelet (e.g. a Morlet or a conical wavelet) is necessary for the detection of oriented features in the signal. Somewhat surprisingly, a directional wavelet is often more efficient in the presence of noise. In the next two chapters, we will review a number of such applications, including some nonlinear extensions of the CWT. First, in the present chapter, we consider various aspects of image processing. Then, in Chapter 5, we will turn to several fields of physics where the CWT has made an impact. Some of the applications are rather technical and use specific jargon. We apologize for that and refer the reader to the original papers for additional information.

4.1

Contour detection, character recognition

4.1.1

The detection principle Exactly as in the 1-D case, the WT is especially useful to detect discontinuities in images, for instance the contour [Mur90,13] or the edges of an object (which are discontinuities in the luminosity), and in particular its corners [204,266]. For that purpose, one may first ignore the directions and perform a pointwise analysis. Then, the simplest choice is an isotropic wavelet, such as the radial Mexican hat ψH given in (3.6) or (3.8). In that particular case, the effect of the WT consists in smoothing the signal with a Gaussian and taking the Laplacian of the result. Thus large values of the amplitude will appear at the location of the discontinuities, in particular the contour of objects. In particular, the corners of the contour, which are point singularities, will be highlighted. In addition, if

125

126

Applications of the 2-D CWT. I: image processing

the wavelet is real, the CWT detects the convexity of each corner: a convex corner gives rise to a sharp positive peak, whereas a concave one yields a negative peak. However, it turns out that other wavelets are useful too. For instance, in the presence of heavy noise, directional wavelets outperform the Mexican hat. On the other hand, if only the corners of the contour are needed, as in character recognition, mixed derivatives of the Gaussian or, even better, the end-stopped wavelets become the first choice. In order to test these properties on a concrete example, we begin by analyzing a simple geometric object, namely, a set with the shape of the letter L, represented by its characteristic function. Thus our test image is the white L-shaped region against a black background, the signal presented in Figure 4.1(a) or, equivalently, in Figure 4.2(a). First, we analyze the effect of scaling on the WT, i.e., going to finer and finer scales. For a pointwise analysis, we choose the isotropic Mexican hat ψH given in (3.6) [13]. The CWT is plotted in Figure 4.1 for three values of the scale parameter a (conveniently taken as powers of 2), a = 2 j , j = 3, 2, 1. Each WT is plotted both in 3-D perspective and in level curves. From these pictures, the following observations can be made. For a large value of a, the WT sees only the object as a whole, thus allowing the determination of its position in the plane. When a decreases, increasingly finer details appear. In this simple case, the WT vanishes both inside and outside the contour, since the signal is constant there. Eventually, only the contour remains and it is perfectly seen at a = 2. This is the analog of the precise localization of discontinuities in 1-D [204]. Of course, if one takes values of a that are too small, numerical artifacts (aliasing) appear and spoil the result. This is only a numerical limitation, however, that could be improved by a finer discretization (but with a longer computing time). We notice that the exterior contour is a sharp negative “wall,” whereas the interior contour is a positive one. The same effect would appear in 1-D if one would consider, for instance, the full WT of a square pulse. The jump from 0 to 1 gives a negative minimum followed by a sharp positive maximum, and the jump from 1 to 0 gives the opposite pattern. Note also that the corners of the figure are highlighted in the WT by sharp peaks. The amplitude is larger at these points, since the signal is singular there in two directions, as opposed to the edges. In addition the WT detects the convexity of each corner. The six convex corners give rise to positive peaks, whereas the concave one yields a negative peak. Here we see again the advantage of using a real wavelet and plotting the WT itself, not its modulus, which is a frequent practice. The conclusion is that the CWT is an efficient edge detector, provided it is evaluated at a scale that is sufficiently small (i.e., high spatial frequency), but still avoiding numerical artifacts (aliasing). The next step is to compare the performances of various wavelets on the same signal, and the L-shape is an ideal benchmark for making the comparison. The result is shown in Figure 4.2 (taken from [Bha99]). Panel (a) gives the signal. Panel (b) is the response of the isotropic Mexican hat wavelet (this is in fact the negative of Figure 4.1(d)!). As noted there already, the response of the wavelet is strongest at the corners, but is also

127

4.1 Contour detection, character recognition

(a)

4.1(a)

4.1(b)

(b)

4.1(b)

4.1(c)

(c)

4.1(c)

(d) Fig. 4.1. The L-shape and its CWT, obtained with an isotropic Mexican hat, presented in 3-D

perspective (left column) and in level curves (right column), at three successive scales: (a) the signal; (b) a = 8; (c) a = 4; (d) a = 2.

128

Applications of the 2-D CWT. I: image processing

(a)

(b)

(c)

(d)

Fig. 4.2. Comparison between the different wavelets applied to the L, all at scale a = 2: (a) the test

image; (b) response of the isotropic Mexican hat wavelet; (c) response of the 2-D Morlet wavelet with |k0 | = 6, θ = 0◦ ; (d) response of the end-stopped wavelet ψE1 at θ = 0◦ (from [Bha99]).

fairly strong at other points along the edges. Panel (c) is the result of applying a 2-D Morlet wavelet, oriented horizontally (θ = 0◦ ), which detects only the vertical edges, as it should (it is an efficient directional filter, as we will see below). Notice the horizontal “leaking” of the response: this is an artifact due to the use of a scale that is too small for the signal (so that the wavelet gets too wide in spatial frequency space). Finally, panel (d) is the response of the end-stopped wavelet ψE1 , again oriented horizontally. This wavelet responds only to endpoints of vertical edges, which in this case are all the corners of the contour. Unlike the previous wavelets, however, it shows no response to other points along the edges.

129

4.1 Contour detection, character recognition

(a)

(b)

(c)

(d)

Fig. 4.3. Analysis of the L-shape with the double end-stopped wavelet ψE2 with |k0 | = 6, θ = 90◦ :

(a) response of the wavelet of width σ = 12; (b) local maxima of (a), thresholded at 50%, superimposed (white crosses) on the input image. The wavelet identifies the short horizontal edges; (c) and (d) the same analysis for the same wavelet of width σ = 36. Now long horizontal edges are selected (from [Bha99]).

In a last case, we analyze the L-shape with the double end-stopped wavelet ψE2 , and the result is shown in Figure 4.3, again taken from [Bha99]. This wavelet is sensitive both to the orientation and to the size of the edges, and its behavior may be understood as follows. The Morlet component of the filter selects linear structures which are perpendicular to the orientation of the wavelet, i.e., here horizontal segments. Then the Mexican hat component, which operates parallel to the linear structures detected by the Morlet component, produces strong responses only for those structures that approach

130

Applications of the 2-D CWT. I: image processing

the characteristic length of the Mexican hat (which is determined by σ ). Thus, the ψE2 wavelet detects linear structures of given size and orientation. In our example, the ψE2 wavelet is oriented vertically, using two different values of σ . Figures 4.3(a) and (b) show the response of the wavelet with σ = 12, first the raw response, then the local maxima of (a), thresholded at 50% and superimposed (white crosses) on the input image. The wavelet identifies the short horizontal edges. Similarly, Figures 4.3(c) and (d) show the raw response and the corresponding thresholded local maxima, respectively, for the wavelet with σ = 36. For this value of σ , the strongest response of the wavelet corresponds to the longer horizontal edge in the input image.

4.1.2

Application to character recognition We will now apply the technique developed in the previous section to the problem of character recognition. Our signal will be a set of simplified characters, modeled by the corresponding characteristic function. At this stage of our investigation, we only consider characters composed of segments or union of rectangles. Let us consider for instance a few simple characters (Figure 4.4). We see that r the L has six corners: five convex and one concave, r the A has 12 corners: six convex and six concave, r the E has 12 corners: eight convex and four concave, r the H has 12 corners: eight convex and four concave. The interesting point is that, in this case, the number of concave corners and convex corners completely characterizes these letters. Therefore, an automatic recognition of these characters requires a fast algorithm for extracting this particular information and encoding it. In what follows we will propose a first step for designing such an algorithm, based on the 2-D CWT. In order to follow the construction in detail, we focus on a thick letter A, represented by its characteristic function. Of course, this object behaves exactly as the academic

Fig. 4.4. A set of simplified characters: the letters L, A, E, H.

131

4.1 Contour detection, character recognition

10

5

0

−5

−10 0 0

200

200

400

400 600

Y

600 X

(a)

1

1

1

-1

-1

-1 -1

-1 -1 1 1

1 (c)

(b) Fig. 4.5. Detecting the contour of the letter A with the radial Mexican hat: (a) the CWT at a = 0.075, in 3-D perspective; (b) the same, in level curves; (c) coding of the same by the signs of the respective corners.

image of the preceding section. If one works again with a radial Mexican hat, and goes down to a sufficiently small scale, the CWT reveals the contour of the letter. Moreover, the corners of the figure are highlighted in the WT by sharp peaks, the sign of which is determined by the convexity of the corresponding corner, since the wavelet is real. The result is shown in Figure 4.5. In panel (a), we show the WT of the letter at scale a = 0.075, in 3-D perspective. Here we see clearly the twelve peaks corresponding to each corner, some positive (for the six convex corners), some negative (for the six concave corners). Panel (b) presents the same result in level curves, and panel (c) shows the coding of the corners by a logical flag (± 1 for concavity or convexity).

132

Applications of the 2-D CWT. I: image processing

This exercise leads to an algorithm for automatic character recognition [17]. The basic idea of the method is to treat only the significant parts of the signal, focusing on the information needed for unambiguous recognition. In the case of simple letters, this information is entirely contained in the high-frequency components, namely corners and edges. Take, for example, the letter A of Figure 4.5. It can be entirely characterized by the succession of its 12 corners and the additional information that consists in deciding whether a corner is concave or convex. That is, twelve points and a logical flag (concavity or convexity) for each point. The following simple algorithm achieves this treatment. It consists in locating the local maxima of the CWT and eliminating everything else by thresholding, and it is able to detect an A unambiguously. r Compute the CWT S(b, a f ) in position representation with the Mexican hat wavelet at the finest relevant scale a = a f . The transform exhibits local extrema at the corners and is positive (negative) for a convex (concave) corner. Compute the absolute extrema of the transform a f )}, m(a f ) = min{S(b, b

r

a f )}. M(a f ) = max{S(b, b

To get rid of the other high-frequency components, threshold the transform using a negative value T− > m(a f ) and a positive value T+ < M(a f ), both directly computed from the CWT. All the values between T− and T+ are set to zero (in the terminology of Donoho [146,147], this is a hard thresholding, see Section 4.6). r We are left with an image, denoted by T S(b, a f ), composed of positive and negative peaks at the position of the corners, which we encode as a vector with components +1 or −1 depending on the local sign of the (thresholded) transform. Using this simple technique we are able to deal with simple shapes, especially with characters that are not corrupted with noise. In the case where we have additive noise or if we need to be more accurate, we use the same treatment at a different small scale, $ % a j )}( j=1,...,N ) . that is a ∈ amin , (amax /amin )1/2 . We obtain a sequence of images {T S(b, Adding these images together gives an image from which one encodes again the local maxima and minima, which are now enhanced against the background noise. The scheme of this algorithm is displayed in Figure 4.6 and applied for recognition in Figure 4.7, for the (noiseless) case of the four letters of Figure 4.4. Panel (a) shows the CWT of the signal with a Mexican hat wavelet, at scale a = 1.5. Panel (b) gives the skeleton of that CWT, thresholded at 90% of the maximal values; as expected, only the top of the peaks survive, and they are shown with crosses for the positive peaks (maxima, convex corners) and circles for negative ones (minima, concave corners). Needless to say, this algorithm works only for a real wavelet and with the values of the CWT itself, not its absolute value. Actually, since only the corners are needed, we may as well use a wavelet that sees only the corners, not the edges. Typically, a directional wavelet (when it is misaligned),

133

4.1 Contour detection, character recognition

Signal

Wavelet ❅ ❘ ❅

✠

aj) S(b,

$ % a j ∈ amin , (amax /amin )1/2

❄

Thresholding

❄

aj) T S(b,

❄

j

aj) T S(b,

Fig. 4.6. Strategy for character recognition.

or a real wavelet such as the gradient wavelets ∂x exp(−| x |2 ) or ∂x ∂ y exp(−| x |2 ), or even better, the end-stopped wavelet ψE1 described in the previous section. The latter has indeed been designed specifically for that purpose. This simple technique may be further improved by adding some denoising and inclusion of a second wavelet capable of dealing with letters of arbitrary shape (for instance, a ring-shaped wavelet sensitive to circular shapes). In addition, the automatic recognition device will need some training. An elegant solution would then be to use the simple wavelet treatment as a preprocessing for some sort of “intelligent” device, such as a neural network. However, when noise is present, a different strategy works better, namely, to use a directional wavelet instead of an isotropic one. The reason is that the detection capability of the former is more robust to noise. This feature may be understood as follows: to specify a direction is an additional element of information, that is present in the signal, but not in the noise, in general, so that the SNR ratio improves. In order to show this, we will analyze below another set of simple letters, namely, A, B and C, in the presence of increasingly strong noise (additive Gaussian noise). Now the technique used here is

134

Applications of the 2-D CWT. I: image processing

Fig. 4.7. Application of the algorithm to character recognition: (a) wavelet tranform of the signal in

Figure 4.4; (b) the same after thresholding: maxima (convex corners) are indicated with crosses, minima (concave corners) with circles.

a particular instance of the general problem of feature detection and recognition, so we postpone the details of the algorithm to the next section. We may also notice that directional wavelets, namely directional derivatives of a smoothing function, have been applied to the same problem of character recognition by Hwang and Chang [230], implementing the wavelet maxima technique of Mallat and Zhong [264,265].

4.2

Object detection and recognition in noisy images Suppose we have an image containing a certain number of targets, embedded in a cluttered environment: how can one detect and identify the various targets in an automated way? Stated explicitly [16], the purpose of automatic target detection and recognition (ATR) is the use of computer processing to detect and recognize signatures in sensor data, especially targets embedded in a cluttered environment, with the aim of neutralizing potential threats to military and civilian populations while minimizing the required resources and the risk to human life (this problem has an obvious military connotation, which explains the jargon used!). Such targets can be tanks, planes, other vehicles, missiles, ground troops, etc. Clutter can be grass, trees, topographical features, atmospheric phenomena (i.e., clouds, smoke, etc.). In general, the situation can be modeled using the following equation: s( x ) = n( x) +

L l=1

Tl ( x ).

(4.1)

135

4.2 Object detection and recognition in noisy images

where n( x ) represents an additive noise (clutter plus measurement noise), Tl ( x ) are targets to be detected and recognized, and s( x ) represents the accessible measured signal. Automatic or assisted target detection and identification requires the ability to extract the essential features of an object from (usually) cluttered environments. However, detection and identification lead to different requirements. Typically, to provide detection and identification of difficult targets while maintaining full surveillance coverage, a coarse resolution sensor is required for detection, while a fine resolution sensor is necessary for recognition (identification). This suggests the use of multiscale techniques, which provide the flexibility to utilize only the resolution required at each level and, perhaps, allow optimal processing for each of the required operations. Many methods are used for the ATR problem: classical pattern matching, modelbased schemes, dyadic wavelets, subband coding, even some attempts using neural networks [155,247]. Here we will describe an approach based on the 2-D continuous wavelet transform, which could offer a great improvement over traditional pattern matching methods. The rationale for using the CWT for ATR is the following. Typical features to be extracted from the image of a target are its position, its spatial extent, and its shape, including its orientation and symmetry. Thus the relevant parameters are position, scale and orientation, that is, exactly those considered in the 2-D continuous wavelet transform. The scale dependence allows sensitivity to variations in sensor resolution, as well as determination of target size, or equivalently the target distance, for instance in optical or infrared imagery. Rotation dependence leads to robust behavior in identifying the orientation of the target. So, unlike other methods, the 2-D wavelet transform incorporates several parameters directly relevant to the essential features of an object. Projection of the transform can thus provide a useful set of image representations for fully automated discrimination. In addition, wavelet methods yield a consistent and efficient image reconstruction algorithm. Indeed, the CWT has been used successfully in a number of situations, notably in infrared imagery. We will discuss this application in the next two sections.

4.2.1

Principle of the ATR wavelet algorithm A simple ATR algorithm based on the 2-D CWT has been proposed in [16]. It consists in a two-stage strategy and relies in an essential way on the successive use of the two basic representations described in Section 2.3.3, the position and the scale-angle representations. A similar technique has been used previously in the analysis of acoustic wave trains in water, see Section 5.3.4. The algorithm reads as follows (Figure 4.8). At the first stage, we compute the CWT in the position representation at all relevant a j , θ j ), j = 1, 2, . . . . For the detection, scales a = a j and angles θ = θ j , that is, S(b, we take the image obtained for each fixed pair (a j , θ j ), threshold it, and add all the images together. Thresholding is performed in an adaptive way, becoming more severe

136

Applications of the 2-D CWT. I: image processing

allow false alarm Input Data

✲

Detection of potential targets: position representation

❄

(a, θ) = (a j , θ j ) j = 1, . . . , N

bi , i = 1, . . . , L

Elimination of ambiguous targets bk , k = 1, . . . , K (K < L) ❄

Target recognition: scale-angle representation

b1 → target T1 (a1 , θ1 ) b2 → target T2 (a2 , θ2 ) ...

Fig. 4.8. A diagram illustrating the two-step strategy for ATR.

for smaller a. Note that other nonlinear transformations (e.g., enhancement, morphological operators) may also be applied. The effect of this procedure is to suppress the clutter information, while preserving the target information. As a result, the latter is reinforced and becomes visually enhanced. Next, we compute the centroids b = bi , i = 1, . . . , L in the resulting composite image. These centroids correspond to the positions of potential targets. False alarms are of course possible, but one may control the false-alarm rate by adjusting the thresholds in order to eliminate spurious false detection. Then, at the second stage, one switches to the scale-angle representation and computes the wavelet transform of the composite image at each remaining centroid b = bk , k = 1, . . . , K (K L). If the centroid bk corresponds to a genuine target Tk , the corresponding wavelet transform will exhibit a unique maximum (ak , θk ), which gives the size and the orientation of the target Tk . Moreover, the signature of each target in the scale-angle representation allows the discrimination between different targets. An academic example of application of the algorithm just described, more precisely the first stage of it, is presented in Figure 4.9. We take our favorite L-shape (a), embedded in a Gaussian white noise (b), with a signal-to-noise ratio of 18. The wavelet transform, with a Mexican hat, is taken at six different scales, a = 1, 2, 4, 8, 16, 32, each image is properly thresholded (at 90%/a), and the six images are added together. The resulting composite image (c) shows the reconstructed object, the noise has largely been suppressed.

137

4.2 Object detection and recognition in noisy images

(b)

(a)

(c) Fig. 4.9. Reconstruction of a signal embedded in noise. (a) The signal; (b) the noisy signal; (c)

reconstruction of the signal with six scales.

Of course, this method is rather primitive, although it is fast and robust. An alternative technique for image denoising relies on the use of directional wavelet packets (Section 2.6). We will make a detailed comparison between the two methods in Section 4.6.

4.2.2

Application to infrared radar imagery: position features The real power of this approach is, of course, better appreciated in real life situations, preferably difficult. A prime example is that of infrared imagery (FLIR or Forward Looking Infrared Radar imagery). Automatic target detection and identification for

138

Applications of the 2-D CWT. I: image processing

FLIR imagery requires the ability to extract the essential features of an object from cluttered environments under the condition that the range to the target is unknown. Moreover, the gray-scale of a target in FLIR imagery displays a great variability: a succession of dark and bright areas depending on the temperature radiating from various parts of the target (for example, a target can be hot or cold). Most of the time, the target does not appear in ideal conditions and it is very difficult to estimate the sensor output probability density function (PDF). One therefore uses algorithms that first extract features of interest (such as structural, spatial and frequency features) and then classifies the objects based on those features (a review of these issues may be found in [134]). Multiscale techniques, such as the CWT, are highly desirable, because they can extract and normalize both the unknown scale and orientation of the target. As indicated above, typical features to be extracted from the image of a target in a cluttered environment are the position, the spatial extent, and the shape of the target, including its orientation and symmetry. This justifies the use of the 2-D CWT in the ATR problem, and in particular for infrared imagery. An additional issue is to determine which wavelet will perform best, an isotropic one (Mexican hat) or a directional one (Morlet or Cauchy). Now, in the presence of heavy noise, it is better to apply the ATR algorithm described above not to the CWT amplitude itself, but rather to one of the partial energy densities discussed in Section 2.3.4. We will then speak of CWT features to be extracted from the image. This opens a choice between two solutions. The most obvious one is to take the position energy density of the signal s( x ), given in (2.56), namely, ∞ da 2π = a, θ )|2 , P[s](b) dθ |S(b, (4.2) a3 0 0 and this we shall do in this section. We test the algorithm on FLIR data from the TRIM2 database, namely, a set of images each of which contains four targets (various types of tanks), seen under 21 different aspect angles (we recall that the aspect angle α is the polar α)). Figures 4.10 and 4.11 present two angle in the position representation, b = (|b|, such images, together with the corresponding receiver operator characteristics (ROC) curves. These curves plot the probability of detection versus false alarm rate, for the whole set of images containing a given type of tank at 21 aspect angles, analyzed in turn with the Mexican hat, the Morlet, and the Cauchy wavelet. The first stage of the ATR algorithm is then applied to the image in Figure 4.11(a) and the results are presented in Figure 4.12. The upper row shows the output of the first step of the algorithm, i.e., the CWT position features, evaluated with the three wavelets successively. Clearly, the targets stand out from their background with more definition than the original. The bottom row then gives the output of the second step of the algorithm, that is, the use of thresholding, morphological transformations, and other nonlinear transformations. The particular image presented corresponds to thresholding at gray-scale value 40. Finally, Figure 4.13 gives the Receiver operator characteristics (ROC) curves obtained

4.2 Object detection and recognition in noisy images

m2

prob. of detection

1

0.75

Cauchy Morlet Mexican Hat

0.5

0.25

0 0

10

20

30

40

(a)

50 60 70 false alarms

80

90

100

110

120

(b)

Fig. 4.10. (a) Original image containing four M2 tanks (left); (b) Receiver operator characteristics (ROC) curves for the set of images containing M2 tanks at 21 aspect angles.

m163

1

prob. of detection

139

0.75

Cauchy

0.5

Morlet Mexican Hat 0.25

0

(a)

0

10

20

30

40

50 60 70 false alarms

80

90

100

110

120

(b)

Fig. 4.11. (a) Original image containing four M163 tanks; (b) Receiver operator characteristics (ROC) curves for the set of images containing M163 tanks at 21 aspect angles.

by combining images with all types of tanks, again for the three wavelets, Mexican hat, Morlet, and Cauchy wavelets. The results show that directional wavelets such as the Morlet and the Cauchy wavelets perform better than an isotropic one such as the Mexican hat. Moreover, the Cauchy wavelet performs better than the Morlet wavelet. This can be understood as follows. The selected images contain objects plunged in clutter noise. Detection of an object of interest supposes, for example, the ability to capture the internal structure (for example, wheels, doors, etc . . . ) of the object, the boundaries (edges, corners) between the

Applications of the 2-D CWT. I: image processing

Fig. 4.12. CWT position densities (top row) and detection results (bottom row) of the original image in Figure 4.11 with the Mexican hat (first column), the Morlet (second column), and the Cauchy wavelet (third column), respectively.

1

prob. of detection

140

0.75

Mexican Hat

0.5

Morlet Cauchy

0.25

0 0

10

20

30

40

50 60 70 false alarms

80

90

100

110

120

Fig. 4.13. Receiver operator curves for the detection algorithms on data containing M1, M2, M163, and M60 tanks.

141

4.2 Object detection and recognition in noisy images

object and its surrounding. This justifies the use of directional wavelets in this context, and explains their better performance over nondirectional filters. This technique might be applied with different wavelets, for example, target-adapted wavelets. It might also be compared to or combined with the CWT-based classification algorithm developed in [287].

4.2.3

Scale-angle features and object recognition The detection algorithm described in the previous section was based on CWT position features extracted from the image. An alternative solution is to use the scale-angle energy density (scale-angle spectrum) of the signal s( x ), given in (2.57), namely, a, θ)|2 . M[s](a, θ ) = d 2 b |S(b, (4.3) R2

In this section, we shall present an algorithm based on scale-angle CWT features, then apply it again to automatic character recognition and target recognition in FLIR images. As we will see, this algorithm allows significant reduction of the data needed for an efficient recognition and it is robust against noise. As in the previous case, we will compare the performance of two wavelets, this time two directional wavelets derived from the Mexican hat, the modulated or Gabor Mexican hat (3.22) and the conical Mexican hat (3.38). Detailed comparative tests show that the conical Mexican hat wavelet outperforms both the Gabor Mexican hat and the usual Morlet wavelet, and all of them outperform traditional methods of character recognition such as template matching [289]. As a first application, we analyze another set of simple letters, namely, A, B, and C, in the presence of increasingly strong noise (additive Gaussian noise). The characters are shown in Figure 4.14(a), at various SNR levels, 20, 25, 30, 35, and 40 dB. Then we compute the continuous wavelet transform of these letters, using the conical Mexican hat wavelet (3.38). The result is presented in Figure 4.14(b) in the scale-angle representation, evaluated at the center of each character. In this way, the directional features of the object are enhanced, and indeed are detected despite the noise. Given any letter in Figure 4.14(a), the problem is to recognize it (in an automated way), that is, to determine which of the 26 letters of the alphabet it resembles most – or actually coincides with. This is an instance of object recognition in a noisy environment, more precisely, identification of a noisy object within a preassigned collection of test objects – here the 26 letters of the alphabet. Here again, the standard technique consists of choosing a certain number of characteristic features that suffice to discriminate unambiguously among the test objects. Then one arranges them in a feature vector and measures the distance between the feature vector of the unknown object and those of each test object. The one that yields the smallest distance gives the answer.

Fig. 4.14. (a) Test data for the recognition algorithm: three characters A, B, and C at various SNR levels, 20, 25, 30, 35, and 40 dB. (b) Continuous wavelet transform of the same, using the conical Mexican hat wavelet. The CWT is presented in the scale-angle representation, evaluated at the center of the character.

143

4.2 Object detection and recognition in noisy images

4.2.3.1

The recognition algorithm We will first describe the algorithm in complete generality. Applications to specific problems will be given in the following sections. Consider a training set, consisting of objects divided into q classes, X j , j = j j j 1, 2, . . . , q, the class X j having p j prototypes, X j = (X 1 , . . . , X l , . . . , X p j ), and an unknown object, Y , that we wish to classify as object X j , for some j. The algorithm for classification is described as follows: r For every j = 1, 2, . . . , q, compute the 2-D CWT scale-angle energy density j j M[X l ](a, θ) for each element X l of the class X j . r For every j = 1, 2, . . . , q, compute the mean and the standard deviation of the elements of the class X j , j 1 j M[X l ](a, θ) p j l=1 " # pj & '2 1/2 1 j σ X j (a, θ) = . M[X l ](a, θ) − µ X j (a, θ ) p j l=1

p

µ X j (a, θ) =

(4.4)

(4.5)

r

Compute the scale-angle energy density M[Y ](a, θ ) of the test object. Select the scale-angle feature vectors V[Y ], V[X j ] for the unknown object Y and for each class X j , i.e., M[Y ](m, n) and µ X j (m, n), σ X j (m, n), m = 1, 2, . . . , N1 , n = 1, 2, . . . , N2 , where N1 , N2 are the chosen number of scales and angles, respectively. r Compute the distance between them, r

d j = d(Y, X j ) ≡ d(V[Y ], V[X j ]), j = 1, . . . , q.

(4.6)

r

Then, if dk is the minimum of the set {d j , j = 1, . . . , q}, the object Y is classified as belonging to the class X k . In principle, one can use any distance, such as the Euclidean distance d(Y, X j ) =

N2 & N1 '2 1 M[Y ](m, n) − µ X j (m, n) , N1 N2 m=1 n=1

or the maximum likelihood distance, which is given by " 2 # N1 N2 j (m, n) M[Y ](m, n) − µ 1 X . 2 log σ X j (m, n) + d(Y, X j ) = N1 N2 m=1 n=1 σ X j (m, n) Clearly this algorithm may lead to false recognition, if two or more distances d j are very close to each other.

4.2.3.2

Application to character recognition The algorithm has been tested [289] on the 26 characters of the alphabet (A,B,C, . . . .) at various noise levels (additive Gaussian noise). In this case, one has q = 26, that j is, each X j is one letter of the alphabet, whereas the p j prototypes X l correspond to different copies of the same letter, distorted, rotated, embedded in different noises. For

144

Applications of the 2-D CWT. I: image processing

No. of orientations = 72

No. of scales = 10

100

100 scales = 10

angles = 72

90

90

80 80

Correct classification

Correct classification

70

60

50

scales = 5

40

70

60

angles = 18

50

30 40 20

10

0 20

30

scales = 2

25

30 35 Noise level (in dB)

(a)

40

20 20

angles = 9

25

30 35 Noise level (in dB)

40

(b)

Fig. 4.15. Results of recognition as a function of noise level for various lengths of the feature vector in the scale-angle plane, using the modulated (Gabor) Mexican hat wavelet with k0 = 6 and eccentricity $ = 2; (a) fixed number (72) of orientations, different numbers of scales; (b) fixed number (10) of scales, different numbers of orientations.

instance, the five letters A on the top row of Figure 4.14(a) are five prototypes of an A, with different levels of noise. The results of the analysis are summarized in Figures 4.15 and 4.16. Both give the performance of the recognition algorithm for various lengths of the feature vector as a function of the noise level (SNR), using the modulated (Gabor) Mexican hat and the conical Mexican hat, respectively. In Figure 4.16 the result of the traditional template matching is added for comparison. In this experiment, it turns out that the conical Mexican hat performs better than the modulated Mexican hat, which in turn performs better than the traditional Morlet wavelet. The same behavior is observed in other cases, for instance in the recognition and classification of targets in FLIR imagery [285–289].

4.2.3.3

Application to radar imagery Other angular features of images have been used successfully in recent work by Kaplan and Murenzi on Synthetic Aperture Radar (SAR) images [241]. The problem here is to have a good estimate of the position of a target (in jargon, pose estimation as part of

145

4.3 Image retrieval

110

100 temp cwt15

Correct classification

90

80

70 cwt10

60

cwt8

50

40

22

24

26

28 Noise level (dB)

30

32

34

Fig. 4.16. Results of recognition as a function of noise level for various lengths of the feature vector in the scale-angle plane; the wavelet now is the conical Mexican hat.

ATR). Since the latter is assumed to be at the center of the image, one fixes the position at b = 0 in the CWT and integrates the energy density over scale only, thus defining the angular energy density amax da E[s](θ ) = |S(0, a, θ )|2 . (4.7) a3 amin The pose estimate is then simply the orientation that maximizes the angular energy density, θ = arg maxθ E[s](θ). This technique yields very good results, using the fast circular convolution algorithm described in Section 2.3.1 with an isotropic Mexican hat wavelet (but, surprisingly, not with a Morlet wavelet).

4.3

Image retrieval

4.3.1

The problem of content-based image retrieval Digital images are everywhere, and they often come in huge quantities: remote-sensing agencies, medical imaging facilities, art museums, travel agencies, law-enforcement

146

Applications of the 2-D CWT. I: image processing

agencies, all have to deal with large collections of archived digital images. Depending on the particular situation, these collections may be homogeneous (as in medical imaging or fingerprint archives), or completely heterogeneous. Thus there arises the problem of image retrieval: given an image, one has to identify those images in the collection that resemble it most (a typical example is the identification of a suspect’s fingerprint among those archived in the police records). Beyond a few dozen images, manual browsing is practically impossible, a purely computerized solution is necessary. Traditional methods of object classification, based on external descriptors, are inoperant for images, especially in heterogeneous collections. One has to resort to a description derived automatically from the image itself, and this leads to a fairly new field of research, called content-based image retrieval. Although several methods have been proposed, we will consider here only the wavelet-based scheme designed by Bhattacharjee in his thesis [Bha99], where an overview of other methods may be found. The roots of this approach lie in the theory of vision initiated by Marr [Mar82], namely the process of comparing images must necessarily be based on low-level information extracted from the images, especially in the case of a heterogeneous collection of images, where no assumption is made about the content of an image. Indeed, methods based on segmentation do not make sense in such a situation. Subdividing the image into blocks, and performing block-wise comparisons of images does not work either, because this approach is not invariant to rotation and cannot support arbitrary subimage-queries. Thus image comparison should be based on visually significant structural features detected in the images. Indeed, this seems to be the case when human subjects view a scene. Psychophysical experiments show that even when analyzing a static scene, our eyes do not remain continually focused on a single retinal image, but rather perform the so-called saccadic movements described in Section 3.3.3. As explained there, the target points of consecutive saccades are points of interest which stand out against the general background of the scene. Thus it remains to identify those key points, and we are facing again the problem of identifying relevant features in an image. According to experiments, the crucial ones are low-level features, which can be classified in various ways (see [Bha99] for a complete discussion). For instance, one may distinguish patches of uniform intensity; edges or lines; and corners or line ends. Among these, the last kind seems to encode the maximum amount of image information. For this reason, it is natural to design wavelets that respond precisely to these features, and the end-stopped wavelets described in Section 3.3.3 are Bhattacharjee’s answer to that question. Before going further into their actual implementation, we shall now briefly describe the whole recognition scheme. The first stage is the feature point detection scheme, which forms the core of the image comparison process. However, image comparison based simply on the positions or color attributes of the feature points would not be very robust. Thus, for each feature point, one constructs a description based on the texture of the immediate neighborhood

147

4.3 Image retrieval

of the point. This is obtained by using suitable filters, namely directional derivatives of a Gaussian, of order one, two and three. The responses of these filters are organized in an ordered set and may be represented as a vector. Thus each feature point detected in the image produces one vector of responses, called a token. These tokens carry information about a small region in the image, including directional information. The last step of the process is the actual measure of similarity between images, based on tokens. This requires a rather complex image indexing strategy. The outcome is an efficient technique for content-based image retrieval, which proceeds iteratively. To quote the author [Bha99]: . . . To use the system, the user presents the query in the form of an image. The system then sorts all images in the collection in order of decreasing similarity to the query, and returns the top few images as the answer-set. The size of the answer-set is specified by the user. From the answer-set, the user may mark the images that are relevant to the user’s needs, and provide this information to the image retrieval system as feedback. The system then refines the query automatically, based on the relevance information provided, and subsequently returns another answer-set of images to the user. This process may be iterated till the user is satisfied.

This image recognition scheme turns out to be both efficient and reasonably economical, as attested by several explicit examples given in [Bha99].

4.3.2

Feature point detection using an end-stopped wavelet To conclude this section, we discuss now in more detail the wavelet-based algorithm for feature point detection [Bha99,76,77] (we quote freely from these works). Both endstopped wavelets ψE1 and ψE2 can be used to select meaningful points in images, and this is why we have discussed them both in Section 3.3.3. However, the author chooses to consider the former only, because, as explained above, he is essentially interested in detecting corner-like features, to which ψE1 responds best. In addition, he works at a fixed scale, since an exhaustive search of the scale space is computationally prohibitive. Furthermore, for a heterogeneous collection of images, to which new images may be added as and when they become available, it is impossible to identify a specific set of scales that will be appropriate for analyzing all images in the collection. Thus, the algorithm applies the ψE1 at the same scale to all images in the collection. There are two negative consequences of this compromise. r The feature point detection scheme is not scale invariant. Consider two images I and 1 I2 , where I2 is a subsampled version of I1 by a factor of two along each dimension. If the ψE1 wavelet is applied at the same scale to both images, we are not guaranteed that the set of feature points detected in I2 will have a one-to-one correspondence with the set of feature points detected in I1 . That is, the system is not invariant to scale changes even by a factor of four. However, as the experimental results demonstrate, the final system is quite robust to small variations of scale, presumably because of the redundancy of information between responses at nearby scales.

148

Applications of the 2-D CWT. I: image processing r

Some of the feature points detected in an image may not be well localized, that is, the local maxima of the wavelet response may not coincide with the corresponding image feature. This happens because the scale of analysis is coarser than the most appropriate local scale, so the local maximum of the wavelet response does not fall on the feature in question, due to the interaction with other neighboring features. Such seemingly spurious points would pose serious problems for recognition systems. However, the performance of the proposed image retrieval system is not appreciably impaired by the use of a fixed scale of analysis. This is because the comparison of images is not based on a recognition process, but rather, is based on a comparison of small image-patches surrounding the feature points. This being said, the algorithm reduces essentially to the standard one described in Section 4.1.2. (i) Transform the input image, at a preselected scale, s0 , with the ψE1 wavelet at N different orientations θ = θ0 , θ1 , · · · , θ N −1 . The result is a set of N response images, each showing strong responses near the end-points of linear structures oriented perpendicular to θ. (ii) For each pixel position, retain only the strongest response value among all the orientations. This produces the so-called maxima image. (iii) Detect peaks of significant local maxima in the maxima image. The coordinates of these peaks give the feature points. In practice, the author uses N = 18, which corresponds to an orientation resolution of 10◦ , to cover the entire semicircle evenly. In fact, the angular resolving power of the ψE1 wavelet is 19.2◦ , as measured by the standard benchmarking technique described in Section 3.4.1, but the response is quite weak away from the axis, so that a significant overlap is recommended. Finally, the scale of analysis chosen is a = 8. We conclude by showing some experimental results obtained with this feature point detection scheme. In particular they demonstrate the robustness of the scheme in the face of rotation and cropping. Figure 4.17 shows the feature points detected for two images of very different kinds. In both cases, the detected points are marked by bright crosses superimposed on the input image. Figure 4.17(c) shows the result for a face image, and Figure 4.17(d) shows the points detected in an ornament image. The ornament images depict very complex hand-drawn artwork, which were designed to be trademarks of publishers in the nineteenth century (these images have been scanned from photocopies of old books, as gray level images). Figure 4.18 demonstrates the robustness of the proposed feature detection scheme towards rotation and cropping. The image in Figure 4.18(a) shows the feature points detected for a subimage of the ornament image shown in Figure 4.17(b). Note that most of the feature points marked in Figure 4.18(a) have corresponding feature points in Figure 4.17(d). Figure 4.18(b) shows the feature points detected for a subimage extracted from a rotated version of the image shown in Figure 4.17(b). The image has

149

4.3 Image retrieval

Fig. 4.17. Results of feature point detection: (a) face test image; (b) ornament test image; (c) feature

points detected in (a); (d) feature points detected in (b). The ornament image shown in (b) is a black-on-white pattern that has been scanned as a gray level image. In both (c) and (d), the white crosses mark the positions of the detected feature points (from [Bha99]).

Fig. 4.18. Robustness to cropping and rotation: (a) feature points detected in a subimage of Figure 4.17(b); (b) feature points marking detected in a rotated, cropped version of Figure 4.17(b). Most feature points are detected in both images (from [Bha99]).

150

Applications of the 2-D CWT. I: image processing

been rotated by 17◦ . Again, most of the feature points in the corresponding section of the original image [Figure 4.17(d)] are also detected in the rotated, cropped version.

4.4

Medical imaging Wavelet applications abound in medicine and biology, both in 1-D and in 2-D, often with the discrete WT [Ald96]. For the 2-D case, one may quote image segmentation, mammography, tomography [68,315], and magnetic resonance imaging (MRI) [217]. Here we will mention only one potential application, still under development, namely the technique of spiral reconstruction in MRI. The problem at hand is the so-called spiral acquisition method in MRI, that is, the Fourier transform of the image to be analyzed is sampled along several interleaving spirals [84]. This raises the question of the completeness of the reconstruction of the original image from such data. This is clearly a case of nonuniform sampling, and completion means that the set of sampling points generate a frame. The question was analyzed and solved by Benedetto [67], using powerful mathematical tools, such as Beurling’s theorem. Actually, 2-D wavelets may also be used in this context, the hope being that they may allow to bypass the infamous “gridding” problem, namely the necessity of adjusting the sampling points on a Cartesian grid in order to apply the FFT algorithm.

4.5

Detection of symmetries in patterns

4.5.1

The tools for symmetry detection Wavelets may be used for evaluating the symmetry of a given pattern under discrete rotations and dilations, as was demonstrated in detail in [24], on which this section is based. The method presented here allows one to determine, in a straightforward and economical way, all the (possibly hidden) symmetries of a given pattern. Of course, invariance under separate rotations or dilation is easy and there are various methods for determining it. But the determination of combined dilation–rotation invariances (helicoidal symmetries) of a given pattern is much more delicate and, in fact, we do not know any other method for doing it. The technique uses, in an essential way, the angular selectivity of the directional wavelets. In order to achieve good precision, one needs a directional wavelet with a very good selectivity in the scale-angle variables. The Gaussian conical wavelet (3.37) is an extremely efficient tool in that respect and we will use it systematically. However, it is a fact that most patterns of interest possess only local or approximate symmetries of this type, i.e., without true periodicity or discrete translational invariance. They are thus quasiperiodic sets, such as quasilattices, planar self-similar (Penrose)

151

4.5 Detection of symmetries in patterns

tilings or diffraction patterns of quasicrystals. These objects have indeed a local symmetry only, but the method applies as well, because the local character of the wavelet transform allows us to treat exact and local symmetries precisely on the same footing. For quasiperiodic objects, the useful information is contained, to a first approximation, only in the scale and angle variables, since there is no spatial periodicity as for regular (crystallographic) tilings [Bar94,61,322]. In such a case, one may ignore the dependence of the CWT on the translation degrees of freedom, represented by the One possibility is to use the scale-angle representation, which consists parameter b. of fixing the position parameter b (Section 2.3.3). However, this may lead to ambiguities, because the result, including its symmetries, depends sensitively on the value of b that has been chosen [24] (an example is given in Section 4.5.2.3). The alternative is to average over all values of b and consider the scale-angle spectrum, as defined in (2.57). a, θ)|2 M[s](a, θ) = d 2 b |S(b, R2 2 rθ−1 k)| 2 d 2 k , 2 |ψ(a (4.8) = (2πa) | s(k)| R2

a, θ) is the wavelet transform of the signal s( where S(b, x ) with respect to a directional wavelet. Clearly, M[s] gives the intensity of the spectrum of s, namely, the contents of |s|2 according to the shape of |ψ|2 . Furthermore, if ψ is averaged locally around arθ−1 (k), is supported in a narrow cone, and then (4.8) “probes” the behavior is directional, ψ of the signal in the direction θ, as the beam of a torchlight exploring a target. This intuitively explains all the results that follow. Positions are not considered in the analysis, because only the modulus of s is used. This is why the method may be interesting in a (quasi)crystallographic context, where only amplitudes of the diffraction spectrum are recorded in experiments. In practice, however, the scale-angle spectrum is often not very readable, because some of the (approximate) symmetries may be rather weak. Therefore, exactly as for the wavelet transform itself, one plots instead the skeleton of the scale-angle spectrum, which in this case reduces to the set of local maxima (isolated peaks in the scale-angle spectrum), and then the possible periodicity properties become clearly visible. In addition, the scale-angle measure is well adapted for analyzing a statistical symmetry [321]. This is a weaker concept of symmetry, which corresponds to the invariance under rotation, or dilation, of the frequency of appearance of any given local configuration inside of the pattern. This is clearly the relevant concept when one is dealing with an approximate symmetry. Let us thus assume that a certain substructure of s interferes positively with the wavelet ψbo ,ao ,θo ; that is, the corresponding wavelet coefficient is large. Assume further that this configuration occurs a certain number of times in s, giving to M[s] a local maximum at a point (a0 , θ0 ). Then, if s is statistically symmetric under a certain rotation

152

Applications of the 2-D CWT. I: image processing

Rρ and a certain dilation Dλ , the local maximum of M[s] has a mirror image at (a0 λ, θ0 + ρ) by the covariance of the CWT (Proposition 2.2.3). The important fact is that not only the maxima are relevant, but all the points of the scale-angle spectrum, because, in the definition of statistical symmetry, any local configuration has importance and gives its contribution to M[s]. Of course, it is nice to detect (statistical) symmetries of the image, but one also wants to know whether one has found all of them. An answer to that question is given by the voting algorithm introduced in [Vdg98], extending a technique of Hwang and Mallat [228]. The idea is to make a vote on the more significant symmetries of the image s under consideration. First, one computes the correlation P[s](τ0 , α0 ) between the scale-angle spectrum M[s] and the version obtained under dilation by a factor λ0 and rotation by an angle α0 , that is, M[s](aλ−1 0 , θ − α0 ). Introducing logarithmic coordinates a = t τ0

e , λ0 = e , and defining M [s](t, θ) = M[s](et , θ), one has M[s](aλ−1 0 , θ − α0 ) = M [s](t − τ0 , θ − α0 ). The correlation P[s] is thus given by: tmax 2π −2 dt dθ M [s](t − τ0 , θ − α0 ) M [s](t, θ ), (4.9) P[s](τ0 , α0 ) = s2 tmin

0

where τ0 ranges from 0 to the width of the logarithmic scale interval. In practice, of course, we take a bounded interval [tmin , tmax ] for t and we work with discrete steps and thus the integration is approximated by a summation over a linear grid ⊂ [tmin , tmax ] × [0, 2π ]. Then the algorithm allocates a vote to the point (τ0 , α0 ) if P[s](τ0 , α0 ) exceeds a given constant K > 0 (which specifies the error that is tolerated). Once a vote has been cast for a point (τ0 , α0 ), one identifies all its integer multiples (nτ0 , nα0 ) that lie within , and give all their votes to (τ0 , α0 ) [Vdg98,24].

4.5.2

Detecting symmetries in 2-D patterns We shall now apply the method developed in Section 4.5.1 to the detection of rotation– dilation symmetries of certain classes of 2-D patterns, following essentially [24,26].

4.5.2.1

Geometric patterns We begin with a simplified version and eliminate the scale dependence by integrating over a, thus ending with the angular spectrum α[s](θ) of the object, defined in (2.60). In general, α[s](θ) is 2π-periodic. However, when the analyzed object has rotational symmetry n, i.e., it is invariant under a rotation of angle 2π /n, then α[s] is in fact 2π/n-periodic. To give a simple example, consider three geometrical figures, a square, a regular hexagon and a rectangle [24]. The square and the hexagon have symmetry n = 4 and n = 6, respectively, and thus their angular spectrum show four and six equal peaks, respectively (Figure 4.19). The width of these peaks is simply the aperture of the support

153

0

4.5 Detection of symmetries in patterns

90

180

(a)

270

360

0

90

180

(b)

270

360

0

90

180

270

360

(c)

Fig. 4.19. Angular measure of regular figures obtained with a Cauchy wavelet (A R P = 20◦ ): (a) a square; (b) a regular hexagon; and (c) a rectangle, with side ratio 2 : 1 (from [24]).

cone (i.e., the ARP) of the wavelet. The case of the rectangle is more interesting. It has symmetry n = 2 × 2 (two mirror symmetries, or rotations by π around both O x or O y), and that is reflected on the graph of its angular spectrum: there are two large peaks corresponding to the longer edges and two smaller peaks corresponding to the shorter ones, and the ratio 2 : 1 between the two equals that of the lengths of the corresponding edges. Indeed, the wavelet catches the direction of the edges, not that of the corners, so that indeed the maxima of α[s] are again at θ = 0◦ , 90◦ , 180◦ , 270◦ , just as for the square, but now the amplitudes are different. This also explains why the peak at 90◦ in the case of the hexagon, panel (b), is slightly higher: the vertical sides in the original figure are sharp, the oblique ones are ragged, for numerical reasons, so that the former give a larger response to the wavelet. This explains why one needs a highly directional wavelet in this case. It is remarkable that the scale-angle spectrum technique works in the presence of severe noise. Let us take again a square pattern and compute its CWT with a directional wavelet, first without noise (Figure 4.20), then with moderate additive Gaussian noise (Figure 4.21, top panels), finally with severe additive Gaussian noise (Figure 4.21, bottom panels). In each figure, we show successively: (c) the angular spectrum (2.60), which reveals the fourfold symmetry of the pattern; and (d) the scale spectrum (2.61), which measures the size of the object (from [289]). Next we proceed to patterns with a genuine combined rotation–dilation symmetry. In this case we need the full scale-angle spectrum M[s](a, θ), which will again be 2π/n-periodic in θ if the pattern has rotational symmetry n. In addition, if the object is invariant under dilation by a factor ao , then M[s] is (log ao )-periodic in log a. Thus in the case of an inflation invariance, M[s](a, θ) is a doubly periodic function in log a and θ (note that, like wavelet transforms themselves, a scale-angle spectrum is usually plotted as a function of log a and θ). The first object we analyze is a “twisted snowflake,” that is, a mathematical snowflake [43,44] with the following modified construction rule: upon each downscaling by a

154

Applications of the 2-D CWT. I: image processing

(b)

(a) 0 10

scale

20 30 40 50 60 0

100

1

0.8

0.8

0.6

0.6

0.4 0.2 0

300

(d)

1

energy

energy

(c)

200 angle

0.4 0.2

0

100

200 angle

300

400

0

0

20

40 scale

60

80

Fig. 4.20. Detection of the symmetries of a square pattern: (a) the signal; (b) the scale-angle spectrum; (c) the angular spectrum; and (d) the scale spectrum (from [289]).

factor of 3, the figure is turned by 36◦ . The scale-angle spectrum of this object, given in Figure 4.22(b), shows precisely the combined symmetry. The set of four maxima at a given scale ao is reproduced, at scale ao /3, but translated in θ by 36◦ .

4.5.2.2

Quasiperiodic point sets An interesting class of point sets is that of the quasilattices based on algebraic numbers (see [Bar94,61] for a systematic analysis). All of them possess a rotational symmetry of order n, where n may be crystallographically allowed (n = 1, 2, 3, 4, 6) or not (e.g. n = 5, 8, 10, 12). Moreover, each pattern is invariant under dilation by a characteristic factor, which is an integer equal to 2 cos(2π /n) in the first case, and an irrational number βn in the second case (thus called quasicrystallographic). But in fact there is more. In many cases, there is in addition a combined rotation–dilation symmetry, typically a rotation by π/n together with a dilation by a factor δn related to βn . As an example, we consider the octagonal pattern shown in Figure 4.23 [24]. It has a global √ symmetry n = 8 and is invariant, by construction, under dilation by a factor 1 + 2. But the scale-angle spectrum M[s](a, θ ) (calculated with a Gaussian conical wavelet) reveals two combined rotation–dilation symmetries, namely a rotation of π/8 √ together with a dilation by a factor δ1 = 2 cos(π/8), or δ2 = 2 cos(π/8), respectively. The remarkable fact is that these two additional symmetries were discovered on the graph of the scale-angle spectrum, not on the tiling itself! (Actually both symmetries

4.5 Detection of symmetries in patterns

(a)

(b) 0 10

scale

20 30 40 50 60 0

100

(c) 0.8 energy

energy

300

1

0.95

0.9

0.85

0.8

200 angle

(d)

1

0.6 0.4 0.2

0

100

200 angle

300

0

400

0

20

(a)

40 scale

60

80

(b) 0 10

scale

20 30 40 50 60 0

100

(c)

200 angle

300

(d)

1

1

0.98

0.8 energy

0.96 energy

155

0.94

0.6 0.4

0.92 0.2

0.9 0.88

0

100

200 angle

300

400

0

0

20

40 scale

60

80

Fig. 4.21. The same analysis as in Figure 4.20, in the presence of additive Gaussian noise: (top) moderate noise [signal-to-noise ratio (SNR) = 26 dB]; (bottom) severe noise (SNR = 22 dB).

Applications of the 2-D CWT. I: image processing

−1.5

log(3)

−2

−2.5

log(a)

−3

−3.5

−4 0

20

40

60

80

100

120

140

160

180

θ

36

(a)

(b)

Fig. 4.22. Analysis of a “twisted snowflake”: (a) the pattern; (b) the scale-angle spectrum M[s](a, θ ), computed with a Cauchy wavelet (m = 4, γ = 10◦ ). Corresponding local maxima are shifted by 36◦ and a scaling ratio of 3 (from [24]).

2 1.8 1.6 1.4 1.2 ln(a)

156

1 0.8 0.6 0.4 0.2 0 0

(a)

20

40

60

80

θ

100

120

140

160

180

(b)

Fig. 4.23. Analysis of an octagonal pattern: (a) the pattern; (b) the local maxima of its scale-angle spectrum M[s](a, θ ); this pattern has a rotation symmetry by π/4, and two distinct √ mixed symmetries, consisting of a rotation by π/8 combined with a dilation by δ1 = 2 cos(π/8), δ2 = 2 cos(π/8), respectively. Homologous maxima are linked by a line segment, continuous for δ1 and dashed for δ2 (from [24]).

157

4.5 Detection of symmetries in patterns

(a)

(b)

Fig. 4.24. Two sets of octagons on the octagonal tiling obtained by successive applications of a √ rotation by π/8 combined with a dilation by (a), δ1 = 2 cos(π/8); (b), δ2 = 2 cos(π/8) (from [24]).

were later derived by a geometrical argument.) A nice way of visualizing the symmetries is to draw successive octagons, representing the orbits of successive points under a rotation by π/4 (Figure 4.24). Note that, for a better visualization of these orbits, we have brought back all successive summits into the first sector 0 θ π/8. This operation reveals an additional difference between the two symmetries. Indeed the pattern on the right is invariant under the combined operation δ2 -dilation + rotation by π/8, and this operation generates a semigroup (every point has a successor, not necessarily a predecessor, i.e., the inverse operation is not a symmetry). This semigroup has apparently infinitely many different orbits (on the portion of the tiling visible on the figure, we have detected 10 different orbits). However, the other combined operation, δ1 -dilation + rotation, is not an exact symmetry, it is only approximate. For instance, some orbits stop after a few iterations, or have gaps. Now comes the question, did we detect all symmetries of the octagonal pattern? The answer is in fact yes, as shown by the result of the voting algorithm described in Section 4.5.1, presented in Figure 4.25. The graph indeed shows the two pure operations, rotation √ by π/4 and dilation (ρ0 ) by a factor 1 + 2, and the points ρ1 and ρ2 corresponding to the combined rotation–dilation operations with dilation ratio δ1 and δ2 , respectively. Note that the pure dilation ρ0 is equivalent under the π/4-periodicity to the product ρ1 · ρ2 .

4.5.2.3

Other examples of aperiodic patterns There are many more examples of patterns that exhibit this kind of combined symmetries. A whole class is that of tilings of the plane, some of which are commonly known

Applications of the 2-D CWT. I: image processing

1.2

1

ρ

0

0.8

ln(a)

158

ρ2

0.6

0.4

ρ1

0.2

0

0

10

20

30

40

θ

50

60

70

80

Fig. 4.25. The set of renormalization parameters ρ = (δ, θ) obtained by the voting algorithm on the

scale-angle spectrum of the octagonal tiling given in Figure 4.23. The point on the horizontal axis corresponds to the π/4-periodicity. The points ρ1 and ρ2 correspond to the combined rotation– dilation operations with dilation ratio δ1 and δ2 , respectively, whereas ρ0 is a pure dilation, equivalent to the product ρ1 · ρ2 under the π/4-periodicity. The other, unmarked, points are translations of the previous ones under both periodicities, in a and θ (from [24]).

under the name of Penrose tilings (these are dual to the preceding type, in the sense that they are obtained by drawing the Voronoi cells of the point set). We show a typical example in Figure 4.26. From the scale-angle spectrum, obtained with a Gaussian conical wavelet (3.37), with parameters m = n = 4, σ = 16, we conclude that this pattern √ has 1 a rotation symmetry by π/5, a dilation symmetry by τ = 2 cos(π/5) = 2 (1 + 5), the golden mean, and a mixed symmetry, consisting of a rotation by π/10 combined with a dilation by λ = 1.36. Incidentally, these examples show why it is safer to integrate over all scales in order to isolate the angular behavior, rather than to fix a certain scale a = ao and consider M[s](ao , θ). If ao coincides with one of the characteristic scales, a1 , a2 , . . . , the result is correct, but if ao falls in between, no maximum will be seen, and the symmetry is not detected. The effect is shown in Figure 4.27. Another interesting class of examples may be found in various pattern-forming phenomena in fluids [194]. Typically nonlinear waves at the surface of a fluid generate a regular pattern, via an unstability and a bifurcation. Most of these patterns have a

159

4.5 Detection of symmetries in patterns

(b)

(a)

Fig. 4.26. Symmetry detection with the CWT: (a) a Penrose tiling; (b) the corresponding scale-angle spectrum M[s](a, θ ), obtained with a Gaussian conical wavelet (γ = 10◦ , m = 4, σ = 16). Homologous maxima are linked by a line segment (from [24]). 550

270

500

260 250

450

240 400 230 350 220 300 210 250

200

200

150 0

190

20

40

60

80

100

(a)

120

140

160

180

180 0

20

40

60

80

100

120

140

160

180

(b)

Fig. 4.27. The scale-angle measure of the octagonal pattern from Figure 4.23, for fixed values ao of the scale: (a) for ln ao = 1.20, the periodicity is obvious; (b) for ln ao = 1.13, between two lines of maxima, the symmetry is not seen (from [24]).

rotational invariance of some order, and some of them are quasicrystalline. As a result, they lend themselves quite naturally to a wavelet analysis [86]. The best known case is the instability demonstrated by Faraday in 1831. The resulting pattern was known to have a rotational symmetry of order n = 12. Our standard analysis indeed yields this symmetry, together with √ the corresponding invariance under dilation by the corresponding factor β12 = 2 + 3 3.73. In addition, we find, as before, a combined √ symmetry of a rotation by 2π /24 = 15◦ together with a dilation by δ = 1.89 β12 ,

160

Applications of the 2-D CWT. I: image processing

which was unexpected [26]. It can actually be proved that such a combined rotation– dilation invariance, with half the rotation angle, is always present in all these “algebraic quasicrystals” [Bar94]. This technique permits one to determine, in a straightforward way, the (possibly hidden) symmetries of a given pattern. This applies to a genuine lattice, but also to a quasilattice, for which the symmetry is only local, for instance, the diffraction spectrum of a quasicrystal. Let us remind the reader that quasicrystals are those remarkable alloys discovered in 1984 [338], whose X-ray diffraction patterns show local n-fold point symmetry for n = 5, 8, 10, or 12. The latter are crystallographically forbidden for being incompatible with translational invariance (the only rotational symmetries compatible with lattice periodicity are of order n = 1, 2, 3, 4, or 6). These diffraction patterns display bright Bragg peaks of unequal intensity and they are self-similar with irrational scaling factors. More precisely, the involved irrationals are the following algebraic numbers: √ τ = 12 (1 + (pentagonal or decagonal quasilattices) √ 5) = 2 cos(π/5) β8 = 1 + √2 = 1 + 2 cos(π/4) (octagonal case) β12 = 2 + 3 = 2 + 2 cos(π/12) (dodecagonal case), that is, precisely the dilation factors discussed above. Similarly, some of the wave functions for transport electrons in quasicrystals are critical: they are neither localized (as would be the case in a random amorphous structure), nor spread out (as for perfect periodic crystals). Moreover, they display self-similarity too. For example, in the √ fivefold case, self-similarity ratios are typically powers of the golden mean τ = 12 (1 + 5). Thus we expect that the present wavelet-based method will yield interesting physical applications in the field of crystallography, in three possible directions. The first one concerns the diffraction patterns, where, at a given resolution level, it is necessary to classify and label the Bragg peaks according to their position and intensity. Secondly, two-dimensional wavelets based on the number τ seem particularly appropriate for the scanning analysis of patterns obtained through tunneling or atomic force microscopy of quasicrystalline surfaces. The third application concerns the determination of electronic wave functions in quasicrystals (explicit construction by using a discrete wavelet basis, adapted to the given symmetry type). Actually, these applications may be made easier if one uses a set of wavelets directly adapted to the symmetry. We will describe some of these in Section 11.4.2.

4.5.2.4

Point sets generated from noncrystallographic Coxeter groups A completely different kind of example is given by quasilattices derived from infinite dimensional extensions of noncrystallographic Coxeter groups. The idea is to generate an aperiodic point set by applying successive reflections and translations to the root system. The example we have analyzed is based on an affine extension of the Coxeter group H2 [27,303]. This group is isomorphic to the dihedral group of order 10 and it has a noncrystallographic root system (that contains explicitly the golden mean τ ). Then,

4.5 Detection of symmetries in patterns

0.9624 log(a/amin)

161

σ3

0.7218

σ4

ρ1

0.4812

σ1

ρ4

0.2406

ρ3

ρ5

ρ2

0 0

(a)

σ2

9

18 27 θ (in degrees)

ρ0 36

(b)

Fig. 4.28. (a) A H2aff -induced quasicrystal centered on (0, 0); the picture shows the part of the

quasicrystal located within a radius of 2; (b) result of the voting algorithm with a Gaussian conical wavelet (γ = 3.5◦ , m = 4, σ = 10) (from [27]).

applying a number of reflections and translations in a two-dimensional subspace of the root space, one generates planar aperiodic point set, called an H2aff -induced quasicrystal and shown in Figure 4.28(a). Next we apply the voting algorithm of Section 4.5.1. The result is shown in Figure 4.28(b). We observe six basic symmetries, labeled by ρ j with j = 0, . . . , 5. Among these, the pure operations may be considered as major symmetries, because the intensity of their maxima is dominant, namely the rotation by π/5, denoted ρ0 , and the dilation by τ , denoted ρ1 . All others are weaker symmetries. One of them, denoted ρ2 , is again a pure rotation symmetry, with angle π/10, but since the intensity observed is much weaker than the intensity of ρ0 , it is not possible to view ρ2 as the generator of the latter. In addition, we have combined symmetries, ρ3 , ρ4 , and ρ5 . The remaining symmetries, labeled by σ j , with j = 1, . . . , 4, are compositions of the basic ones. Indeed, σ3 , σ2 , σ4 , σ1 may be obtained by combining the pure dilation ρ1 with ρ4 , ρ3 , ρ5 , and ρ0 , respectively. It remains to provide a geometrical interpretation of all these symmetries. As building blocks of the quasiperiodic tiling, one may take a decagon contained in Figure 4.28(a), just like the octagons of Figure 4.24, and its successive translations. Each translated decagon is reproduced several times during the successive rotations of an angle of kπ /5 with k = 1, . . . , 9. Thus any symmetry found in this construction will be important for the whole tiling. Furthermore, any symmetry which is present already in the nontranslated decagon will appear with an even stronger intensity, because it appears more frequently and thus leads to a more dominant statistical symmetry. Then one finds that each of the symmetries σ2 and ρ3 corresponds to the ratio of two specific segments in the basic decagon, with the angle between them. In addition, the weaker symmetries σ3 , σ4 , ρ4 , and ρ5 can be traced to similar geometric relations in the figure [27].

162

Applications of the 2-D CWT. I: image processing

4.6

Image denoising A very successful application of wavelets in image processing is to denoising. Most of the authors so far have used the DWT for that purpose, and indeed orthogonal or biorthogonal wavelet bases. In some cases, wavelet frames have also been exploited and also the lifting scheme of Section 2.5.2.4 [Jan00,Jan01]. A notable improvement was obtained by E. Cand`es with a new tool, better adapted to images, called curvelets [Can98,96,97,346] (we will describe these in Section 11.1.4). It turns out, however, that the techniques described in the present book may be useful too. We have already seen an example in Section 4.2, namely, the first stage of the ATR algorithm. Although it is rather primitive, this technique has the advantage of being very simple. A much more elaborate approach is to take directional dyadic wavelet frames, introduced in Section 2.6.3. Without going into technical details, it is instructive to compare the two methods, and this we shall do here. The key notion in every denoising method is thresholding. The idea is that only the relevant features of an image yield significant wavelet coefficients, whereas the noise gives many small coefficients, spread more or less everywhere in the parameter space. Thus it suffices to put to zero all the coefficients that lie below a fixed threshold #. The problem, however, is how to choose the latter. A sophisticated and highly successful technique was introduced by Donoho and Johnstone [146,147]. In addition, one has to choose between two versions: r hard thresholding: here the small coefficients, i.e. |c | < #, are replaced by 0 and jk the rest remains untouched. As a consequence, artificial discontinuities are created. r soft thresholding or wavelet shrinkage: in order to remove these discontinuities, all the remaining coefficients are shifted by ±#, so as to make them continuous. This thresholding technique, initially developed in 1-D, extends to 2-D in an obvious way. The ATR method has been described in Section 4.2, and one observes that it uses a hard thresholding, with a threshold level # fixed arbitrarily. As for the dyadic wavelet packets, we start with a dyadic tight frame generated by a wavelet that satisfies the conditions (compare Proposition 2.4.1, in particular, (2.114)) L−1

ϕ − 2π /L)|2 = |ψ(r )|2 , |(r,

(4.10)

=0

) is the Fourier transform of an isotropic where L is the number of orientations, and ψ(r dyadic wavelet, i.e., +∞ j=−∞

j r )|2 = 1 |ψ(2

(4.11)

163

4.7 Nonlinear extensions of the CWT

and J

j r )|2 = |φ(2 J r )|2 , |ψ(2

(4.12)

j=−∞

where φ is the associated 2-D scaling function [265]. Using a directional frame allows us to put more redundancy in the technique and to benefit from the fact that directional wavelets will emphasize oriented features, such as edges, etc. Even though (2.140) [or (2.141)] is a continuous formula, all computations may be carried out in a discrete setting, either by means of the sampling theorem, or by using the approximate QMFs introduced in Section 2.6.4. It suffices then to compute the wavelet coefficients and to threshold them appropriately. The choice of the threshold is a crucial matter and usually requires an estimation of the standard deviation of the contaminating noise. Finally, reconstructing using the thresholded coefficients yields an estimated, denoised image. In order to illustrate the efficiency of the method, we present in Figure 4.29 a comparison between the ATR technique and the present one. We choose again our familiar L-shape, embedded in increasingly severe Gaussian noise, with standard deviation σ = 14 255, 12 255, and 255, and corresponding PSNR 12.07, 6.04, and 0.01 dB, respectively. In the middle column, we show the image denoised with the ATR method. The CWT is computed with an isotropic Mexican hat over five dyadic scales, 2, 4, 8, 16, and 32. Then, a hard thresholding is applied at each scale layer with a relative (with respect to the maximum) threshold decreasing with scale: 50, 25, 12.5, 6.25, and 3.125%, respectively. Finally, all these modified scale layers are summed, giving an image denoised, in the sense that the object has been considerably enhanced over the noise. In the right-hand column of the figure, we show the result obtained with a directional dyadic wavelet packet, using five scales and eight orientations. The denoised images have PSNRs equal to 25.97, 21.63, and 18.16 dB, respectively. The result is obviously better than the one obtained with the ATR method. As a second example, we show in Figure 4.30 the denoising of the lena image with the directional dyadic wavelet packet method, using five scales and 16 orientations. Panel (a) shows the original image, panel (b) the noisy version, obtained by adding to the original picture a white noise of σ = 25.5 (PSNR: 20.02 dB). Panel (c) shows the denoised image, which now has a PSNR of 31.07 dB.

4.7

Nonlinear extensions of the CWT

4.7.1

Local contrast The intensity of light around us varies considerably, in fact by several orders of magnitude. Our visual system is well adapted to this situation. Indeed it analyzes the spatial organization of the luminous field by relying on the contrast of objects and figures contained in the images. Intuitively, contrast is defined as the ratio between a variation

164

Applications of the 2-D CWT. I: image processing

Fig. 4.29. Comparison between the two denoising methods. The left column (a) shows the signal, with noise increasing from top to bottom, the middle one (b) is the result of the ATR algorithm, the right one (c) that obtained with a directional dyadic wavelet packet.

of luminance and a reference level of luminance. It is mathematically expressed using Weber’s law: CW =

L . L

(4.13)

165

4.7 Nonlinear extensions of the CWT

(a)

(b)

(c)

Fig. 4.30. Denoising of lena with a directional dyadic wavelet packet.

This definition is often used for small patches with a luminance offset L on a uniform background of luminance L. In the case of sinusoids or other periodic patterns of symmetrical deviations ranging from L min to L max , which are also very popular in vision experiments, one generally uses the Michelson contrast [Mic27], namely, CM =

L max − L min . L max + L min

(4.14)

While these two definitions are good predictors of perceived contrast for the abovementioned classes of simple stimuli, they fail when the stimuli become more complex and cover a wider frequency range, for example, Gabor patches [307]. It is also evident

166

Applications of the 2-D CWT. I: image processing

that neither of these simple global definitions is appropriate for measuring contrast in natural images, because the brightest and darkest points would determine the contrast of the entire image, whereas actual human contrast perception varies with the local average luminance. In order to address these issues and provide a quantitative definition of contrast, Peli [308] proposed a local band-limited contrast: C Pj (x, y) =

ψ j ∗ I (x, y) , φ j ∗ I (x, y)

(4.15)

where I (x, y) is the input image, ψ j is a band-pass filter at level j of a filter bank, and φ j is the corresponding low-pass filter. The normalization by the low-pass signal takes into account the local luminance variations. Modifications of this contrast definition have been used in a number of vision models [118,257] and are in good agreement with psychophysical experiments on Gabor patches [307]. The particular form of (4.15) suggests use of the wavelet transform for describing the variations of luminance. Now the WT is a space-scale analysis, and the spatial extension of the wavelets is characterized explicitly by their scale factor. Thus it is possible to define at each scale a different normalization, similar to a local average. So, following Duval–Destin [Duv91,12], one is led to the notion of local contrast, defined by combining the wavelet transform with an adaptive normalization. The latter will be obtained by projecting the signal, at a given scale, on a local weight function, chosen with the same localization properties as the wavelets. This local mean value will be called luminous level. This is the background against which luminance variations are measured, and the WT may be interpreted as a representation of these luminance variations within an image. The resulting contrast analysis is nonlinear, but it presents several advantages. It is particularly well adapted to the processing of positive signals. It also yields a multiplicative reconstruction process, which preserves positivity. Let us give some details and an example of application. Let h ∈ L 1 (R2 ) ∩ L 2 (R2 ) be a non-negative, rotation invariant, weight function, normalized to h L 1 = 1. Given an image, represented by a non-negative function f , the luminous level with respect to the weight function h is defined as = h (b,a) . Ma [ f ](b) h (b,a) x ) = a −2 h a −1 ( x − b) (4.16) | f , ( instead of the usual Note that we use throughout the L 1 -normalization, that is, h (b,a) h b,a (see Section 2.2). This is more natural in this context, since all the functions h (b,a) have the same L 1 -norm. Then we define the local contrast as the ratio of the CWT to the corresponding luminous level (the wavelet ψ is assumed to be also rotation invariant): = Ca [ f ](b)

Fa (b) Ma [ f ](b)

=

ψ(b,a) |f h (b,a) |f

=

ψb,a |f h b,a |f

,

(4.17)

167

4.7 Nonlinear extensions of the CWT 1 ˘ b, ≡ F( a) = ψ(b,a) where Fa (b) | f is the CWT of f with the L -normalization (but the local contrast is independent of the normalization). In order to make sense, this definition requires that the support of ψ be contained in the support of h. The local contrast is nonlinear, but its behavior is controlled by an integral condition. Large absolute values of contrast imply the existence of a region where the luminance signal is very small. A typical example, very natural in the study of vision, is to take for h a Gaussian and for ψ a Mexican hat. But one can do better and take for ψ the difference wavelet associated to h, as given in (3.13). Then the local contrast becomes

= Ca [ f ](b)

h (b,aα) |f h (b,a) |f

− 1,

(4.18)

and the existence condition is simply that the support of h be star-shaped. This formula in turn leads to a multiplicative reconstruction scheme. Indeed, estimates of the luminous level at smaller and smaller scale factors a may be considered as smoothened versions of the image with progressively contracted weight functions h. Then, as for the WT, the approximation of a function at a given scale may be written in terms of the approximation at a larger scale and the complementary signal: Maα [ f ] = Ma [ f ] · (Ca [ f ] + 1), Maα2 [ f ] = Maα [ f ] · (Caα [ f ] + 1)

(4.19)

= Maα [ f ] · (Ca [ f ] + 1) · (Caα [ f ] + 1), and, by recurrence: Maαn [ f ] = Maα [ f ] · (Ca [ f ] + 1) . . . (Caαn−1 [ f ] + 1).

(4.20)

Maαn [ f ] is the nth resolution approximation of f ; it is the image as seen through the smoothing function h contracted by a factor aα n (a < 1). One notices the obvious analogy with the usual multiresolution analysis (Section 1.5). The formalism may be generalized further to the so-called infinitesimal contrast analysis developed in [159]. An interesting application of the notion of local contrast is the design of an algorithm for the matching of stereoscopic images [313,314]. The technique consists in using a localized correlation function for comparing the two images, by means of the 2-D CWT. The locality of the latter provides a good estimate of the disparity between images, at each scale. However, in order to prevent the occurrence of artifacts due to the high sensitivity of the localized correlation function to the local mean value of the signals, one must normalize the latter in an adaptive way, and this leads precisely to local contrast described above. Peli and Duval–Destin’s definition of local contrast as defined above measures contrast only as incremental or decremental changes from the local background, which is analogous to the symmetric (in-phase) responses of vision mechanisms. However, a

168

Applications of the 2-D CWT. I: image processing

(b)

(a)

(c) Fig. 4.31. Peli’s local contrast (c) from equation (4.15) computed for the lena image (a) using an isotropic band-pass filter (b).

complete description of contrast for complex stimuli has to include the antisymmetric (quadrature) responses as well [347]. The problem is illustrated in Figure 4.31, which shows the contrast C P computed with an isotropic band-pass filter for the lena image. It can be observed that C P does not predict correctly the perceived contrast, as it varies between positive and negative values of similar amplitude at the border between bright and dark regions and exhibits zero-crossings right where the perceived contrast is actually highest. This behavior can be understood when C P is computed for sinusoids with a constant C M . The contrast computed using only a symmetric filter actually oscillates between ±C M with the same frequency as the underlying sinusoid, which complicates

169

4.7 Nonlinear extensions of the CWT

establishing a correspondence between such a local contrast measure and data from psychophysical experiments. These examples underline the need for taking into account both the in-phase and the quadrature component in order to be able to relate a generalized definition of contrast to the Michelson contrast of a sinusoid grating. Analytic filters represent an elegant way to achieve this. The magnitude of the analytic filter response, which is the sum of the energy responses of in-phase and quadrature components, exhibits the desired behavior, i.e., it gives a constant response to sinusoid gratings. Unfortunately, extending the Hilbert transform to 2-D is not a straightforward task. However, as already stressed in Chapter 3, directional wavelets offer a pleasant alternative, since the Fourier transform of the wavelet is included in a convex cone with apex at the origin and of aperture strictly smaller than π . This means that at least three such wavelets are required to cover all possible orientations uniformly, but otherwise there is no restriction on the number of filters. There are many applications where isotropy is required. In these cases, it is important to combine the analytic responses defined above into an isotropic contrast measure. Working in polar coordinates (r, ϕ) in the Fourier domain, we choose, as in Section ϕ) satisfying the above requirements and the conditions 2.6.3, a directional wavelet (r, (4.10)–(4.12) above. Note that the function φ in (4.12) need not be a scaling function associated to ψ, but it should at least have the same localization properties in order to provide for a meaningful normalization of the luminance level. Now it is possible to construct an isotropic contrast measure from the energy sum of directional filter responses [374]: /

2 | j ∗ I (x, y)|2 I C j (x, y) = , (4.21) φ j ∗ I (x, y) where j denotes the wavelet dilated by 2− j and rotated by 2π /L. If the directional wavelet belongs to L 1 (R2 ) ∩ L 2 (R2 ), the convolution in the numerator of (4.21) is again a square integrable function, and (4.10) shows that its L 2 -norm is exactly what would have been obtained using the isotropic wavelet ψ. C Ij is thus an orientation- and phase-independent quantity, but being defined by means of analytic filters, it behaves as prescribed with respect to sinusoidal gratings (i.e., C Ij (x, y) ≡ C M in this case). Examples for this isotropic contrast are shown in Figure 4.32. It can be seen that the contrast features obtained with C Ij correspond very well to perceived contrast. The combination of the directional analytic filter responses produces a naturally meaningful phase-independent measure of isotropic contrast. This technique may be applied for improving the contrast in any kind of image. An example of application to a photograph was given in [12]. Here we show one with a medical image (Figure 4.33). The image f is decomposed over N contrast levels, as in (4.20), using the couple Gaussian–DOG. For each level j, one defines the contrast chart as the modulus of the local contrast,

170

Applications of the 2-D CWT. I: image processing

(a)

(c)

(b)

Fig. 4.32. Isotropic contrast of the lena image as described by equation (4.21) at three different levels, 0, 1, and 2.

= |C2 j [ f ](b)|, M j (b) j = 1, . . . , N .

(4.22) 6N

as a measure of the = j=1 M j (b) Then one interprets the product of the N charts, S(b) After thresholding, correlation between the successive scales of the image at the point b. one obtains a binary image or mask. The latter is used in medical imagery, for instance, as a preprocessing to more sophisticated algorithms. It is taken as a priori knowledge and helps to reduce the amount of computation.

4.7.2

Watermarking of images Digital image watermarking consists in embedding a digital signature in an image. This operation is usually performed by slightly modifying the visual information in such a way that the perturbation is invisible to human eyes, but can still be recovered by using an appropriate algorithm. This embedding can be performed directly in the spatial domain, but also in the frequency domain (using DCT† coefficients) or, as we shall see now, in the wavelet domain. Finally, one often asks that the watermark should be robust, that is, it should survive common image alterations: geometrical image transformations, addition of noise, lossy compression or even print-scan procedure. We refer to [214] for further details. The generic picture of an image watermarking application is depicted in Figure 4.34. The inputs of the system consist in the original image, the watermark and an optional public or secret key. The watermark, or digital signature, can be of various nature: number, image or text. The key is used to encrypt the watermark and prevents it being read by unauthorized parties. In the sequel, we will mainly focus on the embedding part of the system and, more precisely, we will see how the wavelet transform can be used in conjunction with a vision model for robust and imperceptible image watermarking. †

DCT = discrete cosine transform (see [Mal99]).

171

4.7 Nonlinear extensions of the CWT

Fig. 4.33. Contrast analysis of a medical image: (a) the original image; (b) the CWT with a Mexican

j = −1; (d) the resulting binary image. Many more hat ( j = −1); (c) the contrast chart M j (b), details are seen on the two bottom images than on the ordinary CWT.

Watermark

Host Data

Digital Watermarking

Secret / Public key Fig. 4.34. Typical image watermarking system.

Watermarked Data

logCT

Applications of the 2-D CWT. I: image processing

ε

CT0

172

CM0

logCM

Fig. 4.35. Contrast masking model describing the relationship between the masker contrast and the target contrast at detection threshold.

In order to model the visibility of the watermark in the original image, there are mainly two effects that need to be taken into account, namely, contrast sensitivity and masking. Contrast sensitivity, as we have seen previously, describes the response of the human visual system to the contrast of a stimulus. Masking, on the other hand, describes the phenomenon in which a signal, the masker, is capable of “hiding” a second signal, the target. In other words, the target visibility depends on the presence of a masker. It is possible to combine contrast sensitivity and masking in a model that describes the relation between the masker contrast and the target contrast at detection threshold. Figure 4.35 shows such a model where we have represented on the horizontal axis the logarithm of the masker contrast CM , and on the vertical axis we have the logarithm of the target contrast CT . The curve is divided into a threshold range, where the target detection threshold is independent of the masker contrast, and a masking range, where it grows as a power of the masker contrast. The mathematical description of this model is given by: C T0 if CM < CM0 , ε CT (CM ) = (4.23) CT0 CM /CM0 otherwise. The model contains three parameters, ε, CT0 and CM0 , which specify the size of the threshold and the masking range as well as the slope of the transducer function. They have to be determined by means of subjective experiments.

173

4.7 Nonlinear extensions of the CWT

This model is applied in [359] to a watermarking scheme based on spatial spreadspectrum modulation, as proposed by Kutter [Kut99]. Each bit to be embedded in the image is represented by a two-dimensional pseudo-random pattern. The statistics of the pattern are bimodal with equal probabilities for −1 and 1. The random patterns of all bits are superimposed as follows: w(x, y) = α(x, y) pi (x, y), (4.24) i

where pi (x, y) are the pseudo-random modulation functions for bit i, α(x, y) is the watermark weighting function, and w(x, y) is the resulting watermark which is added to the image. In this watermarking scheme, the pseudo-random patterns pi are sparse, which means that the superposition of all patterns does not necessarily modify all pixels in the image. To quantify the sparseness, we introduce the density D of the watermark, which is given by the modified number of pixels divided by the total number of pixels in the image. The watermark weighting function α(x, y) is computed using the introduced masking model and the local isotropic contrast measure presented in Section 4.7. For computing the local contrast according to (4.21), we use directional wavelet frames as described in Section 2.6, based on the scaling functions of Table 2.1. The minimum number of orientations required by the analytic filter constraint, i.e., an angular support smaller than π , is three. The human visual system emphasizes horizontal and vertical directions, so four orientations should be used as a practical minimum. To give additional weight to diagonal structures, eight orientations are preferred. We only use the highest frequency band of the pyramidal decomposition, because masking is strongest when masker and target have similar frequencies. Furthermore, higher levels tend to smear the local contrast and are thus not suitable for this kind of application. The watermark weighting function α is now computed as follows: α(x, y) = CT (C0I )(x, y) · φ0 ∗ I (x, y),

(4.25)

where C0I is the local isotropic contrast of the masker image at level 0, CT is the corresponding target contrast threshold as given by our masking model, and φ0 is a low-pass filter. The local amplitude of the watermark at the threshold of visibility is thus determined by the multiplication of the isotropic contrast values with the corresponding low-pass filtered image. Finally, the parameters of our vision model (CT0 , CM0 and ε) have been determined by performing subjective tests [359]. Figure 4.36 shows weighting masks for the lena image at watermark densities of 0.4 and 1, respectively. For illustrative purposes, the figures to the right visualize the segmentation into threshold and masking ranges. The dark areas correspond to regions where only contrast sensitivity is exploited, and the bright areas show regions where the masking effect is dominant.

174

Applications of the 2-D CWT. I: image processing

(a) Weights for D = 0.4

(c) Weights for D =1

(b) Segmentation for D = 0.4

(d) Segmentation for D = 1

Fig. 4.36. Watermark weighting function of the lena image for two different density values (left column). The segmented images (right column) illustrate the threshold and masking ranges of the contrast masking model, represented by dark and bright areas, respectively.

In comparison with other watermarking schemes, this weighting mask based on the simple masking model presented above facilitates the insertion of a watermark with higher energy while preserving the visual quality of the image, leading to a watermark that is more robust. It has also been applied successfully to watermarking the blue channel of color images [251].

5

Applications of the 2-D CWT. II: physical applications

In the previous chapter, we have discussed a number of applications of the 2-D CWT that belong essentially to the realm of image processing. Besides these, however, there are plenty of applications to genuine physical problems, in such diverse fields as astrophysics, geophysics, fluid dynamics or fractal analysis. Here the CWT appears as a new analysis tool, that often proves more efficient than traditional methods, which in fact rarely go beyond standard Fourier analysis. We will review some of these applications in the present chapter, without pretention of exhaustivity, of course. Our treatment will often be sketchy, but we have tried to provide always full references to the original papers.

5.1

Astronomy and astrophysics

5.1.1

Wavelets and astronomical images Astronomical imaging has distinct characteristics. First, the Universe has a marked hierarchical structure, almost fractal. Nearby stars, galaxies, quasars, galaxy clusters and superclusters have very different sizes and live at very different distances, which means that the scale variable is essential and a multiscale analysis is in order, instead of the usual Fourier methods. This suggests wavelet analysis. Now, the main problem is that of detecting particular features, relations, groupings, etc., in images, which leads us to prefer the continuous WT over the discrete WT. Finally, there is in general no privileged direction, nor particular oriented features, in astrophysical images. All this leads us to use the CWT with an isotropic 2-D wavelet. In addition, astrophysical images are very noisy. In particular the bright sky and our own galaxy (the Milky Way) represent noise, which must be removed, with a technique similar to that used in 1-D for the subtraction of unwanted lines or noise in spectra [210]. Here statistical techniques play an essential rˆole. All these considerations characterize the type of wavelet applications that have been developed in astronomy and astrophysics. The first attempt to apply the CWT to astrophysical images is due to the group of A. Bijaoui in Nice, in 1990 (see [80] for a review). In their pioneering paper [343], they

175

176

Applications of the 2-D CWT. II: physical applications

used the CWT for the analysis of galaxy clusters, with the 2-D Mexican hat. Similar techniques were exploited by a large number of authors, especially in the last few years. In this section, we will review some of this work, and also present two novel applications, one to solar physics, the other one to the detection of gamma-ray sources in the Universe.

5.1.2

Structure of the Universe, cosmic microwave background (CMB) radiation In several papers [161,162,193,343,344], the authors exploit galaxy counts to identify galaxy groupings, from compact groups (0.5 degrees or tens of kpc† in extent) to clusters (down to 1 degree, from hundreds of kpc to some Mpc), to large-scale structures or superclusters (5 degrees or tens of Mpc or more), including the determination of a possible hierarchy between them. The same technique allows the detection of voids, that is, large regions (up to 60 Mpc) with very few galaxies, and also to a neat definition of each of these notions. The results of such work leads to the analysis of the largescale structure of the Universe, thus to cosmological considerations. For instance, the distribution of groups of various size and of voids points to a possible fractal structure of the Universe. On the other hand, the multiscale approach yields much valuable insight into the inner structure of individual clusters [193]. Here, as in all papers analyzing galaxy maps, the basic data is a bidimensional distribution of Dirac delta functions, possibly weighted according to some statistical criterion. The same type of data will be used in the next two sections. A byproduct of such hierarchical analyses is the multiscale vision model developed by Bijaoui and his group [78,79] in order to detect and characterize structures of different sizes (for numerical reasons, also linked to the necessity of denoising the images, they later switched to a discrete WT, based on spline wavelets). For instance, they propose in [254] a morphological indicator allowing a comparison between various cosmological models (for instance, cold versus hot dark matter). In the same vein, a group from Santander, Spain, has undertaken a systematic analysis, by wavelet methods, of the COBE data on the cosmic microwave background (CMB) radiation. As a first step, they study the local (i.e., in small sky patches) temperature anisotropies in the CMB, including denoising the images [332,333]. In these papers, the authors use both the CWT and the DWT (the latter especially for denoising). As for the former choice, they first consider a 2-D CWT without a rotation parameter, but with two independent scalings in the x and y directions, then the usual isotropic Haar and Mexican hat wavelets. Next [100,361], they use isotropic wavelets to detect and determine the flux of point sources superimposed on the CMB, in conditions simulating the Planck Surveyor mission. As they point out, the advantage of the wavelet method is that no assumption has to be made regarding the statistical properties of the point †

1 pc = 1 parsec = 3.26 light years.

177

5.1 Astronomy and astrophysics

source population or the underlying emission of the CMB. Since the CMB observations are performed with antennas that are best modeled by a Gaussian beam, it turns out that the isotropic Mexican hat wavelet is in fact optimal for detecting point sources. In a further work [362], a detailed comparison is made of the wavelet method with the standard maximum-entropy method. The conclusion is that the two methods are in fact complementary and can be combined to improve the accuracy of the detection. More recently, the Santander group has turned to a global analysis of the CMB, trying to detect potential non-Gaussian CMB temperature fluctuations. This is an important observation for cosmology, for any non-Gaussianity would be evidence for a departure from standard inflationary theories. Since the data used in these experiments is the full sky COBE-DMR data, it is clear that the sphericity of the data has to be taken into account. As a consequence, one has to resort to spherical wavelets. A first attempt was made by Barreiro et al. [62], using discrete spherical Haar wavelets, constructed with the lifting scheme of Schr¨oder and Sweldens [336], described in Section 2.5.2.4. Then, following the same logic that recommends the use of the CWT with an isotropic Mexican hat, the Santander group introduced the spherical Mexican hat (see below), establishing the superior capability of the latter over the Haar wavelets [267]. The net result of these investigations is that the CMB temperature fluctuations are consistent with a Gaussian distribution, thus vindicating the standard theories [101]. Finally [363], the same group has used the same spherical Mexican hat wavelet for extending their previous work [361] on simulated Planck maps, thus achieving a large catalog of potential point sources. Coming back to the CWT on the 2-sphere S 2 , a mathematically precise transform was constructed in [29] and will be discussed at length in Chapter 9, Section 9.2. The idea is simply to take the plane R2 as the tangent plane at the North Pole of S 2 and lift functions on R2 to functions on S 2 by inverse stereographic projection. Introducing polar coordinates both on the plane and on the sphere, the correspondence reads: θ S 2 & (θ, ϕ) ⇐⇒ (r, ϕ) ≡ (2 tan , ϕ). 2 For square integrable functions, this leads to a unitary map between the respective Hilbert spaces, I −1 : L 2 (R2 , d x) → L 2 (S 2 , sin θ dθ dϕ), namely, (I −1 f )(θ, ϕ) =

θ 2 f (2 tan , ϕ). 1 + cos θ 2

(5.1)

In the case of an isotropic wavelet ψ(r ), with r = | x |, the correspondence is simply (I −1 ψ)(θ ) =

2 θ ψ(2 tan ). 1 + cos θ 2

(5.2)

Then, choosing for ψ the isotropic Mexican hat wavelet ψH , one gets the spherical Mexican hat wavelet:

178

Applications of the 2-D CWT. II: physical applications

ψH,sph (θ ) = 4

1 − 2 tan2 θ2 exp(−2 tan2 θ2 ). 1 + cos θ

(5.3)

It should be noted that, in this parametrization, the scaling x → a −1 x in the tangent plane becomes on the sphere: tan θ2 → a −1 tan θ2 . We refer to Chapter 9 for the full analysis. Another research area where the CWT has been used is the detailed analysis of individual galaxies, notably in the group of P. Frick [173,174]. Of particular importance is the cross-correlation between images obtained at different wavelengths. To that effect, the authors of [174] consider the (normalized) wavelet cross-correlation function or wavelet correlation coefficient, obtained by polarization from the wavelet spectrum (2.61) W[s](a): C[s1 , s2 ](a) =

a) S2 (b, a) d 2 b S1 (b,

(W[s1 ](a) W[s2 ](a))1/2

,

(5.4)

and originally introduced by Hudgins et al. [227]. Using this tool, the authors study a particular spiral galaxy, called NGC 6946, comparing the images of total radio emission, red light and mid-infrared dust emission on all scales. Note that in their treatment they use both the Mexican hat wavelet and, for a better separation of scales, their own isotropic wavelet, called Pet hat and defined in (3.11). In a later work from Frick’s group [175], spherical wavelets (in a somewhat primitive form) are used for isolating coherent structures in the distribution of the Faraday rotation measure of extragalactic radio sources, that is, a weighted integral of the longitudinal magnetic field along the line of sight. In addition, since these sources are given as irregularly distributed points in the sky, they adapt to the 2-D spherical situation the technique of gapped wavelets introduced previously in 1-D [172]. A final application in astrophysics, still under development, is to gravitational lensing, namely, the detection of Einstein arcs in cosmological pictures [82]. Whenever the light from a distant bright object (a quasar) is seen through a galaxy, the latter behaves as a gravitational lens, so that the point source appears as a ring, or a portion of a ring (“arclet”), if the alignment is not exact. By measuring the radius of that ring, one may infer the distance of the source. This may be done in two steps. The center of the arc is obtained with an annular-shaped wavelet, such as the Bessel filter (3.9), Frick’s wavelet (3.11), or the annular Halo wavelet (3.12), used at a rather large scale (e.g. a = 2). This determination is quite robust to noise, in particular, to spurious bright points, that mimic nearby stars. The arc itself is obtained with an isotropic Mexican hat, at a smaller scale (e.g. a = 0.5). By superposing the two transforms and applying a severe thresholding (up to 95%) for eliminating the noise, one obtains an image with three bright spots: two points of the arc, around the endpoints, and the center of the corresponding circle. From this, one can reconstruct the arc unambiguously, and thus one obtains a tool for measuring in a simple way the distance of quasars, for instance.

179

5.1 Astronomy and astrophysics

Before turning to more specific 2-D applications, we hasten to add that 1-D wavelet analysis has been used in various current problems is astrophysics. A case in point is the analysis of the solar neutrino capture rate data from the Homestake experiment [215], a crucial ingredient in the resolution of the celebrated solar neutrino problem.

5.1.3

Application of the CWT in solar astronomy Since 1996, the Extreme-ultraviolet Imaging Telescope (EIT) on board the Solar and Heliospheric Observatory (SoHO) satellite observes the Sun in four wavelengths: 171, ˚ These correspond respectively to particular emission lines of 195, 284 and 304 A. Iron (IX-X, XII, XV) and Helium (II), and thus to temperatures typical of those of the Sun corona in the first three wavelengths, and of the transition region in the fourth one [130]. The Sun corona is physically very complex and contains a huge amount of different events appearing at different locations and scales. Solar astronomers are interested in the physics which can be deduced from them in order to improve our knowledge of the global Sun. One way to achieve this is to make time statistics on features of special solar objects. In addition, because of the large number of EIT pictures (currently greater than 100 000), astronomers aim at an automatic analysis. However, many conceptual problems arise due to the difference between the human description of things and the true (logical) computer vision. These can be summarized into two main questions. r How to define a Sun corona object in simple terms, that is, in sufficiently simple concepts which can be managed by a computer program? r How to determine the relevant characteristics of such an object and how to translate them as simply as possible? After a short description of the common Sun corona objects, we will show that the continuous wavelet transform (CWT) offers tools to answer these fundamental questions (we basically follow [37]). Notice also that Bijaoui and his group have applied their vision model to the analysis of EIT images of the solar corona [316].

5.1.3.1

Special coronal objects The physical objects of the Sun corona result in general from convective motions in the solar mantle and/or of magnetic interactions with hot material. Here is a list of the principal objects, ordered by size, from the smallest to the largest (for more information, see [130,278]). ˚ images, the magnetic network constitutes a Magnetic network: In the red 304 A textured solar background resulting of the advection of small magnetic flux by the convective motion in the solar mantle. Brightenings: Brightenings are visible in all EIT images and are related to magnetic topology changes to a lower energy state. Their typical scale in an image is close

180

Applications of the 2-D CWT. II: physical applications

of the pixel size, but they brighten and fade away on a time scale ranging from several minutes to hours. Flares: A sudden and energetic local brightening in an active region (see below). Bright points: Bright points (BP) are small regions with enhanced emission. They are located above pairs of magnetic features of opposite polarity in the photosphere. We can see them in the quiet corona and in coronal holes. They present a lifetime ranging between two hours and two days. Magnetic loops (or Loops): These objects result from the filling of magnetic field lines with plasma. Because the temperature of this material varies along the loop, ˚ because they are cooler the footpoints of the loop are more precise in the 171 A, ˚ The magnetic loops may than the loop summit, which is better seen in the 195 A. be part of the same active region, connecting two regions of opposite flux, or even join different ARs. Active regions: They show up as a region of large increase in the ultraviolet flux on the image. Their typical size is about 10% of the solar radius. Physically, these active regions (AR) contain hot material in smaller and larger loops around and inside a region of enhanced magnetic flux. Because active regions are deeply related to the well-known Sun spots, they appear in two bands of latitude according to the evolution of the main solar cycle of 11 years: They live at high latitudes at the solar minimum (beginning of the cycle) and move towards the equator at the solar maximum. Coronal holes: Coronal holes (CH) are large regions where the magnetic field lines are open and are advected by the solar wind into interplanetary space. Because the energy is advected away, the CH are colder than the closed magnetic field regions and they appear effectively like dark holes in the EIT images. Their morphology evolves with time and they become very small during the Sun maximum. A visual summary of the solar objects defined above is presented in Figure 5.1. There are also features which are not related to the Sun physics, but either to defects of the SoHO satellite due to its aging, or to its interaction with some external events. The main ones are the cosmic ray hits, which, by the interactions of cosmic rays with the EIT CCD camera, plague the images with many bright pixels or bright straight lines, depending on the cosmic orientation relatively to the CCD surface. We should mention finally that all these images have a noise component, namely, a readout noise coming from the CCD camera, the solar noise and the photon-shot noise (Poisson noise). The global noise is well approximated by Gaussian statistics, because of the high counting effect (central limit theorem).

5.1.3.2

Distribution of small features To start with, let us define the specific wavelet tools needed for the present application, which aims at selecting some of the solar corona phenomena. We restrict ourselves to

181

5.1 Astronomy and astrophysics

˚ wavelength image, the Fig. 5.1. The main Sun corona objects. The top left quadrant is the 304 A ˚ rest corresponds to 171 A.

isotropic wavelets, since directions are irrelevant in the present context. Given an image a) with respect to a Mexican hat and the corresponding s( x ), we consider its CWT S(b, a)|2 . Then, as in Section 2.3.5, we define ridges R j , energy density E[s](b, a) = |S(b, and the corresponding amplitude A j (2.63) and slope S j (2.64). These two parameters are sufficient for the detection and discrimination of small features contained in the image s. A precious tool to that effect is the histogram of the amplitude as a function of the slope, or the slope–amplitude histogram. Let a0 the smallest relevant scale. Choose a sequence {b j , 0 j K − 1} of max a0 ), belonging to ridges {R j , 0 j K − 1}. Then, given the set of ima of E[s](b, all corresponding couples (S j , A j )0 j K −1 , the histogram is built by the following simple algorithm: r determine the desired size of the histogram H, say M × N , and initialize H as the zero M × N matrix; r compute S min and Smax , the minimum and the maximum of all the slopes (S j )0 j K −1 , respectively; r compute A min and Amax , the minimum and the maximum of all the amplitudes (A j )0 j K −1 , respectively; r form the discretized slope S m = Smin + m (Smax − Smin ) for 0 m (M − 1), M−1 n = Amin + n (Amax − Amin ) for 0 n (N − and the discretized amplitude A N −1 1);

182

Applications of the 2-D CWT. II: physical applications r

then, for k = 0, . . . , K − 1: – take the slope Sk and compute the index m such that Sm is the nearest discretized slope from Sk , that is, the index m such that −0.5 < (M − 1)

Sm − Sk 0.5; Smax − Smin

(5.5)

– do the same with the amplitude and determine the index n such that −0.5 < (N − 1)

n − Ak A 0.5; Amax − Amin

(5.6)

– increment the (m, n) entry of H by one, Hmn := Hmn + 1. The histogram H reflects the 2-D distribution of the slope–amplitude couples. Identifying distinct areas inside H is equivalent to detecting different classes of small objects contained in the image s. Before computing any histogram, a practical remark must be made about the difference between the continuous theoretical world and the discretized view of the programming. Indeed, actual computation requires an adequate sampling of the image s and of the wavelet ψ. Therefore, the scale a cannot effectively go to zero in (2.63) and (2.64). Indeed, the wavelet ψ must be sampled sufficiently on the grid determining the is essentially contained image. Thus, there will be a minimal scale a0 for which ψ b a0 in the frequency domain [−π, π ) × [−π, π ) (assuming the sampling period T is equal to 1).

5.1.3.3

Analysis of academic objects To test our method, we begin by analyzing two types of objects that will model small features in EIT images. Take first the smallest possible object, a singularity of height c localized on a point u, represented by a Dirac distribution s( x ) = c δ (2) ( x − u). One readily computes the CWT of s and the corresponding energy density a) = E[s](b,

+ c2 ++ +2 . u − b)) ψ(a −1 ( 4 a

(5.7)

then E[s] is maximum in b = u It is easy to see that, if ψ has a maximum in x = 0, for all scales. The equation of the associated ridge is simply ( u , a) for all a ∈ R+ . The amplitude of this ridge is given by ln Au = −4 ln a0 + ln c (thus it tends to ∞ as a0 → 0) and the corresponding slope has the value −4. The second object is a simple Gaussian localized in w, of width σ and height D, s( x ) = D exp(− 12 | x − w| 2 /σ 2 ). A detailed calculation shows that E[s] has also a vertical ridge localized in b = w with a maximum in a = σ . The amplitude of this ridge is proportional to D and the slope is now positive.

183

5.1 Astronomy and astrophysics

Fig. 5.2. Analysis of singularities and Gaussians. (a) The original academic image; (b) the

slope–amplitude histogram (the logarithm of the amplitude is plotted to reduce the range); (c) the selection of points in the singularity population (triangles), or in the Gaussian population (circles).

These two examples show that the amplitude yields a criterion for selecting small objects according to their intensity. Then the slope decides between a singularity or a larger object modeled by a Gaussian. The procedure is illustrated in Figure 5.2. We analyze an academic image s of size 256 × 256, shown in panel (a), and consisting of a collection of randomly placed singularities and Gaussians of small size, and compute the corresponding slope–amplitude

184

Applications of the 2-D CWT. II: physical applications

histogram, knowing that, in this case, the minimal scale a0 of the Mexican hat is close to 0.9. In Figure 5.2(b), we clearly see two distinct populations in the slope–amplitude histogram. The population on the left-hand side (left dashed circle) corresponds to singularities of s with a slope centered around −4. The right area (right circle) corresponds to the Gaussians. A rough selection of points according to the sign of the slope is made in Figure 5.2(c). Negative slopes are represented by triangles, and positive ones by circles; singularities and Gaussians are effectively selected separately.

5.1.3.4

Application to EIT images We can now apply the preceding technique to the selection of cosmic ray hits and of bright points in the EIT images. The former are well described by singularities, because cosmics burn only a few pixels on the CCD camera of the satellite, and the latter can be modeled by Gaussians of small size. ˚ The analyzed EIT image, shown in Figure 5.3(a), is the top-left quadrant of a 284 A wavelength image of the Sun. The slope–amplitude histogram, computed for a0 = 1, is presented in Figure 5.3(b). We notice that the distinction between populations is not as neat as in the academic example. The reason is that the white noise present in the picture recording has a main effect of spreading the well-defined areas of Figure 5.2(b). Next, we impose on the histogram of Figure 5.3(b) an additional selection criterion. For cosmics, we choose the maxima b j of E[s](., a0 ) such that ln A j > 2 and S j < 0 and, for bright points, those with S j > 0. The amplitude thresholding prevents us from taking singularities that are too faint coming from quantization and Gaussian noise. The result is shown in Figure 5.3(c). The cosmics are detected everywhere in the image (because they are not related to solar physics), while the bright points appear mainly on the solar disk (on-disk). In Figure 5.4, we make a zoom on a particular on-disk area of the Sun. The selection effect is now clearer than in the global image.

5.1.3.5

Conclusion and open questions We have presented in the previous section a simple method based on the CWT to discriminate two kinds of simple events in the Sun corona pictures, the cosmic ray hits and the bright points. However, there is ample room for improving the range and efficiency of the method. First, noise has been suppressed in the slope–amplitude histogram by a hard thresholding. However, a precise statistical study remains to be made on these selections according to the SNR of the analyzed images and their CWTs. Then, one may try to characterize more complex solar phenomena, such as the active regions, the magnetic loops or the textured magnetic network. A possibility is to exploit the full information carried by the vertical ridges of the CWT. Indeed, the method described above uses only the first relevant scale of the latter, that is, only their

185

5.1 Astronomy and astrophysics

˚ wavelength EIT image; (b) Fig. 5.3. Analysis of an EIT image. (a) The top-left quadrant of a 284 A the slope–amplitude histogram; (c) the selected cosmics (triangles) and bright points (circles).

beginning. Information about possible maxima of E[s] along these ridges is interesting too, for instance for determining the typical scale which defines each type object, as in the 1-D analysis of impact experiments [358]. Several hierarchical criteria based on the CWT may also help us to detect the inclusion of small events into larger ones, such as the brightenings within the active regions. The shape of strong response areas in the CWT at different scales could be useful in this context, as found in [344], for instance.

186

Applications of the 2-D CWT. II: physical applications

Fig. 5.4. A closer look on a small on-disk region of the Sun: (a) Bright points selection; (b) cosmics

selection.

Finally, directional wavelets may be useful, too, since many solar events present an anisotropic behavior. The magnetic loops, for instance, are locally equivalent to straight lines characterized by a particular width. At a scale proportional to this width, the anisotropic CWT coefficients should vary for different angles θ relatively to the main direction of this line. A magnetic loop signature could perhaps be found inside this variation (this is a variation on the problem of detection of oriented contours).

5.1.4

Detection of gamma-ray sources in the Universe Another topic where the CWT has been applied successfully is the analysis of the X-ray structure of various objects, such as clusters of galaxies, following a suggestion by Grebenev et al. [201]. This leads to a different class of problems. Indeed, such sources are frequently at the limit of detection, so that statistical considerations become crucial. In particular, we are here often in the photon-counting regime, the photon per pixel statistics is significantly different from Gaussian and most sources are extended. The analysis of such images by wavelet methods was further developed by Damiani et al. [115]. The work reported here uses a similar approach, but goes significantly deeper in the analysis. In particular, considerable care is devoted to the presence of Poisson noise in the photon flux. We again follow [37]. When it comes to the detection and analysis of gamma-ray sources in the Universe, the way data analysis is carried out depends on the energy range explored by the telescope. This is not only due to the nature of emitting objects, and to the difficulty of designing appropriate detectors, but also to the gradually lower photon counting rate

187

5.1 Astronomy and astrophysics

as the energy increases. For example, in average 100 ultraviolet photons from the Sun are expected to be detected in one second by each pixel of the SoHO CCD camera (Section 5.1.3), whereas about 1 gamma photon is recorded by the whole gammaray space telescope EGRET during the same period! This correspondingly decreases accuracy and significance of any statistical decision, like event detection. Equally important for the data analyst, the nature of photon-counting processes induces an intrinsic “noise,” called Poisson noise, requiring more statistical care than the usual Gaussian noise. The problem we address in this section is the detection of sources in the raw data of the above-mentioned telescope EGRET (20 MeV – 30 GeV photons). Sources are pointlike objects like pulsars or active galactic nuclei and appear in the data as a few detected photons coming from the same direction in the sky. The whole issue is to give a meaning to the coincidence of finding these photons together, hence to conclude (or not) that they were produced by chance from the diffuse background (interaction of cosmic rays with interstellar clouds). In addition to the detection significance, the position, magnitude and spectral characteristics of a source are other desirable quantities determined from the data. This may seem a very humble problem to solve, but, as outlined above, the scope of questions one can answer at 1 GeV is considerably restricted compared to the wealth of the analysis in the previous section.

Sample data and the classical solutions Every dot on Figure 5.5 is a detected photon, of energy 100 MeV or above, during the viewing period 21.0 of EGRET. A “position” on this counting map refers to a direction in the sky, and the map is modeled as approximately flat. Such counting maps are very broadly modeled as counting (Poisson) processes from two contributions, the background flux and the flux from the sources. That is, we do not directly observe light Declination

5.1.4.1

200 30

180 25 160

140 20 120 15

100 10

80 5 60

40 0 20 −5

0 0

70

50

60

100

50

150

40

200

30

250

20

10 300 Right Ascension

Fig. 5.5. Detected photons above 100 MeV during EGRET viewing period 21.0.

188

Applications of the 2-D CWT. II: physical applications

intensities, but rather photons that are randomly created from the corresponding physical objects (point-like objects or extended objects like interstellar gases). Moreover, the detector is far from being perfect: direction and energy of an incoming photon are recorded with an error that translates into the convolution of the above-mentioned fluxes by a bell-shaped function, the PSF (point-spread function), which is here more heavy-tailed than a Gaussian [268]. As most recognition tasks in data analysis, gamma-ray source detection is often carried out by a maximum likelihood (ML) method. That is, a parametrized source model is fitted to data through maximization of the probability that the data arose as a realization of the suggested model. This involves a heavy nonlinear optimization procedure and an initial guess from the user to set the parameter values (height, width and position, say) to their optimal values. But one eventually ends up with a very faithful account of the physical properties of each source. There are statistical reasons to think that it is hard to beat the quality of ML estimation, like the minimum variance property (see [Ead71]). The reference for ML source detection in EGRET data is [268]. Our concern, however, is to develop a simpler method of source detection based on the continuous wavelet transform. Roughly speaking, the idea is to group the events in a chosen interval of energies into a single 2-D counting map, as in Figure 5.5, and to take its CWT with an isotropic wavelet, typically a Mexican hat. The source candidates are the maxima of the wavelet transform. Then, based on some statistical criterion, a detection significance will be given to each maximum. The higher the significance, the more likely the candidate to be a true source. In order to give the status of source candidate to the maxima of the wavelet transform, we must make sure that relevant information (the sources) is properly decorrelated from noise (the background). For this purpose, the Mexican hat wavelet is a good choice, because: (i) its isotropic bell shape responds mostly to bell-shaped sources; (ii) its good localization in space allows to discriminate events according to their position, more efficiently than the Laplacian of the heavy-tailed PSF; and (iii) its good localization in the frequency plane permits us to discriminate events according to their relative scale. The statistical performance of a wavelet method is presumably poorer than what can be achieved by ML. However, the latter has two drawbacks, namely, high computational complexity of the implementation and supervision of the optimization process. These issues now simply disappear, since wavelets allow a real-time automatic processing of the same job. To stress again the difference between the two tools, we can think of wavelet methods as providing an initial guess for more thorough ML identification of the source parameters, or as an on-board data processing module to warn the astronomers in case of a sudden gamma-ray burst. Before giving the details of our procedure, let us mention that, as in the analysis of the distant Universe [78,79], related problems like the identification of extended objects, that is, multiscale structures, have also been attacked with the discrete WT

189

5.1 Astronomy and astrophysics

[Sta98,345]. As usual, this allows a faster implementation, but the choice of filter is severely restricted, in particular, the Mexican hat is not admissible. Thus a balance must be made between quality of the analysis and speed of implementation. We can also note at this point that the “almost flat” approximation used above, namely that a direction in the sky (a point on the sphere) can locally be represented by two planar coordinates, is not necessary. If a more global data analysis is required, we can always switch to the genuine spherical wavelet transform, mentioned in Section 5.1.2, and discussed at length in Section 9.2. This makes no conceptual difference in what follows, only the algorithms will be more CPU-time consuming.

5.1.4.2

Decision criteria and results It remains to describe the procedure itself. In what follows, we will concentrate on the wavelet aspects and skip most of the statistical arguments, which may be found in detail in [37]. The problem may be subdivided in a series of questions, each of which requires to choose a decision criterion, following more specific questions. (i) What is the detection criterion? Or in other words, how big should the values of the wavelet transform be to conclude that a peak is indeed a source? Our criterion is based on a physical model of the background interstellar gamma-ray emission, related to the distribution of hydrogen in the galaxy. The idea is to measure the discrepancy between this model and the data in the wavelet domain. Peaks will be considered sources if they significantly overshoot the model. (ii) How to estimate the total photon flux from a source? Intuitively, the bigger the value of the wavelet transform at the position of a source, the larger the flux of the source; and, since the wavelet transform is linear, this relation should be linear too. That is correct, modulo complications due to the presence of an unknown background of magnitude comparable to that of the source. We can however use our coarse a priori background model to remove part of this bias. Our estimator of the flux of a source detected to be at position x is not the value y obs (x) of the wavelet transform of the observed data, at scale a and at position x, but rather =

y obs (x) − W [µB ](x) , W [µS ](x)

where W [µB ](x) and W [µS ](x) are the wavelet transforms at x (and at a given scale, as always) of the modeled background flux and of the modeled source flux, respectively. This estimator can be shown to be asymptotically unbiased (i.e., when the exposure time or all the fluxes tend to infinity), provided the models are correct. Confidence intervals on this statistic can also be derived.

Applications of the 2-D CWT. II: physical applications

(iii) How to estimate the position of a source? Intuitively, the source candidates should be located at the maxima of the wavelet transform. As above, this is true only for flat backgrounds. We account for its nonuniformity in a correction to the quantity to maximize in order to get the position of the source candidate. The position estimator is thus not argmaxx y obs (x) but rather % $ x ∗ = argmax y obs (x) − W [µ B ](x) . x

Again, assuming the models are correct, this subtraction restores asymptotic unbiasedness. Confidence regions corrected in this way may be seen on the example presented in Figure 5.7 below. (iv) How to choose the scale of the wavelet for best performance? Since the physical source is point-like, recorded sources look like the impulse response of the detector, i.e., the PSF. Hence, the best choice for the scale parameter is that leading to a wavelet with a width comparable to that of the PSF. This width does not vary much from source to source, it only depends on the energy of the incoming photons. A source emitting proportionally more at high energies than low energies is said to be “hard,” or to have a low spectral index and has a rather peaked PSF: it is best detected at small scale. On the contrary, a “soft” source has a rather flat PSF and is best detected at a larger scale. This dependence of the optimal scale parameter on the spectral index is illustrated on Figure 5.6. Number of σ

190

7

6

5

4

3

2 0

2

4

6

8 10 Scale parameter a

Fig. 5.6. Significance for the detection of a source, expressed in number of sigmas, as a function of

the scale parameter of the wavelet. The wavelet is centered at the position of the source. Each curve refers to a different spectral index, from lower (peaked curve) to higher (flat curve). The position of the maxima of these curves changes as the spectral index changes.

5.1 Astronomy and astrophysics

Most of the 270 EGRET sources are not identified. The intrinsic resolution of high energy gamma-ray detectors is not good enough to provide strong constraints on the source position; it is therefore difficult to find counterparts at other wavelengths. Their nature is still a mystery, that the next generation telescope GLAST will help to solve. A large fraction of identified sources consists in active galaxies whose nucleus is a massive black hole (up to 109 MSun ) surrounded by an accretion disk of matter falling in the gravitational well. In addition, strong jets of ultra-relativistic matter and radiation are emitted perpendicularly to the disk. Active galactic nuclei (AGN) detected in gammarays above 100 MeV have a jet pointed towards the Earth. Their emission is very variable, so that they are often undetected when they are in a quiescent state and then, in a short time, they become very bright. To give an example, EGRET viewing period 21.0 is a high latitude observation in which an AGN is in flaring state. 3EG J0237+1635 was detected at 10 σ in [213] and is 16 σ here. Several other sources are also present above 4 σ . The procedure described above has been applied and results are shown in Figure 5.7. One can see that all but one of the sources are seen. One should also note that the bright AGN position is slightly wrong, because of the presence of a faint source in the vicinity that is not detected. To detect this kind of source, the algorithm must be applied a second time after the addition of the significant sources to the background. This attempt at developing an alternative method to the usual maximum likelihood estimation will probably prove to be relevant in years to come. Indeed, the next generation gamma-ray telescope, GLAST, is to be launched in March 2006 and its complexity Declination

191

30

25

20

15

10

5

0

−5 70

60

50

40

30

20

10

Right Ascension

Fig. 5.7. Detected sources using the wavelet analysis of the EGRET viewing period 21.0 for E>100 MeV, the contours give the significance level from 4 to 16 σ . Superimposed stars give the positions of the third EGRET catalog sources detected over 4 σ (from [213]).

192

Applications of the 2-D CWT. II: physical applications

(parameters to take into account, volume of the data stream) will make it impossible to build a source catalog following EGRET’s old-school procedure. The algorithms must be made more efficient in some way. Wavelets will not be the key to the whole problem, of course, but will hopefully help develop alternative viewpoints.

5.2

Geophysics

5.2.1

Geology: fault detection As we have seen in Section 1.7, wavelet analysis was born in geophysics, as the empirical method designed by J. Morlet for analyzing the recordings of microseisms used in oil prospection. Thus it was to be expected that wavelets would soon find applications in other geophysical problems. It was indeed the case, as can be seen from the reviews [Fou94] or [250], where mostly 1-D applications are discussed, however. Then an interesting application of 2-D directional wavelets to geology was initiated in 1995 by Ouillon [Oui95,298]. The object to be analyzed is a system of geological faults covering a large area in the Arabian peninsula, which shows a self-similar behavior over scales from a few meters to hundreds of kilometers. Standard methods for analyzing such a system are based on renormalization group techniques or on the multifractal formalism (see Section 5.4). What the authors propose here is a continuous wavelet analysis, with directional wavelets, combined with a multifractal analysis. The motivation for this choice is that the relevant information to be measured is the anisotropy of the fault field, and the variation of this anisotropy with scale. In order to understand the idea, let us consider a synthetic so-called en e´ chelons fracture [298], depicted in Figure 5.8(a). At a small scale, the dominant orientation of this object is vertical, but at large scale, one sees only an oblique line pointing NE at 45◦ . Analyzing it with an isotropic wavelet would reveal these details, without focusing on directions, whereas an anisotropic one will enhance the direction response. Thus the authors of [298] chose the latter, namely an anisotropic Mexican hat (see Section 3.3.1). The originality of the method is an optimized local filtering of the WT, as follows. The wavelet used is an anisotropic Mexican hat with anisotropy factor $. One computes the WT of the image for a number of couples (θ, $). Then, for each point in the signal, one selects the pair (θ, $) that gives the largest value of the CWT among all those computed, the Optimum Anisotropic Wavelet Coefficient (OAWC). Thus, for each point, the OAWC selects the local filter which best matches the signal at the chosen resolution (scale). Then one thresholds the OAWC map in order to keep only the most significant features. The ridges of the remaining map correspond to the dominant structures detected. In the case of a fault array, these ridges (called virtual rupture lines or VLR) correspond to the faults as seen on a map. Finally, one draws a histogram of

193

5.2 Geophysics

Fig. 5.8. NOAWC analysis of the synthetic en e´ chelons fracture. Top row: (a) the signal; (b) the

NOAWC map at the resolution of 2 pixels; (c) the same at the resolution of 4 pixels. The bottom row shows the corresponding orientation roses (from [Gai00]).

the azimuths θ of the optimal wavelets associated with the points of the VLRs. This histogram, called a rose by geologists, depicts clearly the anisotropy of the object and its variation with scale. In further papers [120,181], this OAWC method was further improved by adding an adaptive normalization, in the sense that each OAWC is divided by its theoretical maximum corresponding to a perfect match between the wavelet and the object. The so-called NOAWC so obtained is a local indicator of the quality of the match. To give an example, we present in Figure 5.8 the NOAWC analysis of the en e´ chelons fracture signal (from [Gai00]). Panels (b) and (c) show the NOAWC map at the resolution of 2 and 4 pixels, respectively, with the VLRs enhanced in white, and on the bottom row, the corresponding orientation roses. Whereas the 2 pixels rose points at 90◦ (vertical orientation), that at 4 pixels resolution points at 60◦ . The interesting information is then the critical scale corresponding to a brutal shift in the orientation of the rose (see [Gai00] and [298,299] for more details). The NOAWC method has been applied successfully to the analysis of geological fault arrays and to the so-called rock fabric analysis, where “fabric” means the “complete spatial and geometrical configuration of all those components that make up a deformed rock” [Gai00,182]. As an example, we present in Figure 5.9 the NOAWC multiscale analysis of a 150-km-wide fault field, taken from [Gai00]. As in the previous

194

Applications of the 2-D CWT. II: physical applications

Fig. 5.9. NOAWC analysis of a real map of a fault field. (a) The original map; (b)–(f) NOAWC maps at scale a = 2, 4, 8, 16, 32. On the right, we show the corresponding orientation roses (from [Gai00]).

case, we show in the successive panels the NOAWC maps at smaller and smaller resolutions, together with the corresponding orientation roses. The dominant direction of the latter clearly varies with scale. The critical scales where transitions take place are then determined by a multifractal analysis. We refer to [299] and [Gai00] for further details. As already mentioned, the authors of all these papers use for the NOAWC method an anisotropic Mexican hat, which has a rather poor directional selectivity. However, the elliptical shape of the “footprint” of the wavelet plays an essential role in the method. This suggests the use of a Morlet wavelet instead of a Mexican hat. We should expect a much better precision, but the experiment has yet to be done.

5.2.2

Seismology As it is well-known [Bur98], the wavelet saga started with Jean Morlet, a French geophysicist working in oil prospection for Elf-Aquitaine. The technique consists of the sending of an impulse into the ground (by an explosion or any other means) and analyzing the signal reflected by the various discontinuities in the underground, down

195

5.2 Geophysics

x

y

x

t

t (a)

(b)

Fig. 5.10. (a) A seismic section; (b) a seismic block (from [Bou97]).

to 8000 m. These correspond to abrupt changes in density and composition of the rock, a necessary (but by far not sufficient!) condition for the presence of oil or gas. Clearly the resulting signal will be extremely noisy. Then, on a purely empirical basis, Morlet had the idea of representing this signal by a linear superposition of contributions obtained by dilating/contracting a fixed mother function (the analyzing wavelet). The method worked reasonably well, but it took a year of work between J. Morlet and the theoretical physicist A. Grossmann to understand its exact mathematical structure – namely, the content of Chapter 1 [199,205]. From there the theory of wavelets expanded in all directions, as we have seen in the preceding pages, until the loop was completed in the Ph.D. thesis of E. Bournay Bouchereau [Bou97], where she looked again at the very problem of seismic exploration treated by Morlet. The raw data are the so-called seismic sections, that is, 2-D plots where the vertical t axis represents twice the time needed by the wave to reach the corresponding rock layer (thus depth) and the x axis the horizontal distance between two successive receptors. A typical seismic section is shown in Figure 5.10(a). For a comprehensive study, one groups together a collection of parallel sections, thus getting a 3-D seismic block, as shown in Figure 5.10(b). Now, given a section, or a portion thereof, the goal is to detect the geological faults it contains. The technique developed in [Bou97] consists in taking the 2-D continuous WT of the section with a directional wavelet, either a Morlet wavelet or a separable directional one, on the model of (3.24). (The CWT is used here, instead of a discrete WT, in order to maintain translation covariance, which is essential for pattern identification.) This CWT easily detects faults,

196

Applications of the 2-D CWT. II: physical applications

(a)

(b)

Fig. 5.11. Detection of faults in a seismic section: (a) a section; (b) faults detected with a Morlet wavelet analysis (from [Bou97]).

which can be considered as lines of singularities. In the presence of a high noise (deep layers), a preliminary adaptive filtering improves the efficiency of the detection. An example of fault detection is given in Figure 5.11. In addition to the work just described, we ought to mention other applications of directional wavelets in seismology. For instance, the NOAWC method of the previous section has been used for the study of natural seismological events (spatial distribution of hypocenters of an earthquake sequence) [Gai00]. Even 3-D wavelets (see Section 9.1), namely a 3-D Mexican hat, have been used for the description of seismicity of a large area in the western Alps [75]; and, needless to say, 1-D wavelet analysis has also been applied to seismic time series, in particular, the arrival time of the various components, S-wave, P-wave, etc. (see, for instance, [104,297,Oon00]). Finally, it is amusing to note that essentially the same technique based on 2-D directional wavelets has been used for certain problems in metallurgy [236], owing to the similarity between a metallurgical image (inner structure of a piece of metal) and a seismic image.

5.2.3

Climatology Before concluding this section, we have to say a few words on the use of wavelet analysis in climatology. Most of the applications are in 1-D, typically time series analysis, using

197

5.3 Applications in fluid dynamics

often the discrete WT. We refer to [Fou94] or [250] for a review. An exception is the work of Hudgins et al. [227] on atmospheric turbulence, which uses the CWT in an essential way (this is the paper in which they introduced the wavelet cross-correlation function or wavelet cross-spectrum). Another one in the same spirit is that of van Milligen [273] on turbulence in fusion plamas. In both cases, the wavelet method turns out to be superior to the standard Fourier techniques. As for 2-D examples, the prime domain is again that of turbulence in fluids, that we will discuss in Section 5.3.1. In addition to the latter, an interesting application was made by Kumar [249], namely, to determine the so-called scale space anisotropy of geophysical fields. By this, one means that such fields are not only highly anisotropic over a wide range of scales, for dynamical reasons, but, in addition, different scale features are oriented in different directions (exactly as for the geological fault arrays described above). A typical example is provided by hurricanes, where the scale anisotropy is obvious. In many other cases, however, the anisotropy is present in a subtle way, that cannot be properly detected by classical techniques, such as spectrum- or correlation-based techniques. As an alternative, Kumar uses a 2-D Morlet wavelet analysis to characterize scale space anisotropy in radar-depicted spatial rainfall, by studying the fraction of energy in different directions at different scales. For that purpose, he introduces the relative scale-angle spectrum (2.58), which is sufficiently sensitive to reveal the subtle presence of scale space anisotropy in random fields. In the particular example treated here, the author is able to conclude that “ . . . a rainfall field might show an anisotropic structure that might not be obvious from a typical spectral analysis and may have wider implications in modeling and sampling problems.” Another application, closely related to the previous one, is the use of 2-D wavelets for enhancing thin-line features in meteorological radar reflectivity images [212]. Thin-line features in reflectivity correspond to surface wind convergence lines that can potentially lead to the initiation of thunderstorms. Thus the detection, preferably automatic, of such features is an important ingredient in the short time forecasting of thunderstorms. It turns out that a directional wavelet is required, namely a 2-D Morlet wavelet or a separable substitute built on the model (3.24). Once again, we see the superior discrimination power of directional wavelets in physical applications! In addition to the directional aspects analyzed in the previous applications, it is a fact that many meteorological phenomena have a distinctly fractal behavior. Clouds are a good example, but several artificial examples, such as random surfaces, share the same property. Thus, it is not surprising to find several applications of wavelets to such fractal structures. We will discuss some of them in Section 5.4.1.

5.3

Applications in fluid dynamics The wavelet transform, both continuous and discrete, has been successfully applied to the analysis of 2-D developed turbulence in fluids, especially for localization of coherent

198

Applications of the 2-D CWT. II: physical applications

structures in the distribution of energy or enstrophy. This topic is briefly described in Section 5.3.1. In addition, we will describe here two other applications of 2-D wavelets in fluid dynamics, which rely on the possibility of local filtering, both in direction and in position, with directional wavelets.

5.3.1

Detecting coherent structures in turbulent fluids Turbulence in fluids is a phenomenon that has resisted analysis until now. After more than a century of research, no real theoretical understanding of the dynamics of a turbulent flow has been achieved. There only exist various statistical or phenomenological models, which are widely used in practical applications, but lack a genuine justification. Even the very definition of the terms used does not always achieve a consensus among physicists. On the other hand, there is a huge amount of experimental data. In order to understand the rˆole of wavelets in this context, we have to go back to the basics. For a general review, we refer to [165]. The starting point is the system of Navier–Stokes (NS) equations governing the evolution of an incompressible Newtonian fluid: ∂ v v + 1∇ p = ν + ( v · ∇) v + F ∂t ρ · v = 0, ∇

(5.8) (5.9)

supplemented by adequate initial and boundary conditions. In these equations, v ≡ v( x , t) is the velocity, p ≡ p( x , t) the pressure, F the external force per unit mass, ρ a constant density, and ν the constant viscosity. The NS equations (5.8)–(5.9) are × v, which measures the local often expressed in terms of vorticity, namely ω =∇ rotation rate of the fluid. In dimension 2, the NS equations are formally the same, but the velocity field reads v( x ) = (u(x, y), v(x, y)) and the vorticity reduces to the pseudoscalar ω = ∂x v − ∂ y u. Then the fundamental quantities are the total energy and the total enstrophy, defined as, respectively 1 1 2 2 E(t) = d x | v ( x , t)| , Z (t) = d 2 x |ω( x , t)|2 , 2 2 and their Fourier transforms, the energy and enstrophy spectra ( is the volume occupied by the fluid). Fully developed turbulence is the regime of very large Reynolds numbers Re ∼ 1/ν, ν → 0 (in practice, in aeronautics, meteorology or combustion, for instance, Re varies v bebetween 106 and 1012 ). In this regime, the nonlinear advection term ( v · ∇) comes dominant, by several orders of magnitude. As with the semiclassical limit → 0 in quantum mechanics, this changes the character of the equation. As a result, not only is there no analytical solution known, but the NS equations cannot be solved numerically in this regime with present day computers, unless some drastic simplifications are made. Instead, since turbulent flows are highly unpredictable, one

199

5.3 Applications in fluid dynamics

has to use statistical models, requiring some basic assumptions, such as statistical homogeneity and isotropy, or ergodicity, which allows one to replace ensemble averages by space averages. All this led, for example, to the celebrated 1941 cascade model of Kolmogorov [165]. Yet turbulent fluids often exhibit coherent structures, that is, structures in the energy or the enstrophy spectrum that persist through a large range of scales (vorticity tubes, often called filaments), but are highly unstable. Clearly, the mere existence of these invalidates the statistical assumptions. In addition, the averaging processes, while satisfactory at low Reynolds numbers, ignore the coherent structures, since they have a small extent in space and in time. These are, however, an essential aspect of fully developed turbulence, and thus statistical models are inadequate for Re ! 1. This situation led Marie Farge to introduce, back in 1988 [163], wavelet methods for detecting and analyzing the time evolution of such coherent structures. Since then, she and her collaborators have devoted a huge amount of research work in this direction. Many different techniques have been used, wavelets (CWT and DWT), wavelet packets, multifractal techniques. In retrospect, the basic idea is always to separate the coherent structures, which are analyzed with wavelets, from the background flow, treated by statistical methods. This is, of course, not the place to go into the details of this considerable, but rather specialized work, which represents one of the most spectacular applications of wavelet analysis in physics. We refer the interested reader to the extensive review papers by M. Farge et al., which contain references to the original work [164,165,335]. In addition, a 1-D application to intermittent turbulence in atmospheric data is given in [211].

5.3.2

Directional filtering We will turn instead to applications that use specifically directional wavelets. As a consequence of their good directional selectivity, the Morlet and Cauchy wavelets are quite efficient for directional filtering. In order to illustrate the point, we analyze in Figure 5.12 a pattern made of rods in many different directions (a). Applying the CWT, with a Cauchy wavelet in a fixed direction (here horizontal), selects all those rods with roughly the same direction (b), whereas the other ones, which are misaligned, yield only a faint signal corresponding to their tips, in agreement with the behavior discussed above. Since this is in fact noise, one performs a thresholding to remove it, thus getting a clear picture (c). In this way, one can count the number of objects that lie in any particular direction. Note that the same pattern was analyzed with a Morlet wavelet in [18,19], and the result is slightly less neat. Figure 5.13 presents another example of directional filtering, this one with the Gaussian conical wavelet (3.37). The picture represents bacteria, seemingly at random. However, after filtering successively at −10◦ , 45◦ , and 135◦ , one realizes this latter orientation is significantly more populated.

200

Applications of the 2-D CWT. II: physical applications

(a)

(b)

(c) ◦

Fig. 5.12. Directional filtering with a Cauchy wavelet (A R P = 20 ) oriented at θ = 0◦ : (a) the

pattern; (b) the CWT; (c) the same after thresholding at 25%.

5.3.3

Measuring a velocity field In the first example [Wis93,375], the aim is to measure the velocity field of a 2-D turbulent flow around an obstacle. Velocity vectors are materialized by small segments, by the technique of discontinuous tracers. Tiny plastic balls are seeded into the flow and illuminated by a “plane of light,” in order to get a 2-D image. Then two successive photos are taken with a fast CCD camera, with exposure times of 700 and 6000 µs, respectively. In this way one gets a “dot-bar” signature for each tracer, which materializes

201

5.3 Applications in fluid dynamics

Fig. 5.13. Another example of directional filtering, with a Gaussian conical wavelet: (a) the original image, representing bacteria; (b) filtering at −10◦ ; (c) the same at 45◦ ; (d) the same at 135◦ .

the direction and the length of the local velocity (see an example in Figure 5.14, taken from [Wis93]). In order to get sufficiently many data points, one superposes several such pictures, typically 16. First one computes the WT of the resulting image with a Morlet wavelet, which selects those vectors that are closely aligned with the wavelet. Then a second analysis is performed with a wavelet oriented in the orthogonal direction, thus completely misoriented with respect to the selected vectors. Now the WT sees only the tips of the vectors and their length may be easily measured. The same two operations are then repeated with various successive orientations of the wavelet. Using appropriate thresholdings, the complete velocity field may thus be obtained, in a

202

Applications of the 2-D CWT. II: physical applications

Fig. 5.14. The dot-bar signature of tracers in the fluid flowing from left to right (from [Wis93]).

totally automated fashion, with an efficiency sensibly better than with more traditional methods. Two examples of reconstructed velocity fields from [Wis93] are given in Figure 5.15, corresponding to a quasi-laminar flow and a turbulent flow around an obstacle (again the flow comes from the left; units are normalized to the size of the experimental area). Notice that the analysis gives in principle both the modulus and the phase of the WT. But here, contrary to the simple applications like contour detection [13], the phase cannot be exploited, the data are too noisy. Thus one loses some precision on the orientation. Nevertheless, the method is remarkably efficient. In the same vein, the 2-D CWT (again Cartesian only) has been proposed for improving the method of holographic particle velocimetry, which consists in measuring the velocity of particles in a fluid by exploiting holograms of fluid volumes [10].

5.3.4

Disentangling of a wave train A second example originates from underwater acoustics. When a point source emits a sound wave above the surface of water, the wave hitting the surface splits into several components of very different characteristics (called respectively “direct,” “lateral,” and “transient”). The resulting wave train is represented by a linear superposition of damped plane waves, and the goal is to measure the parameters of all components. This phenomenon has been analyzed successfully with the WT both in 1-D [334] and in 2-D [18], and the extension to a 3-D version is straightforward. Let us give some details of the method in the 2-D case. The signal representing the underwater wave train is taken as a linear superposition of damped plane waves: f ( x) =

N

cn ei kn ·x e−ln ·x ,

(5.10)

n=1

where, for each component, kn is the wave vector, ln is the damping vector, and cn a complex amplitude. Then, using successively the scale-angle and the position

203

5.3 Applications in fluid dynamics y 0,5 0,4 0,3 0,2 0,1 0,0 −0,1 −0,2 −0,3 −0,4 −0,5 −9,68

−8,68

−7,68

−6,68

−5,68

−4,68

x

(a) y 2,0 1,5 1,0 0,5 0,0 −0,5 −1,0 −1,5 −2,0 −2,0

x −1,5

−1,0

−0,5

0,0

0,5

1,0

1,5

2,0

2,5

3,0

3,5

(b) Fig. 5.15. Two examples of reconstructed velocity fields: (a) a quasi-laminar flow; (b) a turbulent flow around an obstacle (from [Wis93]).

representations described in Section 2.2.3, one is able to measure all the 6N parameters of this signal with remarkable ease and precision. The method proceeds in three steps and uses explicitly the phase space interpretation. First one computes the CWT of the signal (5.10) with a Morlet wavelet. By linearity, the result is the linear superposition of the contributions of the various components. Moreover, each component is the product of two factors, where the first one depends on b only and the second one on (a, θ) only: a, θ ) = F(b,

N n=1

ˇ cb,n Fn (a, θ).

(5.11)

204

Applications of the 2-D CWT. II: physical applications

Actually, the resulting function may be written explicitly in terms of the phase space variables introduced in Section 2.3.2, for instance, Fˇn (a, θ ) ≡ Fˇn ( v ). Now we go to the scale-angle representation and consider the WT (5.11) for fixed Then a straightforward calculation shows that, for each term in this superposition, b. a −1 Fˇn (a, θ ) admits a unique local maximum. Now, in the full transform (5.11), each term has its own local maximum, but these need not be well separated: one maximum may hide another one, totally or partially. This masking effect will happen, for instance, when: r one component has a much bigger amplitude, |c | ! |c |, for all m not equal to n b,n b,m (total masking); r two wave vectors are close to each other, k k , but with different amplitudes, n m |cb,n | > |cb,m | (partial masking). In both cases, the two waves can be separated, by increasing the selectivity of the wavelet (for instance, using a Morlet wavelet with a more anisotropic modulus). If the two waves have close wave vectors (kn km ) with similar amplitudes (|cb,n | |cb,m |), but different damping vectors (kn = km ), then they can still be separated, by changing the Otherwise the method will fail, the two waves interfere inextricably, observation point b. none of them dominates the other one. When the masking effect is not too important, the maxima will be sufficiently prominent that the interferences between the different components will become negligible (in the modulus) and one may write: a, θ )| |F(b,

N

ˇ |cb,n | | Fn (a, θ)|,

(5.12)

n=1

One then reverts to the position representation, choosing for (a, θ ) each maximum successively. In each case, the filtering effect of the CWT essentially eliminates all components except one, which is then easy to treat. In this way, one is able to measure all the 6N parameters of the signal easily. A striking example is given in [18], illustrating the power of the method as well as the rˆole of the anisotropy factor $ of the Morlet wavelet. The signal is the superposition of four damped plane waves, with different wave vectors kn , except that the directions of k1 and k4 differ by 20◦ only. As a result, wave # 1 partially masks wave # 4: when the analysis is performed with a Morlet wavelet with $ = 1, the corresponding maxima in the scale-angle representation are not well separated [Figure 5.16(a)]. When one uses instead a wavelet with $ = 5, the two maxima are clearly identifiable and and can be localized precisely (b). Notice that the “footprint” of the wavelet is not an ellipse, because the radial coordinate used is a −1 , not a. Now the procedure allows to reconstruct each of the four components almost perfectly [18]. Only wave # 4 keeps some trace of interference with wave # 1, the others are indeed pure waves. In order to remove this effect, one should first subtract wave # 1 from the signal and redo the analysis.

205

5.4 Fractals and the thermodynamical formalism

2 3

2

3

1

1

4

4

(a)

(b)

Fig. 5.16. Disentangling of a four component wave train with a Morlet wavelet: the four maxima (a) with $ = 1; (b) with $ = 5 (from [18]).

5.4

Fractals and the thermodynamical formalism

5.4.1

Analysis of 2-D fractals: the WTMM method Many physical phenomena require a wide range of scales for a complete description of their properties. The paradigm, of course, are fractals, which are complex mathematical objects that have no natural length scale. More precisely, a fractal, be it in 1-D or in 2-D, is by definition self-similar under dilation, either globally (genuine fractal) or locally (multifractal). Physical examples abound. For instance, all kinds of random walks used to mimic various noisy dynamical behaviors, financial time series, geologic shapes (such as the fault systems descibed in Section 5.2.1), interfaces that develop in growth processes far from equilibrium, fractal growth processes (such as the so-called diffusion-limited aggregates), electrodeposition clusters, etc. A fractal is in general a very irregular object (for instance, its support may be a Cantor-like set), hence it should be represented by a singular measure, rather than a function. In order to cope with such situations, the so-called multifractal formalism has been developed. Now, central concepts of the theory, such as generalized fractal dimensions or spectrum of singularities of the measure, are closely related to ideas from statistical mechanics (as a matter of fact, the standard “box counting” method is already of a statistical nature). Thus one speaks also of a thermodynamical formalism of fractal analysis. A review of this formalism may be found in [47] or [293]. Since scaling is the most significant operation in the context of fractals, the CWT is a natural tool for analyzing them. Clearly the continuous version of the WT is essential here, since the characteristic scaling ratio is unknown a priori. The first step is to extend

206

Applications of the 2-D CWT. II: physical applications

the CWT to singular measures. This was achieved in 1-D by Holschneider [221] and Arn´eodo et al. [45], then extended to 2-D by Arn´eodo and his group in Bordeaux, including one of us (R.M.) [43]. We briefly describe the key steps of the formalism. We take the most general case, namely, an object described by a fractal measure µ on R2 . The standard thermodynamical formalism (spectrum of generalized fractal dimensions) yields only statistical information about the object as a whole. To get precise local information requires a wavelet transform. The CWT of the measure µ with respect to the wavelet ψ is defined as T [µ](b, a, θ) = dµ( x ) ψ(a −1r−θ ( x − b)). (5.13) In the isotropic case, we write simply x − b)). T [µ](b, a) = dµ( x ) ψ(a −1 (

(5.14)

Assume now that the measure has the following scaling behavior (self-similarity) around the point xo : x , $)), µ(B( x , λ$)) ∼ λα(xo ) µ(B(

λ > 0,

(5.15)

xo ) is the local scaling exponent. where B( x , $) is a ball of radius $ around xo and α( Using the covariance property of the CWT (Proposition 2.2.3), it is easily shown that the WT of the measure µ scales in the same way: λa, θ ) ∼ λα(xo ) T [µ]( a, θ), T [µ]( xo + λb, xo + b,

λ → 0+ .

(5.16)

This relation is the key to the wavelet analysis of fractals. For instance, the local exponent versus log a, for a small enough. α( xo ) may be obtained by plotting log |T [µ](a, θ, b)| This would suffice for an exact (global) fractal, such as a numerical snowflake, for which α is constant over the whole object. For a genuine multifractal, α( xo ) varies from point to point, and then (5.16) allows one to compute the generalized fractal dimensions and the singularity spectrum of the object. This wavelet approach to the thermodynamical formalism of fractal analysis has been developed systematically by Arn´eodo and his collaborators in Bordeaux. It is instructive to give some hints about the basic ideas of the formalism. We first go back to the standard multifractal formalism. Given a singular measure µ, it can be described in terms of its q-moments as follows. Suppose we cover the support of µ with N ($) boxes Bi ($) of size $. Then the scaling behavior of µ will be deduced from the partition function Z(q, $) =

N ($) i=1

q

µi ($),

where µi ≡ µ(Bi ($)), q ∈ R.

(5.17)

207

5.4 Fractals and the thermodynamical formalism

In the limit $ → 0+ , Z(q, $) behaves as a power law: Z(q, $) ∼ $ τ (q) , $ → 0+ . Finally, the spectrum of generalized fractal dimensions is obtained from the exponents τ (q) by the relation Dq = τ (q)/(q − 1), and these in turn completely characterize the singular behavior of the original measure µ, in particular, the local H¨older exponents that describe it (see [47,292]). Now the rationale for reinterpreting this formalism in terms of wavelets is that the WT tends to “forget” the regular part of the signal (because of the vanishing moments) and focus on the singular part. A naive way of achieving this would be to take as partition function, instead of (5.17), Z(q, a) = d x |T [µ]( x , a)|q , q ∈ R (5.18) (in 1-D, one uses simply |T (x, a)| as WT modulus, see (2.66)). This is a bad choice, however, since Z(q, a) might diverge for q < 0. Instead one replaces the integral over x by a discrete sum over the local maxima of T [µ]( x , a), for fixed a, that is, precisely the WTMM introduced in Section 2.3.5. The justification of this choice is that [262] (i) the maxima lines (ridges) have the same scaling behavior as the WT itself, and (ii) each maxima line l = (bl (a), a) points, as a → 0, to a point bl (0) which corresponds to a singularity of µ and, in addition, the WT modulus scales along the line as |T [µ](bl (a), a)| ∼ a α(bl (0)) .

Thus the wavelet plays the role of a generalized “oscillating box” and the scale a defines its size. Actually, the definition of the partition function can be further refined by using explicitly the WTMMM, as follows. Let L(a) denote the sets of ridges that exist at scale a and contain a maximum at a scale a a. Then one defines finally the partition function q

Z(q, a) = |T [µ]( x , a )| . (5.19) sup l∈L(a)

( x ,a )∈l,a a

In fact, introducing the “sup” amounts to adapting the size of the wavelet along the ridge so as to avoid divergences. Here again, the exponents τ (q) are defined from the power-law behavior of Z(q, a) as a → 0: Z(q, a) ∼ a τ (q) , a → 0+ . (In the analogy with thermodynamics, q and τ (q) play the rˆole of inverse temperature and free energy, respectively.) The WTMM technique has been applied successfully to a wide variety of examples [43,44,47], that cover both artificial fractals (numerical snowflakes, diffusion limited

208

Applications of the 2-D CWT. II: physical applications

aggregates, recursive fractal functions) and natural ones (electrodeposition clusters, various arborescent phenomena, fully developed turbulence data, clouds). The method permits the measurement of the fractal dimensions and the unraveling of universal laws (mean angle between branches, azimuthal Cantor structures, etc.). In 2-D, it should be remarked that the analysis uses exclusively an isotropic wavelet (usually a 2-D Mexican hat), and thus there is no θ dependence in (5.16). However, this may not be the end of the story. Indeed we shall exhibit in Section 4.5.2 below a fractal (“twisted snowflake”) whose structure requires a directional wavelet for its complete determination. In more recent work, the attention has focused on 2-D applications, around the theme of rough surfaces: fractional Brownian surfaces, anisotropic self-affine rough surfaces [Dec00,53], synthetic multifractal rough surfaces [129], cloud structure [52,54]. This last item opens a whole world of applications to physical processes, in particular to meteorology, since fractal objects abound there. Another application, closely related to the previous analysis, concerns the analysis of real rough surfaces, that is, metal surfaces obtained after various kinds of machining processes. Here too, the 2-D wavelet transform yields a useful tool [235]. Actually the object to analyze is essentially the texture of the surface, which brings us to another field of application, namely, texture analysis, that we will discuss in Section 5.5. A last point to notice in this context is that, considering the heavy computational cost of the 2-D CWT, the Bordeaux group has designed an ingenious hardware version, called the Optical WT [Arn95,46]. The technique, based on Fraunhofer diffraction, a familiar tool in optics, amounts to obtaining the WT with a binary approximation to the isotropic Mexican hat, that is, using as isotropic wavelet the Bessel filter described in (3.9)– (3.10). With this tool, they obtained beautiful pictures of CWT analyses of diffusionlimited aggregates [Arn95,43,46]. The technique could not, however, be pushed to full implementation for lack of a sufficiently fast CCD camera – or, equivalently, a sufficiently large budget! Actually, several optical implementations of the WT have been proposed in the literature. However, they are often limited to 1-D data or, in the 2-D case, to Cartesian coordinates (translations and separate scaling factors in the x and y directions). We refer the interested reader, for instance, to a feature issue of Applied Optics dedicated to this topic [255]. An alternative optical approach, based on a special type of grating and able to reproduce many types of wavelets, has been presented in [269]. The isotropic Mexican hat has also found applications in optics proper, for instance, in the determination of the wave aberration coefficients of a rotationally invariant optical system, from the measured data of wave front deformations [340]. The result is that the wavelet method is more efficient and more robust to noise than standard least squares methods. Another application of the 2-D CWT in optics is to moir´e interferometry, a well-established optical technique for measuring displacement and strain in materials, based on phase analysis of interference fringe patterns [238]. Thus a complex wavelet is needed, and the authors resort to a modified Morlet wavelet, obtained by replacing

209

5.4 Fractals and the thermodynamical formalism

the Gaussian by a cubic spline, as advocated by Unser [357], and once again, the CWT method proves superior to the standard Fourier techniques.

5.4.2

Shape recognition and classification of patterns The characterization of a 2-D shape from its outlines is an important problem in several applications of image analysis, such as character recognition, machine parts inspection for industrial applications, characterization of biological shapes such as chromosomes and neural cells, and so on. Furthermore, in the field of human vision and perception, 2-D shape analysis also plays a central role in psychophysics and neurophysiology. There are two general approaches to shape characterization: region based, which deals with the region in the image corresponding to the analyzed object; and boundary based, where the shape is characterized in terms of its silhouette [Pav77]. The former is intrinsically 2-D, dealing directly with planar primitives and concepts, and thus 2-D wavelets may be used directly. The latter, however, mimics 2-D operations through 1-D representations, and is referred to henceforth as contour characterization. An alternative to standard techniques consists of representing the shape by the complex curve that describes its boundary, and applying the 1-D CWT to this complex signal [Ces97,21], as outlined in Figure 5.17. This leads to the so-called W-representation,

SEGMENTATION

CONTOUR EXTRACTION

300 t

250 x(t)

200 CONTOUR TRACKING

150 100 50 0 0

y(t)

200

400

600

800

1000 t

Fig. 5.17. Basic framework for 2-D shape characterization from its outline (from [21]).

210

Applications of the 2-D CWT. II: physical applications

which allows an easy way of performing a number of standard tasks (for instance, in machine vision), such as detection of dominant points, shape partitioning, natural scales analysis. As compared to other techniques, the W-representation has the following useful characteristics, which are desirable for purposes of shape analysis [275] and follow directly from the basic properties of the CWT: r uniqueness, because of the invertibility of the CWT; r invariance under translation, scaling and rotation; r robustness to local modifications of the shape; r efficiency and ease of implementation. Notice that an essential ingredient of the technique is the wavelet-based fractal analysis discussed above. In particular, one resorts systematically to the different types of local maxima lines defined in Chapter 2, Section 2.3.5. More precisely: r The algorithm for the detection of dominant points, e.g. corners, is based on the vertical maxima lines, namely, their position and some relevance measure, for instance, their length. r The detection of periodic patterns and the so-called natural scales is based on the horizontal maxima lines, since this amounts essentially to determining the instantaneous frequencies in the signal. r Finally, the fractal behavior of certain contours is analyzed with the standard technique sketched in Section 5.4.1. This analysis has numerous applications. An original one is the classification of neurons according to their complexity, that is, their fractal dimension [Ces97]. Of course, shape analysis is a whole different world. For an up-to-date review of it, we refer to the recent monograph of L. da F. Costa and R. M. Cesar, Jr. [Cos01].

5.5

Texture analysis The determination and classification of textures in images is an old and difficult problem, with many potential applications, notably in computer vision. Numerous methods have been designed to that effect. Most of them are of a statistical nature, such as Markov random fields, but one has also used Gabor analysis [234] and various kinds of wavelet transforms. We shall concentrate on the latter, of course. Some proposals have been made with the standard (Cartesian) discrete WT (see [166] for instance), but, since most textures possess directional features, it is more natural to use oriented wavelets for attacking the problem. An example is the steerable wavelet pyramid developed by Simoncelli and his collaborators (see Section 2.7). This technique was used in [317] as a basis for a parametric texture model based on joint statistics of wavelet coefficients. It has been considerably developed by Do and Vetterli [Do01,140,141], using appropriate generalizations of the ridgelets and curvelets (these

211

5.5 Texture analysis

will be described in Chapter 9 and Section 11.1), and combining them with waveletdomain hidden Markov models. Alternatively, one can use the plain 2-D CWT with an oriented wavelet. Actually a similar solution was proposed long ago by Rao and Schunck [323], using the first derivative of a 2-D Gaussian, (3.15), already considered by Canny [98] (see Section 3.3.1.1). This is indeed an oriented wavelet, but it is not directional in the technical sense. (In addition, one may find in [323] a survey of the early literature on texture determination.) Texture analysis with a genuine directional wavelet (a Morlet wavelet) was done first by Gonnet [195], using the characterization of the instantaneous frequency of the signal as a vector field [196] (Section 2.3.5). Further results were obtained by Murenzi et al. [290], using the same wavelet (truncated Morlet, also called Gabor filter) and the scale-angle representation. A more efficient technique yet is that of the directional dyadic wavelet packets [360] (Section 2.6.4). Substantial advances in the classification of textures have been obtained recently along this line by Menegaz et al. [270]. As usual in the framework of computer vision, a pattern is represented by a feature vector, as explained in Chapter 4, and the classification is made in defining similar images as those whose feature vectors are close to each other, in the sense of some notion of distance (often Euclidean or L 2 ). In order to use directional wavelet packets for this problem, one simply includes the directional details at each scale among the components of the feature vectors. The latter become longer, but the efficiency of the method for texture discrimination increases significantly. A related topic is the so-called shape from texture problem, which can be formulated as follows. We are given a 2-D photograph of a 3-D surface, which displays a pattern or a texture, more or less regular. The image gives a distorted view of this texture, which depends of the geometry of the surface. The goal is to reconstruct the original surface from the distorted image. The problem is usually split into two steps. First one measures the local distortion of the image, then one recovers the original surface. For the first step, some assumptions are usually made about the type of surface and texture. It turns out that the CWT is an efficient tool for the estimation of distortions, as first proposed by Super and Bovik [348] and Hwang et al. [229]. The planar surface is assumed to have a homogeneous texture, which is modeled as a linear superposition of plane waves. A further paper [231] considers a plane containing several textures with different orientations, a situation which requires first a segmentation step. In both cases, the authors rely on the properties of the ridges of the 2-D wavelet transform (Section 2.3.5). Further work along the same line was made in [179], exploiting the well-known covariance properties of the CWT under translation, rotation, and scaling. Here too, the authors consider a single homogeneous (sinusoidal) texture, and study its deformations. An alternative to this deterministic model is to describe the texture by a stationary random process. This approach has been developed by Clerc and Mallat [Cle99,106,107] for the general case of a curved surface. In addition, they consider general distortions,

212

Applications of the 2-D CWT. II: physical applications

for instance, anisotropic scalings. Since the conventional CWT is no longer covariant under such transformations, they introduce instead the so-called warplet transform, replacing the familiar global scaling by a distortion matrix which is not a multiple of the identity (the name refers to the fact that the image presents a warped view of the surface). This technique, while conceptually interesting, leads to high computational costs, in particular, the formalism of [179] cannot cope with it.

5.6

Applications of the DWT For the sake of completeness, we conclude this chapter with some remarks on the applications of the DWT. As we said already, the latter is used in the majority of applications, but this is not the main subject of the present book. Thus we will give only a few indications. As with other methods, wavelet bases may be applied to all the standard problems of image processing. The main problem of course is data compression, and for achieving useful rates one has to determine which information is really essential and which one may be discarded with acceptable loss of image quality. Significant results have been obtained in the following directions. r Representation of images in terms of wavelet maxima [264], as a substitute for the familiar zero-crossing schemes [Mar82]. r In particular, application of this maxima representation to the detection of edges, and more generally detection of local singularities [262]. r Image compression and coding using vector quantization combined with the WT [41]. r Image compression, combining the previous wavelet maxima method for contours and biorthogonal wavelet bases for texture description [176]. r Image and signal denoising, by clever thresholding methods [144]. Some applications are less conventional. For instance, a technique based on the biorthogonal wavelet bases [108] has been adopted by the FBI for the identification of fingerprints. The advantages over more conventional tools are the ease of pattern identification and the superior compression rates, which allows one to store and transmit a much bigger amount of information in real time. The full story may be found in [87]. Another striking application is the deconvolution of noisy images from the Hubble Space Telescope, by a technique combining the DWT with a statistical analysis of the data [Bou93,85,328]. The results compare favorably in quality with those obtained by conventional methods, but the new method is much faster. One should also quote a large amount of work under development in the field of High Definition Television, where wavelet techniques are being actively exploited; here again the huge compression rates make them specially interesting. As for applications of the multidimensional DWT more specifically oriented to physics, we like to mention two. The first one is in quantum field theory (although

213

5.6 Applications of the DWT

it was done before the wavelet techniques were born): various perturbation expansions (the so-called “cluster expansion”) used in the analysis of Euclidean field theory models are in fact discrete wavelet expansions [65]. Actually the summation over scales, indexed by j, was originally motivated by renormalization group arguments. In the same domain, we may note that wavelet bases have been used also ( [66] and references therein) for estimating the time evolution of solutions of some wave equations (Klein-Gordon, Dirac, Maxwell or the wave equation), or even to expand solutions of the equations in terms of dedicated “wavelets” (although the functions introduced in the last case seem rather far away from genuine wavelets [239]). The other application resorts to solid state physics, namely the Quantum Hall Effect (quantization of the electric conductivity) that occurs when a 2-D electron gas is submitted to a strong transverse magnetic field. Here orthonormal wavelet bases may be used for generating localized orthonormal bases for the lowest Landau level, a necessary step towards the analysis of the Hall effect [14,36,57,58].

6

Matrix geometry of wavelet analysis. I

6.1

Group theory and matrix geometry of wavelets In Chapters 1 and 2, we have studied systematically the continuous wavelet transform in one and two dimensions, respectively. As already emphasized there, the properties of the transforms in the two cases are remarkably similar. In 2-D we have formalized them in the three propositions 2.2.1, 2.2.2 and 2.2.3, and essentially the same statements may be made in 1-D. A moment’s reflection shows that one could write out, without difficulty, an entirely parallel mathematical description in any dimension n 1. Clearly there must be some unifying principle underlying the picture. The question is, of course, what is this principle? As so often in such situations, the answer is to be found in group representation theory, i.e., by looking at the underlying geometry of the space of signals. The various transformations (translation, rotation, zoom, etc.) that a signal may undergo, determine a set of mathematical symmetries, which, interestingly enough, can be expressed in simple matrix terms and, as will be made clear in the following, the signal space itself – as a mathematical object – emerges as a consequence of this geometry. But we have been using group theory all along! Indeed, to draw on a literary analogy, like Moli`ere’s Monsieur Jourdain speaking in prose without knowing so, we have been using group-theoretical language throughout our analysis! It is the aim of the present chapter to demonstrate this fact. By so doing, we intend to achieve three goals. First, group theory will provide us with a unifying picture to view all the mathematical properties of signals that we have derived by hand, so to speak. Second, we hope to convince the reader that the group-theoretical approach is not only aesthetic, it is also simpler, in that it allows us to understand the deeper mathematical structures involved in a simple language. Finally, rewriting our results in this language allows us to easily extend the concept of a CWT to more general manifolds (i.e., more complex spaces of signal parameters), which we shall explore in later chapters: the n-dimensional space Rn , space–time R1+1 or R1+3 , the two-sphere S 2 or the n-sphere S n , etc. Indeed, as recalled in the Prologue, this was the decisive factor in extending the CWT from one to two dimensions in R. Murenzi’s thesis [Mur90], the key step being to identify the

214

215

6.1 Group theory and matrix geometry of wavelets

relevant group and its realization in the space of signals. To set the stage, let us analyze this mathematical structure in some detail. We shall do this first in 1-D, in this chapter, and then take up the 2-D problem in the next chapter. We emphasize that the next three chapters, while interesting and intellectually satisfying, are not prerequisites for the last three, Chapters 9 to 11, which will treat various extensions of the wavelet transform studied so far. In addition, we have collected in the Appendix some pertinent definitions and results from group theory, that we hope would help the group theoretically uninitiated reader to understand the material and to appreciate better the breadth of its scope.

6.1.1

The 1-D CWT revisited We start with the basic 1-D transformation (1.4): x −b −1/2 ψ(x) → ψb,a (x) = |a| ψ , b ∈ R, a = 0, a and rewrite it in the form ψb,a (x) = |a|−1/2 ψ (b, a)−1 x ,

(6.1)

(6.2)

where we have introduced the affine transformation of the line, consisting of a dilation (or scaling) by a = 0 and a (rigid) translation by b ∈ R: x = (b, a)y = ay + b,

(6.3)

and its inverse y = (b, a)−1 x =

x −b . a

(6.4)

Writing φ = ψb,a and making a second transformation on φ we get φ(x) → φb ,a (x) = | a | − 2 φ((b , a )−1 x) 1

= | aa | − 2 ψ((b , a )−1 (b, a)−1 x) x − (b + ab )

− 12 = | aa | ψ . aa 1

(6.5)

This shows us that the effect of two successive transformations is captured in the composition rule (b, a)(b , a ) = (b + ab , aa ), which, if we represent these transformations by 2 × 2 matrices of the type a b (b, a) ≡ , a = 0, b ∈ R, 0 1

(6.6)

(6.7)

216

Matrix geometry of wavelet analysis. I

is reproduced by ordinary matrix multiplication. The point to be noted about these matrices is that the product of two of them is again a matrix of the same type and so also is the inverse, −1 a b a −1 −a −1 b −1 (b, a) = = 0 1 0 1 of such a matrix. Furthermore, the 2 × 2 identity matrix is also in this class. In other words, the set of matrices (6.7) constitute a group, called the (full) affine group and denoted G aff . Note also, that if we consider only those matrices in (6.7) for which a > 0, then this set is also stable under multiplication and inverse taking and hence, it constitutes a subgroup of G aff , denoted G + aff . Coming back to the relation (6.2), we + observe that G aff or G aff consists precisely of the transformations we apply to a signal: translation (time-shift) by an amount b and zooming in or out by the factor a. Hence, the group G aff relates to the geometry of the signals. Next let us study the effect of the transformation given by the group element (b, a) on the signal itself. Writing, ψ → U (b, a)ψ ≡ ψb,a , we may interpret U (b, a) as a linear operator on the space L 2 (R, d x) of finite energy signals, with the explicit action, x −b . (6.8) (U (b, a)ψ)(x) = |a|−1/2 ψ a Additionally, for each (b, a), the operator U (b, a) is unitary, i.e., it preserves the Hilbert space norm of the signal: ∞ d x | ψ(x) | 2 . ψb,a 2 = ψ2 = −∞

More interestingly, the association, (b, a) → U (b, a) is a group homomorphism, preserving all the group properties. Indeed, the following relations are easily verified: U (b, a)U (b , a ) = U (b + ab , aa ) U ((b, a)−1 ) = U (b, a)−1 = U (b, a)† U (e) = I, with e = (0, 1), the unit element.

(6.9)

We say that the association (b, a) → U (b, a) provides us with a unitary representation of G aff . Note that we may also write, U (b, a) = Tb Da , where Tb , Da are the well-known shift and dilation operators, familiar from standard time-frequency analysis (see also (2.7)–(2.12)): (Tb s)(x) = s(x − b),

(Da s)(x) = |a|− 2 s(a −1 x). 1

217

6.1 Group theory and matrix geometry of wavelets

We shall see later that the representation U (b, a) is in a sense minimal or irreducible, in that the entire Hilbert space of finite energy signals L 2 (R, d x) is needed to realize it completely. However, let us first attend to another question which is pertinent here, namely, why is it that G aff is made to act as a transformation group on R (see (6.3) and (6.4)), even without manifestly identifying R with any set of signal parameters? The answer to the above question lies in realizing that this space is intrinsic to the group itself. Indeed, let us factor an element (b, a) ∈ G aff in the manner a b 1 b a 0 (b, a) = = , (6.10) 0 1 0 1 0 1 and note that the first matrix on the right-hand side of this equation basically represents a point in R. We also note that the set of matrices of the type appearing in the second term of the above product is a subgroup of G aff . Dividing out by this matrix, we get (b, a)(0, a)−1 = (b, 0), which enables us to identify the point b ∈ R with an element of the quotient space G aff /H , (where H is the subgroup of matrices (0, a), a = 0). Next we see that, since a b 1 x a ax + b 1 ax + b a 0 = = , 0 1 0 1 0 1 0 1 0 1 the action of the group G aff on its quotient space G aff /H is exactly the same as its action on R as given in (6.3). Thus, the parameter space R on which the signals ψ(x) are defined is a quotient space of the group and hence intrinsic to the set of signal symmetries. We shall see below that the parameter space on which the wavelet transform of ψ is defined can also be identified with a quotient space of the group. In fact this space will turn out to be a phase space, in a sense to be specified later. Before moving on, let us reemphasize that the philosophy which seems to be emerging here is that the group (of signal symmetries) is the determinative quantity and all aspects of the signal and its various transforms emanate from it. We come back now to the point made earlier, that the representation U (b, a) was irreducible. We shall see that it also enjoys a second crucial property, that of being square integrable. The group G aff has a natural action on itself (by matrix multiplication from the left), according to which, for a given (b0 , a0 ) ∈ G aff , a general point (b, a) ∈ G aff is mapped to (b , a ) = (b0 , a0 )(b, a) = (b0 + a0 b, a0 a). It is not hard to see that the measure dµ(b, a) =

db da , a2

(6.11)

is invariant under this action: db da db da = . a2 a 2

(6.12)

218

Matrix geometry of wavelet analysis. I

We call the measure dµ the left Haar measure of G aff . In a similar manner we could obtain a right Haar measure dµr (invariant under multiplication from the right), which would turn out to be dµr (b, a) = a −1 db da. It is important to realize, that while these two measures are (measure theoretically) equivalent, they are not the same measure. The function *(b, a) = a −1 , for which dµ(b, a) = *(b, a) dµr (b, a), is called the modular function of the group. The square-integrability of the representation U (b, a) now means that there exist signals ψ ∈ L 2 (R, d x) for which the matrix element U (b, a)ψ | ψ is square integrable as a function of the variables b, a, with respect to the left Haar measure dµ, i.e., dµ(b, a) |U (b, a)ψ|ψ|2 < ∞, (6.13) G aff

and a straightforward computation would then establish that the function is also square integrable with respect to the right Haar measure. Furthermore, it is a fact that the existence of one such (nonzero) vector implies the existence of an entire dense set of them. Indeed, the condition for a signal to be of this type is precisely the condition of admissibility required of mother wavelets. To derive the admissibility condition, and also to verify our claim of irreducibility of the representation U (a, b), it will be convenient to go over to the Fourier domain. It is not hard to see that, on the Fourier-transformed space, the unitary operator U (b, a) (b, a), with explicit action, transforms to U (b, a)ψ (ξ ) = |a|1/2 ψ(aξ )e−ibξ (b ∈ R, a = 0). U (6.14) , dξ ) the image of The Fourier transform is a linear isometry, and we denote by L 2 (R 2 L (R, d x) under this map. It follows that the operators U (b, a) are also unitary and , dξ ) ∈ L 2 (R that they again constitute a unitary representation of the group G aff . Let ψ be a fixed nonzero vector in the Fourier domain. We will now show that the set of all , dξ ) and this is what will (b, a)ψ as (b, a) runs through G aff is dense in L 2 (R vectors U . Indeed, let constitute the mathematically precise statement of the irreducibility of U , dξ ) be a vector which is orthogonal to all the vectors U (b, a)ψ: χ ∈ L 2 (R (b, a)ψ = 0, ∀b ∈ R, a = 0. χ |U Using (6.14) we get, (b, a)ψ = |a|1/2 χ |U

∞

−∞

(aξ ) e−ibξ = 0. dξ χ (ξ ) ψ

(aξ ) = 0, almost every (ξ ) ψ By the unitarity of the Fourier transform, this yields χ where, for all a = 0. Since ψ ≡ 0, this in turn implies χ (ξ ) = 0, almost everywhere. , dξ ) which are stable under the action of all the Thus, the only subspaces of L 2 (R , dξ ) itself and the trivial subspace containing just the zero (b, a) are L 2 (R operators U , dξ ) is sort of a minimal space for the representation. vector. In other words, L 2 (R

219

6.1 Group theory and matrix geometry of wavelets

The unitarity of the Fourier transform also tells us that the representations U (b, a) and (b, a) are equivalent and since U (b, a) is irreducible, so also is U (b, a). (Note, this U is also clear from the fact that the linear isometry property of the Fourier transform implies that (b, a)ψ, χ | U (b, a)ψ = χ |U respectively.) χ , ψ denoting the inverse Fourier transforms of χ , ψ, Now we address the question of square integrability. We require that da db ψ| 2= |U (b, a)ψ| a2 G aff da

) ψ(aξ )ψ(ξ ) ) ψ(ξ db eib(ξ −ξ ) ψ(aξ = dξ dξ |a| da )|2 )|2 |ψ(ξ = 2π dξ |ψ(aξ |a| ∞ dξ 2 |ψ(ξ )| < ∞ = 2π ψ2 −∞ |ξ | (the integral over b yields a delta distribution, which can be used to perform the ξ integration and the interchange of integrals can be justified using standard distribution theoretic arguments). This means that the vector ψ is admissible in the sense of (6.13) if and only if ∞ dξ 2 cψ ≡ 2π (6.15) |ψ(ξ )| < ∞. −∞ |ξ | From this discussion we draw two immediate conclusions. First, there is a dense set of which satisfy the admissibility condition (6.15). Second, the admissibility vectors ψ condition (1.10) or (6.15), cψ < ∞, simply expresses the square integrability of the representation U . (Note that a vector ψ ∈ L 2 (R, d x) is admissible if and only if its , dξ ), on L 2 (R Fourier transform satisfies (6.15).) Defining an operator C !1 2π 2 (C ψ)(ξ ) = ψ(ξ ), (6.16) |ξ | and denoting by C its inverse Fourier transform, we see that the vector ψ is admissible if and only if cψ = Cψ2 < ∞ .

(6.17)

This operator, known as the Duflo–Moore operator, is positive, self-adjoint and unbounded. It also has an inverse. A straightforward computation, using (6.14) and (6.16) now shows that if a vector ψ satisfies (6.17), then so also does the vector U (b, a)ψ, for any (b, a) ∈ G aff . A word now about the form of the representation U (b, a). How does one arrive at it? In fact, given the way the group acts on R, x → ax + b, the representation U (b, a) is

220

Matrix geometry of wavelet analysis. I

recognized as being the most natural, nontrivial way to realize a group homomorphism onto a set of unitary operators on the signal space L 2 (R, d x). (Unitarity is required in order to ensure that the signal ψ and the transformed signal U (b, a)ψ both have the same total energy.) Indeed, given any differentiable mapping T : Rn → Rn , the operator U (T ), on the Hilbert space L 2 (Rn , d n x), defined as x )), (U (T ) f )( x ) = |det[J (T )]|− 2 f (T −1 ( 1

where J (T ) is the Jacobian of the map T , is easily seen to be unitary. (Recall that d(T ( x )) = |det[J (T )]| d x .) This provides the rationale for defining the representation U (b, a) by (6.8). Of course, the interesting point here is that this representation turns out to be both irreducible and square integrable. But then, why is square integrability of the representation a desirable criterion for wavelet analysis? In order to answer this question, let us take a vector ψ satisfying the admissibility condition (6.17) and use it to construct the wavelet transform of the signal s: S(b, a) = ψb,a | s. As we already know, the total energy of the transformed signal is given by the integral dµ(b, a) | S(b, a) | 2 , (6.18) E(S) = G aff

and we would like this to be finite, like that of the signal itself. An easy computation now shows that E(S) = Cψ2 s2 = cψ s2 ,

(6.19)

which means that the total energy of the wavelet transform will be finite if and only if the mother wavelet can be chosen from the domain of the operator C, i.e., if and only if it satisfies the square integrability condition (6.13). However, this is not the whole story, for let us rewrite the above equation in the expanded form E(S) = dµ(b, a) s | ψb,a ψb,a | s G aff ! = s | dµ(b, a) | ψb,a ψb,a | s G aff

= cψ s | I s, I being the identity operator on L 2 (R, d x) and where, for any nonzero vector φ ∈ L 2 (R, d x), the quantity | φφ | /φ2 denotes the one-dimensional projection operator along this vector. (We have also interchanged two integrations with the taking of a scalar product, without justifying it, but the manipulation can easily be justified using standard

221

6.1 Group theory and matrix geometry of wavelets

absolute convergence techniques.) Using the well-known polarization identity for scalar products, we infer from this the operator relation 1 dµ(b, a) | ψb,a ψb,a | = I, (6.20) cψ G aff also known as the resolution of the identity equation. It is immediately clear that (6.20) is completely equivalent to the square integrability of the representation U (b, a) as expressed in (6.13). The resolution of the identity also incorporates within it the possibility of reconstructing the signal s(x) from its wavelet transform S(b, a). To see this, let us act on the vector s ∈ L 2 (R, d x) with both sides of (6.20). We get 1 dµ(b, a) ψb,a ψb,a | s = I s = s, cψ G aff implying s(x) =

1 cψ

dµ(b, a) S(b, a)ψb,a (x),

almost everywhere,

(6.21)

G aff

which is the celebrated reconstruction formula we encountered before. Summarizing, we conclude that square integrability (which is a group property) is precisely the condition which ensures, in this case, the very desirable consequences of (i) the finiteness of the energy of the wavelet transform, and (ii) the validity of the reconstruction formula (two properties also shared by the Fourier transform of a signal). The resolution of the identity condition (6.20) has independent mathematical interest. First of all, it implies that any vector in L 2 (R, d x) which is orthogonal to all the wavelets ψb,a is necessarily the zero vector, i.e., the linear span of the wavelets is dense in the Hilbert space of signals. This fact, which could also have been inferred from the irreducibility of the representation U (b, a), is what enables us to use the wavelets as a basis set for expressing arbitrary signals. In fact we have here what is also known as an overcomplete basis. Second, this overcomplete basis is a continuously parametrized set, meaning that this is an example of a continuous basis and a continuous frame. There is a host of other useful mathematical properties of the wavelet transform and spaces of transforms, which emanate from square integrability. We proceed to examine a few.

6.1.2

The space of all wavelet transforms A finite energy wavelet transform S(b, a) is an element of the Hilbert space L 2 (H, dµ). Here we have written H = R × R∗ (R∗ = real line with the origin removed). Although, H and G aff are homeomorphic as topological spaces, we prefer to denote them by different symbols, for presently we shall identify H with a phase space of G aff , arising from its matrix geometry. Using (6.18) and (6.19) to compare the L 2 -norm of the

222

Matrix geometry of wavelet analysis. I

wavelet transform S (as an element in L 2 (H, dµ)) to the L 2 -norm of the signal s (as an element in L 2 (R, d x)), we get S2 = cψ s2 ,

(6.22)

which just means that, up to a constant, the wavelet transform preserves norms (i.e., energies). We define a map Wψ : L 2 (R, d x) → L 2 (H, dµ), by the relation $ %− 1 $ %− 1 (Wψ s)(a, b) = cψ 2 ψb,a | s L 2 (R,d x) = cψ 2 S(b, a).

(6.23)

This map is linear and, in view of (6.22), an isometry, so that its range, which is the set of all wavelet transforms corresponding to the mother wavelet ψ, is a closed subspace of L 2 (H, dµ). We denote the range by Hψ : $ % Hψ = Wψ L 2 (R, d x) ⊂ L 2 (H, dµ). (6.24) From the defining equation (6.23) we infer that Hψ consists of continuous functions over H and hence is a proper subspace of L 2 (H, dµ). It is worthwhile reiterating here the fact that the condition of Wψ being an isometry implies, not only that the wavelet transform (with respect to the mother wavelet ψ) of any signal s ∈ L 2 (R, d x) is an element in Hψ , but also that every element in Hψ is the wavelet transform of some signal s ∈ L 2 (R, d x).

6.1.2.1

An intrinsic characterization of the space of wavelet transforms Is there some convenient, intrinsic way to characterize the subspace Hψ ? To answer this question we appeal to the resolution of the identity and a bit of group theory. The final characterization will be spelled out in Theorem 6.1.1. Multiplying each side of equation (6.20) by itself we find, 1 dµ(b , a ) dµ(b, a) | ψb,a K ψ (b, a ; b , a ) ψb ,a | = I, (6.25) cψ H×H where we have written 1 ψb,a | ψb ,a . K ψ (b, a ; b , a ) = cψ

(6.26)

Acting on the signal vector s with both sides of (6.25) and using (6.23), we obtain ! 1

dµ(b, a) dµ(b , a ) K ψ (b, a ; b , a )S(b , a ) ψb,a (x) = s(x), cψ H H almost everywhere (the change in the order of integrations being easily justified by Fubini’s theorem). Comparing the above equation with the reconstruction formula (6.21) we obtain the interesting identity dµ(b , a ) K ψ (b, a ; b , a )S(b , a ) = S(b, a), (6.27) H

223

6.1 Group theory and matrix geometry of wavelets

for almost all (b, a) in H (with respect to the measure dµ). This then is the condition which characterizes wavelet transforms coming from the mother wavelet ψ. It is also known as the reproducing property of the integral kernel K ψ . As we know, the kernel K ψ : H × H → C is called a reproducing kernel. It has the easily verifiable properties: K ψ (b, a ; b, a) > 0, for all (b, a) ∈ H,

(6.28)

K ψ (b, a ; b , a ) = K ψ (b , a ; b, a), dµ(b

, a

) K ψ (b, a ; b

, a

) K ψ (b

, a

; b , a ) = K ψ (b, a ; b , a ),

(6.29) (6.30)

H×H

the last relation being again the reproducing property (6.27) in a different guise (and in fact following directly from it). The last two equations hold pointwise, for all (b, a), (b , a ) ∈ H. Next let us compute the wavelet transforms of the wavelets ψb,a themselves. Denoting the transforms by Sb,a we find Sb,a (b , a ) = ψb ,a | ψb,a = cψ K ψ (b , a ; b, a).

(6.31)

Since the vectors ψb,a , (b, a) ∈ H, are overcomplete in L 2 (R, d x), the wavelet transforms Sb,a must also be overcomplete in Hψ . One also has the easily verifiable resolution of the identity, $ %−2 cψ dµ(b, a) | Sb,a Sb,a | = Iψ , (6.32) H

where we have written Iψ for the identity operator on Hψ . Thus, any vector F ∈ L 2 (H, dµ) which lies in the orthogonal complement of Hψ must satisfy dµ(b , a ) K ψ (b, a ; b , a )F(b , a ) = 0. H

All this goes to say that the reproducing kernel K ψ defines the projection operator Pψ from L 2 (H, dµ) to Hψ : dµ(b , a ) K ψ (b, a ; b , a )F(b , a ), F ∈ L 2 (H, dµ), (6.33) (Pψ F)(b, a) = H

equations (6.29) and (6.30) mirroring the conditions Pψ = P∗ψ = P2ψ (star denotes the adjoint). Stated differently, an arbitrary vector F ∈ L 2 (H, dµ) can be uniquely written as the sum F = Fψ + Fψ⊥ , of a part Fψ ∈ Hψ and a part Fψ⊥ orthogonal to it. The operator Pψ , acting on F, projects out the part Fψ (which is a wavelet transform). It is natural to ask at this point if Fψ⊥ could also be written as the wavelet transform with respect to some other mother

224

Matrix geometry of wavelet analysis. I

wavelet. As will be seen below, generally Fψ⊥ can be written as an infinite sum of orthogonal wavelet transforms, corresponding to different mother wavelets. To proceed further, we go back to the affine group, G aff , and note that there is a natural unitary representation of it on the Hilbert space L 2 (H, dµ), given by its natural action on H. This representation U (b, a), called the left regular representation, acts in the manner (U (b, a)F)(b , a ) = F((b, a)−1 (b , a )) b − b a , , =F a a

F ∈ L 2 (H, dµ).

(6.34)

The unitarity of this representation, U (b, a)F2L 2 (H,dµ) = F2L 2 (H,dµ) , is guaranteed by the invariance of the measure dµ (see (6.12)). However, the left regular representation is by no means irreducible, since as we shall see below, the subspace Hψ carries a subrepresentation of it. The isometry Wψ (see (6.23)) maps the unitary operators U (b, a) onto unitary operators Uψ (b, a) = Wψ U (b, a)Wψ−1 on L 2 (H, dµ). Computing the action of these operators, using (6.8) and (6.23), we find b − b a

(Uψ (b, a)F)(b , a ) = F (6.35) , , F ∈ Hψ. a a This is the same action as that of the operators U (b, a) of the left regular representation, except that now it is expressed exclusively in terms of vectors in Hψ . This means, first of all, that the subspace Hψ is stable under the action of the operators U (b, a) and, secondly, that restricted to this subspace, it gives an irreducible unitary representation of G aff .

6.1.2.2

Decomposition of the space of all finite energy wavelet transforms Let ψ and ψ be two different mother wavelets. This means that they are both vectors in the domain of the operator C (with their Fourier transforms satisfying (6.15)). How are wavelet transforms, taken with respect to these mother wavelets, related? In particular, denoting by Sψ the wavelet transform of the signal s, taken with respect to the mother wavelet ψ, and by Sψ the wavelet transform of the signal s , taken with respect to the mother wavelet ψ , we would like to evaluate the overlap

dµ(b, a) Sψ (b, a) Sψ (b, a) I (ψ, ψ ; s , s) = H

dµ(b, a) s | ψb,a ψb,a | s. (6.36) = H

225

6.1 Group theory and matrix geometry of wavelets

Let us begin by assuming that s, s are taken from a class of smooth functions (e.g., the Schwartz class, S(R)), which is dense in L 2 (R, d x). Then db da b,a | s | ψ b,a ψ s I (ψ, ψ ; s , s) = ∗ a2 R R ! db da 1

(ξ ) ψ (aξ ) 2 e ibξ = dξ | a | s a2 R R∗ R ! 1

−ibξ

2 s(ξ ) ψ(aξ ) . dξ | a | e × R

We exploit the smoothness of the functions s, s to use the identity 1

db eib(ξ −ξ ) = δ(ξ − ξ ), 2π R which holds in the sense of distributions, and then perform the ξ -integration to obtain da (aξ ) )ψ s(ξ ) I (ψ, ψ ; s , s) = 2π dξ ψ(aξ s (ξ ). R R a Changing variables, we get " # ! (y) ψ ψ(y)

dy dξ s(ξ ) s (ξ ) · I (ψ, ψ ; s , s) = 2π |y| R R ψ ψ |C s | s. = C Thus,

dµ(b, a) Sψ (b, a)Sψ (b, a) = H

H

dµ(b, a) s | ψb,a ψb,a | s

= Cψ | Cψ s | s.

(6.37)

Using the continuity of the scalar product s | s in s and s , we may now extend the above expression to all signals s, s ∈ L 2 (R, d x). Equation (6.37) is a general orthogonality relation for wavelet transforms. In particular, if Cψ and Cψ are orthogonal vectors, then the corresponding wavelet transforms are also orthogonal in L 2 (H, dµ). We may also write this equation in the form of an operator identity on L 2 (R, d x):

dµ(b, a) | ψb,a ψb,a | = Cψ | Cψ I, (6.38) H

which clearly is a generalization of the resolution of the identity (6.20). At the risk of being pedantic, we would still like to emphasize that the above orthogonality relation implies: r If s and s are signals which are orthogonal vectors in L 2 (R, d x), then their wavelet transforms S and S , whether with respect to the same or different mother wavelets, are orthogonal as vectors in L 2 (H, dµ).

226

Matrix geometry of wavelet analysis. I r

Spaces of wavelet transforms, Hψ , Hψ , corresponding to mother wavelets ψ, ψ which satisfy the orthogonality condition Cψ ⊥ Cψ , are orthogonal subspaces of L 2 (H, dµ). Equation (6.38) is a remarkable result. Acting on a signal s ∈ L 2 (R, d x) with both sides of this equation, and assuming that Cψ | Cψ = 0, we get 1

dµ(b, a) Sψ (b, a)ψb,a , ψb,a = U (b, a)ψ , s= Cψ | Cψ H where Sψ (b, a) = ψb,a | s is the wavelet transform of s computed with respect to the mother wavelet ψ. Thus, although the wavelet transform is computed with respect to the mother wavelet ψ, it can be reconstructed using the wavelets of any other mother wavelet ψ , so long as Cψ | Cψ = 0. Moreover, up to a multiplicative constant, the reconstruction formula is exactly the same as that in which the same wavelet ψ is used both for analyzing and reconstructing (see (6.21)). This indicates, that in some sense, analysis and reconstruction are independent of the mother wavelet chosen. Let us choose a set of mother wavelets {ψn }∞ n=1 such that the vectors φn = Cψn form an orthonormal basis of L 2 (R, d x), φn | φm = Cψn | Cψm = δnm ,

n, m = 0, 1, 2, . . . , ∞.

(6.39)

Such a basis is easy to find and we shall construct one below. If Hψn , n = 0, 1, 2, . . . , are the corresponding spaces of wavelet transforms and K ψn the associated reproducing kernels, then Hψn ⊥ Hψm , for n = m, and dµ(b

, a

) K ψn (b, a ; b

, a

) K ψm (b

, a

; b , a ) = δnm K ψm (b, a ; b , a ). H

(6.40)

More interestingly, it is possible to show that the complete decomposition, L 2 (H, dµ)

∞

Hψn ,

(6.41)

n=1

of the space of all finite energy signals (on the parameter space H) into an orthogonal direct sum of spaces of wavelet transforms, holds. Thus, in an L 2 -sense, any element S ∈ L 2 (H, dµ) has the orthogonal decomposition, S(b, a) =

∞

Sn (b, a),

almost everywhere,

n=1

into orthogonal wavelet transforms Sn , with respect to a basis of mother wavelets. This result, which can be proved by direct computation in the present case, is actually a particular example of a much more general result on the decomposition of the left regular representation of a group into irreducibles. For a more detailed mathematical discussion of this point, we refer the reader to [Ali00]. The components Sn (b, a) have the form

227

6.1 Group theory and matrix geometry of wavelets

Sn (b, a) = U (b, a)ψn | sn ,

n = 0, 1, 2, . . .

(6.42)

for some signal vectors sn ∈ L 2 (R, d x), which, in general, are different for different n. We also have the relations dµ(b , a ) K ψn (b, a ; b , a )Sm (b , a ) = δnm Sm (b, a). (6.43) H

Finally, we construct an explicit example of a basis set of mother wavelets satisfying (6.39). Let Hn (ξ ), n = 0, 1, 2, . . . , ∞, be the Hermite polynomials, normalized in the manner 0, if m = n, −ξ 2 (6.44) dξ e Hm (ξ ) Hn (ξ ) = √ 2n n! π , if m = n. R The first few are: H0 (ξ ) = 1,

H1 (ξ ) = 2ξ,

H2 (ξ ) = 4ξ − 2,

H3 (ξ ) = 8ξ 3 − 12ξ,

2

etc.

In the Fourier domain, define n (ξ ) = ψ

1 3 4

π 2

n+1 2

ξ2

√ | ξ | 2 e− 2 Hn (ξ ). n! 1

(6.45)

Then, it is easily verified that n 2 2 ψ < ∞, L (R,dξ ) and ψ ψ n = 2π m | C C

R

dξ n (ξ ) = δmn . ψm (ξ ) ψ |ξ |

Thus, in the inverse Fourier domain, the vectors ψn are in the domain of the operator C, satisfying the condition for being mother wavelets, while from the well-known properties of Hermite polynomials, the vectors φn constitute an orthonormal basis of L 2 (R, d x). More generally, since the range of C is dense in L 2 (R, d x), we can take any 2 orthonormal basis, {φn }∞ n=1 of L (R, d x), chosen from vectors in this range and then ∞ {ψn = C −1 φn }n=1 will be the desired wavelet basis. We collect the above results into a theorem: Theorem 6.1.1 The wavelet transform of the space of signals L 2 (R, d x), with respect to the mother wavelet ψ, is a closed subspace of L 2 (H, dµ). This subspace has a reproducing kernel K ψ , which is the integral kernel of the projection operator, Pψ : L 2 (H, dµ) → Hψ . The Hilbert space L 2 (H, dµ) can be completely decomposed into an orthogonal direct sum of an infinite number of subspaces Hψn , each a space of wavelet transforms with respect to a mother wavelet ψn . The vectors ψn are constructed by

228

Matrix geometry of wavelet analysis. I 2 taking an orthonormal basis {φn }∞ n=1 of L (R, d x), chosen from the range of the Duflo– −1 Moore operator C, and writing ψn = C φn .

6.1.2.3

Decomposition into orthogonal channels The above theorem can be used to analyze a given wavelet transform into orthogonal channels along the lines of [105]. Referring back to (6.39), since the mother wavelets {ψn }∞ n=1 form a complete, linearly independent set, any mother wavelet ψ can be written as a linear combination, ψ=

∞

an ψn ,

an = φn | Cψ = ψn | C 2 ψ.

(6.46)

n=1

Hence if s ∈ L 2 (R, d x) is any signal vector and Sψ (b, a) its wavelet transform with respect to ψ, then clearly Sψ (b, a) =

∞

Sn (b, a) ,

where

Sn (b, a) = an U (b, a)ψn | s.

(6.47)

n=1

In this way, the wavelet transform of the signal Sψ (b, a) has been decomposed into a set of mutually orthogonal wavelet transforms Sn (b, a) (of this same signal). We call this a decomposition into orthogonal channels. Note that although the wavelet transforms are orthogonal in L 2 (H, dµ), the mother wavelets ψn are not orthogonal in L 2 (R, d x) (see (6.39)). We shall see in the next chapter (at the end of Section 7.2.3), that in the case of a 2-D wavelet transform, it will actually be possible to obtain a decomposition into orthogonal (angular) channels using a family of mother wavelets which are themselves mutually orthogonal.

6.1.3

Localization operators Let ψ be a mother wavelet and Hψ the corresponding space of wavelet transforms. Let

⊂ H be a measurable set (with respect to the measure dµ). We associate to this set an integral kernel aψ : H × H → C:

dµ(b

, a

) K ψ (b, a ; b

, a

)K ψ (b

, a

; b , a ), (6.48) aψ (b, a; b , a ) =

and an operator aψ ( ) on Hψ acting via this kernel: dµ(b , a ) aψ (b, a ; b , a )S(b , a ), (aψ ( )S)(b, a) = H

S ∈ Hψ .

(6.49)

If we compute the matrix element of this operator for the wavelet transform S, using the properties of the reproducing kernel (see (6.29)–(6.30)), we easily obtain, dµ(b, a) | S(b, a) | 2 . (6.50) S | aψ ( )S =

229

6.1 Group theory and matrix geometry of wavelets

which shows that this operator is bounded, positive and self-adjoint. We call aψ ( ) a localization operator, since in view of the above relation, the quantity p S ( ) =

S | aψ ( )S S2

(6.51)

is the fraction of the wavelet transform which is localized in the region of phase space. If S(b, a) = ψb,a | s, then we shall also write ps ( ) for the above probability, for indeed, it measures the concentration of the phase space content (e.g., time–frequency content) of the signal s in the region . As a set function, p S ( ) has the properties of a probability measure: p S (∅) = 0, ∅ = empty set, p S (H) = 1, 7 8 p S ( i ) = p S ( i ), if i

j = ∅, whenever i = j, i∈J

(6.52)

i∈J

J being some index set. Since this holds for all S ∈ Hψ , we say that the operators aψ ( ) themselves constitute a positive operator-valued measure, or POV-measure, satisfying the properties: aψ (H) = Iψ , aψ (∅) = 0, 7 8 aψ ( i ), if i

j = ∅, whenever i = j, aψ ( i ) = i∈J

(6.53)

i∈J

where the sum on the right-hand side of (6.53) has to be understood in the weak sense, i.e., in the sense of (6.52). It is instructive to see how p S ( ) changes if the set gets transformed under the action of the group G aff . Since, for any (b0 , a0 ) ∈ G aff , U (b0 , a0 )ψb,a | ψb ,a L 2 (R,d x) = ψb,a | U (b0 , a0 )∗ ψb ,a L 2 (R,d x) ,

(6.54)

using the group properties (6.9) and the definition of the reproducing kernel in (6.26), we find that it satisfies the following covariance property: b − b0 a K ψ (b0 + a0 b, a0 a ; b , a ) = K ψ (b, a ; , ), (6.55) a0 a0 i.e., K ψ ((b0 , a0 )(b, a) ; b , a ) = K ψ (b, a ; (b0 , a0 )−1 (b , a )).

(6.56)

Let (b0 , a0 ) denote the translate of the set by (b0 , a0 ): (b0 , a0 ) = {(b0 + a0 b, a0 a) ∈ H | (b, a) ∈ }. Then, taking note of the action (6.35) of the left regular representation of G aff on wavelet transforms and exploiting the invariance of the measure dµ, we easily find that Uψ (b, a)S | aψ ( )Uψ (b, a)S = S | aψ ((b, a)−1 )S , i.e., we have the operator identity

S, S ∈ Hψ ,

230

Matrix geometry of wavelet analysis. I

Uψ (b, a)aψ ( )Uψ (b, a)∗ = aψ ((b, a) ).

(6.57)

This is a group covariance condition satisfied by the localization operators aψ ( ), and is generally known as an imprimitivity relation (see, e.g., [Ali00], for a more detailed discussion). For the probability measure p S ( ), this condition implies the transformation property p S ((b, a) ) = pUψ (b,a)−1 S ( ),

or,

ps ((b, a) ) = pU (b,a)−1 s ( ).

(6.58)

Physically, this relation means that the fraction of the signal s, localized in the transformed set (b, a) , is the same as the fraction of the transformed signal U (b, a)−1 s localized in the original set .

6.2

Phase space analysis We turn our attention to a different way of understanding the wavelet transform, namely, as a function on a phase space (in a sense to be made clear in the sequel). First let us (b, a) of the full affine group (see (6.14)) to the connected restrict the representation U + affine group G aff characterized by a > 0. We immediately see that this representation is no longer irreducible for this smaller group. Indeed, consider the two subspaces + (R ) = { , dξ ) | H f ∈ L 2 (R f (ξ ) = 0, ∀ ξ < 0}, 2 H − (R ) = { f ∈ L (R, dξ ) | f (ξ ) = 0, ∀ ξ > 0},

(6.59)

, dξ ) of the representation U (b, a). From (6.14) it is evident that of the carrier space L 2 (R vectors in any one of these subspaces are mapped to vectors in the same subspace under (b, a), when we only consider elements (b, a) ∈ G + the action of the operators U aff . This means that each one of these subspaces carries a unitary representation of this smaller group and, as before, we can show that both these representations, which we denote by + (b, a) and U − (b, a), respectively, are irreducible but unitarily inequivalent. In fact, U these are the only two nontrivial, unitary irreducible representations of G + aff . Moreover, , dξ ) is the orthogonal direct sum of these two subspaces: the Hilbert space L 2 (R + (R − (R , dξ ) = H ) ⊕ H ). L 2 (R

(6.60)

(b, a) In other words, we have here a complete decomposition of the representation U + of the connected affine group G aff and we write − (b, a). (b, a) = U + (b, a) ⊕ U U

(6.61)

In the inverse Fourier domain, the representation U (b, a) similarly breaks up (because of the unitarity of the Fourier transform) into two irreducible representations, U+ (b, a) and U− (b, a), on the two subspaces

231

6.2 Phase space analysis

H+ (R) = { f ∈ L 2 (R, d x) | f (ξ ) = 0, ∀ ξ < 0}, 2 H− (R) = { f ∈ L (R, d x) | f (ξ ) = 0, ∀ ξ > 0},

(6.62)

of L 2 (R, d x), respectively. These spaces are known as Hardy spaces [55,207,208]. Elements of H+ (R) (respectively, H− (R)) extend to functions analytic in the upper (respectively, lower) complex half-plane, and accordingly they are called upper (respectively, lower) analytic signals [Lyn82,Pap77].

6.2.1

Holomorphic wavelet transforms It is an interesting fact that, for appropriate choices of a mother wavelets, the wavelet transforms of signals in the spaces H± (R) can (up to a factor) become holomorphic functions. We study this property in some detail in this section. For any ν 0, consider + (R ), ∈H the mother wavelet ψ 12 ν+1 2ν ξ 2 e−ξ , for 0 < ξ < ∞, π (ν+1) )= ψ(ξ 0, otherwise, ψ 2 = 1, C

(6.63)

(ν + 1) being the usual Gamma function. The wavelets for this vector have the form b,a (ξ ) = ψ

2ν π (ν + 1)

! 12

ν

a 1+ 2 ξ

ν+1 2

e−iξ z ,

where z = b + ia,

(6.64)

and the reproducing kernel is ν

b,a | ψ b ,a = K ψ (b, a ; b , a ) = ψ

2ν (ν + 1) (aa )1+ 2 . i 2+ν π (z − z)2+ν

(6.65)

Computing the wavelet transform of a signal s ∈ H+ (R), b,a | s= S(b, a) = ψ

2ν π (ν + 1)

! 12 a

1+ ν2

∞

dξ ξ

1+ν 2

s(ξ ). eiξ z

(6.66)

0

ν

This, apart from the factor a 1+ 2 , is a holomorphic function of z = b + ia. Indeed, writing this function as !− 12 2ν ν F(z) = a −(1+ 2 ) S(b, a) π (ν + 1) ∞ 1+ν s(ξ ), z = b + ia, dξ eiξ z ξ 2 = 0

we have,

(6.67)

232

Matrix geometry of wavelet analysis. I

+ + | F(z) | = ++ 2

∞

1+ν 2

+2 + + + s(ξ ) ++ = ++

dξ ξ e ∞ s2 dξ ξ 1+ν e−2ξ a , iξ z

0

∞

dξ ξ

0

1+ν 2

e

−ξ a iξ b

e

+2 + s(ξ ) ++

by the Cauchy–Schwarz inequality

0

=

s2 (ν + 2) . (2a)2+ν

Thus, the convergence of the integral representing F(z) is uniform in any bounded open set containing z and differentiation with respect to it, under the integral sign, is permissible, implying that F is holomorphic in z, on the complex, upper half plane which we identify with H+ = R × R∗+ (where, R∗+ = (0, ∞)). We shall call F(z) a holomorphic wavelet transform. Furthermore, since db da 2 | S(b, a) | = dµν (z, z) | F(z) | 2 , a2 H+ H+ where we have introduced the measure, dµν (z, z) =

(2a)ν db da, π (ν + 1)

(6.68)

the set of all holomorphic wavelet transforms constitutes a closed subspace of the Hilbert space L 2 (H+ dµν ) of functions supported on the upper half plane. We denote this subspace by Hνhol and note that it is also a reproducing kernel Hilbert space, with reproducing kernel ν K hol (z, z ) =

(ν + 2) 1 . 2+ν i (z − z )2+ν

(6.69)

One has indeed ν dµν (z, z) K hol (z, z ) F(z ) H+

2ν (ν + 1) = i 2+ν = F(z).

H+

a ν da db

F(z ) (z − z )2+ν

The vectors ηz , z ∈ H+ , with ν ηz (z ) = K hol (z , z),

(6.70)

which are the holomorphic wavelets, are again overcomplete in Hνhol and satisfy the resolution of the identity: ν dµν (z, z) | ηz ηz | = Ihol ( = identity operator of Hνhol ). (6.71) H+

233

6.2 Phase space analysis ν There is also the holomorphic representation of G + aff on Hhol , unitarily equivalent to ν U+ (b, a). Denoting this by Uhol (b, a), we easily compute its action: z−b ν ν (b, a)F)(z) = a −(1+ 2 ) F (Uhol . (6.72) a

The appearance of the holomorphic Hilbert spaces of wavelet transforms is remarkable in many ways. First of all, their existence is related to a geometrical property of the half plane H+ , which is a differential manifold with a complex K¨ahler structure. This means, from a physical point of view, that it has all the properties of being a phase space of a classical mechanical system and, additionally, that this phase space can be given a complex structure (consistent with its geometry). In particular, it has a metric and a preferred differential two-form, which gives rise to the invariant measure dµ and using which classical mechanical quantities, such as Poisson brackets, may be defined. We will not go into the details of this here, but only point out the existence of a potential function in this context. Consider the function, (z, z ) = − log[−(z − z )2 ].

(6.73)

This function is called a K¨ahler potential for the space H+ and it generates all the interesting quantities characterizing its geometry, such as the invariant two-form and the invariant measure. Indeed, we immediately verify that

ν

e(1+ 2 ) (z,z ) =

i 2−ν K ν (z, z ). (ν + 2) hol

(6.74)

Next we define (z, z) =

1 ∂ 2 (z, z) db ∧ da , dz ∧ dz = i ∂z ∂z a2

(6.75)

which is the invariant two-form (under the action of G + aff ). This gives the invariant measure dµ of the group and furthermore, ν

e−(1+ 2 ) (z,z) = 4(−1)ν (2a)ν db ∧ da,

(6.76)

from which follows the measure with respect to which the holomorphic functions F(z) are square integrable and form a Hilbert space. (Recall that if u and v are two vectors in a vector space V , then u ∧ v is the antisymmetric tensor product, u ∧ v = u ⊗ v − v ⊗ u. The differentials, dz, da, etc., are considered as being elements in the dual of the tangent space of the manifold – in this case G + aff – at each point.) It ought to be emphasized here that Hνhol contains all holomorphic functions in L 2 (H+ , dµν ). Note also that, in view of (6.67), any such function can be obtained by computing the Fourier transform of 1+ν a function f (ξ ) = ξ 2 s(ξ ) and then analytically continuing it to the upper half plane, where s is a signal in the Fourier domain, with support in (0, ∞). Additionally, it ought to be noted that, for each ν > 0, we get a family of holomorphic wavelet transforms,

234

Matrix geometry of wavelet analysis. I

so that depending on the value of ν, the same signal s can be represented by different holomorphic functions on phase space.

6.2.2

Matrix analysis of phase space We have said earlier that the variables (b, a) parametrizing the space H+ , and in terms of which the wavelet transform is written, should be identified as phase space variables. In this section, we proceed to elaborate on this. To begin, let us determine the Lie algebra of the group G + aff . This group has two subgroups, formed by matrices of the type t e 0 1 t t (e , 0) = and (1, t) = , t ∈ R, 0 1 0 1 and a general element of the group can be obtained by multiplying two such matrices. Consider now the following two matrices X 1 , X 2 , which generate the Lie algebra gaff of this group: 1 0 0 1 + d t ++ d , X2 = , (6.77) (e , 0) t=0 = (1, t) + t=0 = X1 = dt dt 0 0 0 0 and satisfy the commutation relation, [X 1 , X 2 ] = X 1 X 2 − X 2 X 1 = X 2 .

(6.78)

Exponentiating these matrices we get, a 0 1 b (log a)X 1 bX 2 = , e = . e 0 1 0 1

(6.79)

The Lie algebra of the group is the two-dimensional vector space spanned by X 1 and X 2 and equipped with the commutation relation (6.78). A general element in the Lie algebra can be written as, 1 x2 x 1 2 X = x X1 + x X2 = (6.80) , x 1 , x 2 ∈ R. 0 0 Any group element can be obtained by exponentiating a suitable element of this Lie algebra. This is made clear if we write x1 x2 x1 a b (e − 1) e x1 X = . (6.81) (b, a) = e = 0 1 0 1 and note the inverse map from the group to the algebra: X = log(b, a) = x 1 X 1 + x 2 X 2 ,

x 1 = log a,

x2 =

b log a . a−1

(6.82)

235

6.2 Phase space analysis

Since every X ∈ gaff is mapped to an element (b, a) ∈ G + aff by the exponential map (6.81), we identify the domain of this map with the full real plane and use x = (x 1 , x 2 ) ∈ R2 as the coordinates for the elements of the Lie algebra. A group has a natural action on its Lie algebra, called the adjoint action. For (b, a) ∈ + G aff this action, which we denote by Ad(b,a) , is defined by 1 1 2 x −bx + ax . (6.83) Ad(b,a) X = (b, a)X (b, a)−1 = 0 0 The matrix of thistransformation, computed in the basis {X 1 , X 2 }, and acting on the x1 ∈ R2 is vectors x = x2 M(b, a) =

1 −b

0 . a

(6.84)

As a vector space, the Lie algebra has a dual space, which we denote by g∗aff . On it the * adjoint action of G + aff induces, by duality, the coadjoint action, denoted Ad(b,a) . To compute this action we take the dual basis {X ∗ 1 , X ∗ 2 } in g∗aff and write ageneral element γ1 in it as X ∗ = γ1 X ∗ 1 + γ2 X ∗ 2 . We identify X ∗ with the vector γ = ∈ R2 . On γ2 such vectors, the coadjoint action of the group is given (by definition) by the transposed inverse of the matrix M(b, a). We write −1 % $ 1 ba T M * (b, a) = M(b, a)−1 = . (6.85) 0 a −1 The determinants of these transformations are related as *

det[Ad(b,a) ] = a = det[Ad(b,a) ]−1 . Explicitly, a point γ ∈ R2 transforms under the coadjoint action as, −1 γ + ba γ 1 2 γ → γ = M * (b, a)γ = , a −1 γ2

(6.86)

(6.87)

and for fixed γ0 ∈ R2 , its orbit under G + aff is the set: 2 Oγ∗0 = {γ = M * (b, a)γ0 | (b, a) ∈ G + aff } ⊂ R .

(6.88)

Such an orbit is called a coadjoint orbit of the group G + aff . Orbits of different points either coincide entirely or are disjoint. In this way, the entire dual space g∗aff becomes the union over disjoint coadjoint orbits. Indeed, we easily establish the following orbit structure:

236

Matrix geometry of wavelet analysis. I

(i)

(ii)

(iii) (iv)

0 The orbit of the vector , 1 γ1 ∗ (6.89) ∈ R2 | γ2 > 0} = R × R∗+ . O+ = {γ = γ2 0 The orbit of the vector with the same matrices, −1 γ1 ∗ (6.90) ∈ R2 | γ2 < 0} = R × R∗− . O− = {γ = γ2 α α , The orbits of vectors , one for each α ∈ R. Such orbits are singletons, 0 0 and we denote them by Oα∗ . It is obvious that ∗ ∗ ∪ O− ∪α∈R Oα∗ = R2 . O+

(6.91)

We note also that each of the two orbits (i) and (ii) is homeomorphic to the group ∗ = H+ = R × R∗+ . As a manifold, this space itself. Consider the first one of these, O+ has a symplectic structure, i.e., there exists a preferred nondegenerate, antisymmetric, closed differential two-form on it, which is invariant under the coadjoint action. Indeed, it is trivially verified that the two-form (γ ) =

dγ1 ∧ dγ2 γ2

(6.92)

satisfies this condition. This also gives the invariant measure on this phase space, in ∗ these coordinates. We look upon O+ as an abstract differential manifold, with (γ1 , γ2 ) representing a particular choice of coordinates. In this context, we study two other possible choices of coordinates and see how the two-form (6.92) appears in these new coordinates. As the first of these coordinate transformations, we write q γ2−1 γ1 γ → η = = , (6.93) p γ2 which maps R × R∗+ onto itself. Under the coadjoint action of the group element (b, a), these coordinates transform as (q, p) → (q , p ) with, q = aq + b,

p = a −1 p.

The invariant two-form is simply ( η) = dq ∧ dp,

(6.94)

237

6.2 Phase space analysis

giving also the invariant Liouville measure in these coordinates. We call these coordi∗ nates the canonical or Darboux coordinates of the phase space O+ , in view of the form, familiar from classical Hamiltonian mechanics, assumed by the two-form . This also allows us to identify the variables q and p as position and momentum, respectively. We also immediately recognize the coadjoint action as inducing canonical transformations on phase space. As the second coordinate transformation, we choose −1 ξ γ γ 1 1 2 γ → ξ = = , (6.95) ξ2 γ2−1 which again maps R × R∗+ onto itself. Under the coadjoint action, these coordinates change in the manner: ξ1 = aξ1 + b,

ξ2 = aξ2 ,

(6.96)

with the invariant two-form now being (ξ ) =

dξ1 ∧ dξ2 . ξ22

(6.97)

The invariant measure arising from this should be compared to the left invariant measure, dµ, of the group G + aff (see (6.11)). Indeed, if we identify ξ1 , ξ2 with group parameters, then the transformation (6.96) is just the left multiplication in the group: (ξ1 , ξ2 ) = ∗ (b, a)(ξ1 , ξ2 ). It is because of the possibility of coordinatizing the orbit O+ in this particular way, that we may legitimately look upon the group elements themselves as phase space variables and the wavelet transform as a transform on phase space. ∗ Similar considerations apply to the other nontrivial orbit O− . We shall see in the next chapter that the two-dimensional wavelet transform can also be analyzed in a completely analogous manner. Before concluding this section, we should point out one crucial fact about the orbit structure of G + aff . The composition rule (6.6) equips this group with the structure of a semidirect product, R R∗+ , in which the subgroup R∗+ has an action on the subgroup R, in this case by simple multiplication. This action and α ∈ R∗ ), on the induces an action (again by multiplication, x → αx, for x ∈ R + dual space R (which we naturally identify with R). Under this action the dual space splits up into three orbits, the open half-spaces R∗+ , R∗− and the singleton {0}. The first two orbits are open and free, in the sense that they are both homeomorphic to the subgroup R∗+ itself and the map identifying the subgroup with the orbit is continuous . The fact and open. Furthermore, the union of the two open free orbits is dense in R ∗ that the orbits O± are open and free, is what leads to the representations U± (b, a) being square integrable (see, for example, [Ali00] for a detailed discussion of this point). ∗ Geometrically, the coadjoint orbits O± (of the whole group G + aff ) are the cotangent ∗ bundles of the orbits R± (of the subgroup R∗+ ).

238

Matrix geometry of wavelet analysis. I

6.3

The case of Gabor wavelets For the sake of completeness and comparison, we briefly look at the case of Gabor wavelets, or short-time Fourier transforms, for they too can be understood in group theoretical and phase space terms which are similar to those of the standard wavelets considered above. The one difference here is that the phase space of signals is not the entire group in question (the Weyl–Heisenberg group) and square integrability of the representation, giving rise to the Gabor wavelets, has also to be understood in the light of this fact. Another remarkable difference here is that for building Gabor type wavelets, any vector in the Hilbert space of signals can be used as a mother wavelet – the admissibility condition is trivial. The Gabor transform was introduced in (1.3). It was obtained by taking a window function ψ ∈ L 2 (R, d x), translating by b (in time) and modulating it in frequency 1/a, to obtain the Gabor wavelets or gaborettes. It had the form: ψb,a (x) = ei(x−b)/a ψ(x − b). The Gabor transform of a signal s is then defined in the same way as the wavelet transform: S(b, a) = ψb,a | s = d x e−i(x−b)/a ψ(x − b) φ(x). (6.98) R

It is a remarkable fact that the Gabor transform, and the 1-D continuous wavelet transform are built on exactly the same pattern (and this similarity persists in higher dimensions). Both are based on transformation groups of signals in L 2 (R, d x), the action of the group being implemented by unitary operators, s → U (b, a)s: r for the Gabor transform, the group is the Weyl–Heisenberg group G , acting as in WH (1.3); r for the continuous wavelet transform it is the ax + b or affine group, acting as in (1.4), and in fact the translation parts are identical in the two cases.

6.3.1

Group theoretical analysis In order to understand the Gabor transform group theoretically, let us first rewrite the gaborettes by adopting a slightly different notation and also introducing an additional phase variable. We write q for b and p for 1/a and denote the phase variable by θ : q

ψθ,q, p (x) = eiθ ei p(x− 2 ) ψ(x − q),

θ ∈ R.

The gaborettes would then correspond to setting θ = − pq/2. Writing

(6.99)

239

6.3 The case of Gabor wavelets

ψθ,q, p = U (θ, q, p)ψ,

(6.100)

an easy computation shows that U (θ, q, p) defines a unitary operator on the space of signals. Computing the effect of two such operators in succession, on a vector ψ yields the composition rule, U (θ1 , q1 , p1 ) U (θ2 , q2 , p2 ) = U (θ1 + θ2 + ξ ((q1 , p1 ); (q2 , p2 )), q1 + q2 , p1 + p2 ), (6.101) where, 1 (6.102) ( p1 q2 − p2 q1 ), 2 which in fact defines the multiplication rule between elements of the Weyl–Heisenberg group G WH . The function ξ is called a multiplier. The Weyl–Heisenberg group is a three-parameter group, homeomorphic to R3 . An arbitrary element g of G WH is of the form

ξ ((q1 , p1 ); (q2 , p2 )) =

g = (θ, q, p),

θ ∈ R,

(q, p) ∈ R2 ,

and the multiplication law in the group is g1 g2 = (θ1 + θ2 + ξ ((q1 , p1 ); (q2 , p2 )), q1 + q2 , p1 + p2 ).

(6.103)

The multiplier ξ equips the group with the structure of a central extension, of the group of translations of R2 , corresponding to phase space translations (i.e., translations q and modulations p). This just means that the subgroup consisting of the elements + = {g = (θ, 0, 0) | θ ∈ R}, is the center of the group G WH , i.e., these elements commute with every element g in the group, g(θ, 0, 0) = (θ, 0, 0)g, and it is the introduction of ξ which extends the commutative group R2 , with elements (q, p), into the noncommutative Weyl–Heisenberg group. The Weyl–Heisenberg group is unimodular, the measure dµ = dθ dq d p being invariant under both the left action, (θ, q, p) → (θ0 , q0 , p0 )(θ, q, p), and the right action, (θ, q, p) → (θ, q, p)(θ0 , q0 , p0 ). As with the affine group, the Weyl–Heisenberg group also has a matrix realization. It is given by the 4 × 4 matrices 1 12 ζ T ω θ 0 −1 (θ, q, p) = ω= , (6.104) ζ I2 0 , 1 0 0 0 T 1 with θ ∈ R,

ζ =

q p

∈R , 2

0 =

0 0

,

I2 =

1

0

0

1

.

Also, as for the affine group, we can compute the generators of the Lie algebra of G WH , by considering the one-parameter subgroups of elements of the type, g1 (t) = (t, 0, 0), g2 (t) = (0, t, 0) and g3 (t) = (0, 0, t), with t ranging through R. Computing

240

Matrix geometry of wavelet analysis. I

+ d gi (t) +t=0 , dt we obtain the three elements which generate the Lie algebra, gWH : T 1 T 1 T 0 − 2 e3 0 1 e 0 2 2 X0 = , X2 = , , X1 = O e1 O e2 O 0 Xi ≡

where 0 0 = 0 , 0

1 e 1 = 0 , 0

0 e 2 = 1 , 0

(6.105)

0 e3 = 0 , 1

and O is the 3 × 3 zero matrix. The matrices X i satisfy the commutation relations [X 0 , X 1 ] = [X 0 , X 2 ] = 0,

[X 1 , X 2 ] = X 0 .

(6.106)

A general element of gWH can be written as x 1 , x 2 , x 3 ∈ R;

X = x 0 X 0 + x 1 X 1 + x 2 X 2,

the commutation relations (6.106) then define a Lie bracket, [X, Y ] = X Y − Y X , between any two elements X, Y ∈ gWH . Next, noting that for any X ∈ gWH , (X )2 is the null matrix (which, in group theoretical terms, is stated by saying that the group G WH is nilpotent), we see that e X = (x 0 , x 1 , x 2 ) = I4 + X ∈ G WH .

(6.107)

Thus, the group and the Lie algebra can be given the same parametrization. In order to make the connection with gaborettes, we need to find unitary irreducible representations of G WH on the Hilbert space of signals, L 2 (R, d x). That is, we need to find a set of unitary operators, U (θ, q, p), for all (θ, q, p) ∈ G WH , which realize a group homomorphism, are stable under inverse taking and map the identity element (0, 0, 0) of the group to the identity operator on L 2 (R, d x). But this is already done by the operators defined in (6.100), so that U (θ, q, p) realize a unitary representation of the group. The fact that it is also irreducible can be proved in much the same way in which we proved irreducibility for the representation of the affine group in Section 6.1.1. More generally, it can be shown that any (nondegenerate) unitary irreducible representation of G WH is characterized by a real number λ = 0 and may be realized on the Hilbert space L 2 (R, d x) by the operators U λ (θ, q, p): q

(U λ (θ, q, p)s)(x) = eiλθ eiλp(x− 2 ) s(x − q),

φ ∈ L 2 (R, d x).

(6.108)

Two representations U λ (θ, q, p) and U λ (θ, q, p) are unitarily inequivalent if λ = λ . For the construction of gaborettes we shall mostly work with the representation for which λ = 1 and denote it simply by U instead of U 1 . (The other representations will be used in our discussion of holomorphic Gabor transforms below.)

241

6.3 The case of Gabor wavelets

A general element (θ, q, p) in G WH can be factorized as (θ, q, p) = (0, q, p)(θ, 0, 0). Thus, the quotient space G WH /+ (where the phase is factored out) is identifiable with R2 . It is this space which plays the rˆole of a phase space for the Weyl–Heisenberg group. We parametrize an element in G WH /+ by (q, p) ∈ R2 and since for a fixed element (θ0 , q0 , p0 ) ∈ G WH we again have the factorization 1 (θ0 , q0 , p0 )(0, q, p) = (00 , q + q0 , p + p0 )(θ0 + ( p0 q − pq0 ), 0, 0), 2 the action of the group on the quotient space is simply given by (q, p) → (q + q0 , p + p0 ). The invariant measure under this action is just the Lebesgue measure dq d p. For any ψ ∈ L 2 (R, d x), let us define the vectors qp ψq, p = U (− , q, p)ψ, (6.109) (q, p) ∈ G WH /+. 2 These are just gaborettes, expressed in the parameters q, p. The corresponding Gabor transform of a signal s is ∞ d x e−i p(x−q) ψ(x − q) s(x). (6.110) S(q, p) = ψq, p | s = −∞

Once again, computing the total energy of the transform, dq d p |S(q, p)|2 = 2π ψ2 s2 , E(S) =

(6.111)

R2

which should be compared to (6.19). From this there follows the resolution of the identity (exactly as in the derivation of (6.20)) 1 dq d p |ψq, p ψq, p | = I, (6.112) 2πψ2 R2 and the reconstruction formula (see (6.21)), 1 dq d p S(q, p)ψq, p (x), s(x) = 2πψ2 R2

for almost all x.

(6.113)

In the physical literature, gaborettes, and indeed also wavelets, are referred to as coherent states, of their respective groups [Kla85, Ali00, 6]. Note, however, that unlike the wavelets ψb,a , which were defined for all elements (b, a) of the affine group, the gaborettes ψq, p are only defined for points in the quotient space G WH /+. Moreover, the two (equivalent) conditions (6.111) and (6.112) imply that the representation U (θ, q, p) is square integrable with respect to this space: qp dq d p |U (− , q, p)ψ|ψ|2 < ∞, ∀ψ ∈ H. (6.114) 2 G WH /+ Again, one ought to emphasize here that the above admissibility condition is satisfied by all vectors ψ in the signal space, but that it is defined with respect to a quotient

242

Matrix geometry of wavelet analysis. I

space. Indeed, it also follows from the above that, for any vector ψ in the Hilbert space of signals, dq d p dθ |U (θ, q, p)ψ|ψ|2 = ∞, G WH

so that there is no vector which is admissible with respect to the entire group. We shall return to the subject of building wavelet transforms on general quotient spaces in the next chapter (see Section 7.1.5). To complete this cycle of properties, we could again verify that the space Hψ of all Gabor transforms, corresponding to a particular window function ψ, is a closed subspace of L 2 (G WH , dq d p), which is also a Hilbert space with a reproducing kernel: K ψ (q, p; q , p ) =

1 ψq, p | ψq , p . 2π ψ2

(6.115)

This relation is the exact analog of (6.26). The complete decomposition of the space L 2 (G WH /+, dq d p), of all Gabor transforms, could be worked out in the manner of (6.41), with very little change in the derivation.

6.3.1.1

Holomorphic Gabor wavelets It is possible to construct spaces of holomorphic Gabor transforms and associated holomorphic gaborettes, in much the same way as we constructed holomorphic wavelet transforms in Section 6.2.1. We show below the existence of one such space and later we shall indicate how others may be obtained. However, in order to do so, it is first necessary to modify somewhat the definition of the Gabor transform. We proceed by first choosing the window function x2

ψ(x) = (π )− 4 e− 2 , 1

ψ2 = 1.

(6.116)

Next, we define the modified gaborettes, q

ψ(0,q, p) (x) = (U (0, q, p)ψ)(x) = (π )− 4 ei(x− 2 ) p e− 1

(x−q)2 2

qp

= ei 2 ψq, p

(6.117)

and use them to define the modified Gabor transform of a signal s: S(q, p) = ψ(0,q, p) | s = e−i

qp 2

S(q, p).

(6.118)

Then S obviously satisfies the finiteness of energy condition and moreover, the modified gaborettes also satisfy the resolution of the identity. Thus, signal reconstruction is possible using these modified gaborettes as well. Introducing the complex variable z = q − i p, it is easy to verify that |z|2

e− 4 S(q, p) = 1 π4 Writing

∞

−∞

z2

d x e− 2 (x−z) e 4 s(x) ≡ S(z). 1

2

(6.119)

243

6.3 The case of Gabor wavelets |z|2

e− 4 F(z), S(z) = √ 2π

(6.120)

we see that F(z) is an entire analytic function of z. These functions constitute the holomorphic Hilbert space Hhol = L 2 (C, dν(z, z)), where dν is the measure |z|2 dq d p dν(z, z) = e− 2 . 2π This Hilbert space is also a reproducing kernel Hilbert space, with the kernel given by zz

K hol (z, z ) = e 2 .

(6.121)

We call the ψ(0,q, p) holomorphic gaborettes and the functions F(z) holomorphic Gabor transforms. Other spaces of holomorphic Gabor transforms can now be similarly constructed by replacing the window function (6.116) by a Gaussian with standard deviation λ−1 , λ > 0 and using the representation (6.108) of the Weyl–Heisenberg group. Consider the window function !1 λ2 4 − λ2 x 2 ψ(x) = e 2 , ψ2 = 1, λ > 0, (6.122) π and define the generalized gaborettes λ2 ψ(0,q, p) (x) = (U (0, q, p)ψ)(x) = π λ

! 14

q

eiλ(x− 2 ) p e−

λ2 (x−q)2 2

,

(6.123)

where U λ is the unitary representation of the Weyl–Heisenberg group defined in (6.108). It can then be verified that the resolution of the identity λ dq d p |ψ(0,q, p) ψ(0,q, p) | = I, (6.124) 2π R2 holds. The generalized Gabor transform is now S(q, p) = ψ(0,q, p) | s,

s ∈ L 2 (G WH /+, dq d p),

(6.125)

which, by virtue of (6.124), enjoys the finiteness of energy condition 2π s2 . S2 = λ Introducing the complex variable √ p z = λq −i√ , λ (6.123) may be rewritten as: λ2 ψ(0,q, p) (x) = ψ(0,z) (x) = π

! 14

e−

λ|z|2 4

e

λz 2 4

√ − λ2 ( λx−z)2

The generalized Gabor transform now becomes

.

(6.126)

244

Matrix geometry of wavelet analysis. I

λ2 S(q, p) ≡ S(z) = π

! 14 e

− λ|z| 4

2

e

λz 2 4

∞

λ

d x e− 2 (

√ λx−z)2

s(x),

(6.127)

−∞

so that 2π F(z) = λ

! 12 e

λ|z|2 4

2

λz S(z) = e 4

∞

λ

√

d x e− 2 (

λx−z)2

s(x)

−∞

is the analytic Gabor transform and it then follows that 2 dq d p | S(q, p)| = dνλ (z, z) |F(z)|2 , G WH /+

C

where we have introduced the measure λ − λ|z|2 e 2 dq d p. dνλ (z, z) = 2π Thus, F is a vector in a Hilbert space of entire analytic functions, which we denote by Hλhol and note that this space of holomorphic Gabor transforms is a subspace of L 2 (C, dνλ ). The Hilbert space Hλhol has the reproducing kernel λ (z, z ) = e K hol

λzz 2

,

(6.128)

so that, for any F ∈ Hλhol , λ dνλ (z, z) K hol (z, z )F(z ) = F(z). C

Finally, the complex plane C is also a K¨ahler manifold, with potential function zz , 2 and invariant two-form (z, z) =

1 ∂ 2 (z, z) dz ∧ dz = dq ∧ d p, i ∂z ∂z so that, once again, we verify the relations (see (6.74)–(6.76)) (z, z) =

λ (z, z ) = eλ (z,z ) , K hol

and e−λ (z,z) = e−

λ|z|2 2

dq ∧ dp,

from which follows the measure dνλ . Thus, group theoretically, the analysis of Gabor wavelets using the Weyl–Heisenberg group runs entirely parallel to the analysis of the 1-D CWT using the affine group, except for the following two differences. First, here every vector in L 2 (R, d x) is admissible (the group theoretical reason being that G WH is a unimodular group). Second, the square integrability of the representation U (θ, q, p) is not on the entire Weyl–Heisenberg

245

6.3 The case of Gabor wavelets

group G WH itself, but on the quotient G WH /+ of the group by its center +. However here, as with the affine group, it is the square integrability of the representation in question which leads to the finiteness of the energy of the Gabor transform and enables us to reconstruct the signal from its transform.

6.3.2

Phase space considerations It is instructive to carry out a phase space analysis of the Weyl–Heisenberg group, for it sheds light on both the differences and the similarities between it and the affine group. In particular, it will clearly emerge why the phase space over which the Gabor transform is built is two-dimensional, although the group itself is a three-dimensional manifold. Let (θ, q, p) ∈ G WH and X ∈ gWH . We compute the adjoint action of the group on the Lie algebra, using (6.107) to get

x 0 = x 0 + px 1 − q x 2 ,

x 1 = x 1,

x 2 = x 2, where

X = Ad(θ,q, p) X = x 0 X 0 + x 1 X 1 + x 2 X 2 = (θ, q, p) [x 0 X 0 + x 1 X 1 + x 2 X 2 ] (θ, q, p)−1 . Representing X by the vector in R3 with components x 0 , x 1 , x 2 , the adjoint action is given by the matrix

1

p

M(θ, q, p) = 0

1

0

0

−q

0 . 1

(6.129)

Note that there is no dependence of M on the phase θ . The adjoint action on the dual vectors γ is effected by the matrices M * (θ, q, p), which are the transposed inverses of the matrices M(θ, q, p). Writing γ = M * (θ, q, p)γ , we get the dual transformation equations γ 0 = γ0 , γ1 = − pγ0 + γ1 , γ2 = qγ0 + γ2 . Thus, the orbits of vectors under the coadjoint action fall into two categories:

246

Matrix geometry of wavelet analysis. I

λ 0 3 (i) The orbits of vectors of ∈ R , 0 = , one for each λ = 0: 0 0 Oλ∗

λ = {γ = | x ∈ R2 }, x

(6.130)

which are planes orthogonal to the γ0 -axis. 0 ∈ R2 . These orbits are singletons consisting of the (ii) The orbits of vectors , λ λ 0 vector itself. We denote them by Oλ∗ and note that together they form a set λ of Lebesgue measure zero in R3 . Clearly, the union of all the orbits is the entire dual space, g∗WH , of the Lie algebra, now identified with R3 . The nontrivial orbits Oλ∗ , λ = 0, can each be identified with the quotient space G WH /+, since under the coadjoint action the phase subgroup + (consisting of elements of the type (θ, 0, 0)) is stable. These are the (two-dimensional) phase spaces of the problem and from the general theory of group representations it is known that each unitary irreducible representation of G WH is associated to one such orbit. As stated earlier, these representations U λ (θ, q, p) can be realized as in (6.108). The invariant two form under the coadjoint action on the orbits Oλ∗ is just dγ1 ∧ dγ2 and for λ = 1, this action is simple: γ1 → γ1 − p , γ2 → γ2 + q. The preceding discussion makes clear the mathematical sense in which the space of parameters of the Gabor transform S(q, p) is to be thought of as a phase space. The space of parameters is a coadjoint orbit of the group, which has the structure of a classical mechanical phase space.

7

Matrix geometry of wavelet analysis. II

The last chapter has already familiarized us with the use of group theoretical methods for the construction and analysis of wavelets and gaborettes. We aim in this chapter to first indicate the general applicability of these techniques and then to look at the case of the two-dimensional continuous transform, using the SIM(2) group. Later, we look at general matrix groups of the type that can be used for constructing other types of wavelet transforms in two dimensions. We shall be led, in this manner, to studying a class of semidirect product type groups, certain coadjoint orbits of which are isomorphic to the group itself. In all these cases, the common features of such a matrix-group analysis will be: (a) the group will refer to a set of possible symmetry transformations which the signal may undergo; (b) the space over which the signals are defined (as L 2 -functions) is intrinsic to the group; (c) the parameters in terms of which the wavelet transform is expressed are the parameters of the group itself, i.e., symmetry parameters of the signal, and (d) these parameter spaces, which arise as coadjoint orbits of the group, are also identifiable with phase spaces of signals. Referring back to the 2-D wavelet transform introduced in Chapter 2, we shall see that this transform is again related to a square integrable representation of a matrix group. The coadjoint orbit of this group (there is only one nontrivial orbit in this case) will again allow us to carry out a phase space analysis. As in the 1-D case, square integrability will enable us to obtain a resolution of the identity, lead to finiteness of the total energy of the wavelet transform and yield a reconstruction formula for the signal. In order to put the discussion in the context of a more general framework, we begin with a word about the choice of an appropriate group for building wavelet transforms and the general rationale for appealing to group theory in the first place.

7.1

A group-adapted wavelet analysis As mentioned earlier, a group-theoretical approach enables one to exploit mathematically the symmetries underlying the particular geometry which the signal space may have. If symmetries exist, it is natural to try to build these into the wavelet transform itself. This generally implies finding a continuous wavelet transform by exploiting a

247

248

Matrix geometry of wavelet analysis. II

representation of the group on some Hilbert space, the key ingredient required for such a construction being square integrability of the representation. Assume that the class of finite energy signals under consideration can be realized as functions on a manifold Y , i.e., s ∈ L 2 (Y, dµ) ≡ H. This manifold could in fact be the space of some parameters (e.g., frequency, time, position, etc.) of the signal. As examples, Y could be space Rn , the 2-sphere S 2 , space–time R × R or R × R2 , and so on. Assume that we measure the signals with the help of a probe η : s → η[s]. Usually such probes are taken to be linear functionals, representing the action of a measuring apparatus, a reference frame, etc. Mathematically, the measurement of the signal is given usually by some sort of an overlap integral that is, in the present case, an inner product s → ψ | s, with ψ representing the probe. Note that, if we were to restrict the signals to smooth functions on Y , measurements could also be represented by distributions of some type.

7.1.1

Some generalities Suppose there is a group G of symmetries of the signal, which acts as a set of transformations of the manifold Y . This means that any element g ∈ G, representing a specific symmetry operation on the parameter space (e.g., rotation of the signal through some angle, translation by some amount, etc.), induces a transformation of Y : y → g[y]. Successive transformations of this kind can be composed: g[g [y]] = gg [y] and the composite transformation gg is again an element of G. To each transformation, y = g[y], there exists an inverse transformation, y = g −1 [y ], g −1 ∈ G and in addition the identity transformation e for which e[y] = y, for all y ∈ Y is also a member of the group G. We assume further that this action is transitive, i.e., for any pair y, y ∈ Y , there is at least one g ∈ G such that g[y] = y . It should be noticed, however, that the transformation group G acting on Y is in general not unique, its choice may depend on the problem at hand. In the case of the 1-D CWT, for instance, we could choose between the full affine group G aff (a = 0) and its connected subgroup G + aff (a > 0). However, when talking about wavelets, the groups that we consider will always contain a dilation as a symmetry of the signal. Furthermore, in all the cases that will concern us here, the group action will be realized through matrices acting on a vector space into which Y can be embedded. Recall, this means that the set of matrices constituting this group is closed under multiplication, inverse taking and that the identity matrix is also a member of this set. The action of the group on Y induces an action on the signal space and there are two possible ways in which this could happen. (1) Action on the signals themselves. This is the active point of view: s → sg , in which one measures the transformed signal, which we denote by sg , with the fixed probe η. (2) Action on the probes. This is the passive point of view: η → ηg , in which one measures the fixed signal s with the transformed probe ηg . Now, if the manifold Y is globally G-invariant, consistency requires that the two points of view be equivalent,

249

7.1 A group-adapted wavelet analysis

ηg [sg ] = η[s], ∀ g ∈ G.

(7.1)

If now signals are taken to be vectors s in the Hilbert space H ≡ L 2 (Y, dν), where dν is some convenient measure, one naturally identifies probes with linear functionals on H, that is, η[s] ≡ η | s, for some fixed vector η ∈ H. Next, if one imposes that the action of G be linear, one ends up with a linear representation of G in H, sg = U (g)s, ηg = U (g)η. This means that to each g ∈ G one associates an operator U (g) on H, and the mapping g → U (g) is a group homomorphism (see (6.9)). The consistency condition (7.1) now requires that the U (g) be unitary operators, i.e., one has a unitary representation of G on the space H of signals: U (g)s | U (g)s = s | s , ∀ g ∈ G, s, s ∈ H

⇒

U (g)−1 = U (g)∗ .

(7.2)

Being a representation of G, the operators U (g) must also satisfy U (g1 )U (g2 ) = U (g1 g2 ),

U (g −1 ) = U (g)−1 ,

U (e) = I (= identity operator of L 2 (Y, dν)).

(7.3)

for all g1 , g2 in G and where e denotes the unit element of G (i.e., ge = eg = g, ∀ g). As an additional requirement, the representation U (g) will be assumed to be irreducible. Technically this means that for any nonzero vector ψ ∈ H, the set of vectors U (g)ψ, g ∈ G, span the Hilbert space. (At times a weaker condition, e.g., the existence of just one such nonzero vector ψ may be sufficient.) Irreducibility also means that H is a minimal space for realizing the symmetries unitarily. As a measure space, the group G has generally two invariant measures defined on it, a left Haar measure, dµ, invariant under g → g0 g, for fixed g0 ∈ G and a right Haar measure, dµr , invariant under g → gg0 . dµ(g0 g) = dµ(g) ,

dµr (gg0 ) = dµr (g),

g ∈ G.

The existence of these measures is a general property of all topological groups; for the affine group these were explicitly written down as dµ = db db/a 2 and dµr = db da/a (see (6.11) and the discussion following). The two invariant measures on G are equivalent, but generally not equal. In particular, for all wavelet related groups, which include dilations, the two Haar measures are different. We shall usually work with the left Haar measure, although everything we do could just as well be done using the right Haar measure. The space L 2 (G, dµ) is usually taken to be the Hilbert space of all wavelet transforms and it will turn out that finding wavelet transforms of signals implies mapping the signal space L 2 (Y, dν) isometrically into a (closed) subspace of L 2 (G, dµ). Given this setting, one may derive a wavelet analysis on Y , adapted to the symmetry group G, in which the wavelet transform of a signal s would be an associated function S, defined on the group (the group being identified with a phase space for the signal) or some other phase space related to the group. In case the group itself can be identified with a phase space (as was the situation with the affine group or as it will be for the 2-D

250

Matrix geometry of wavelet analysis. II

wavelet transform discussed below), the obvious choice of the function S is to write it as S(g) = U (g)η | s. Two questions immediately arise: first, does the transform S(g) have finite energy and second, is it possible to reconstruct the signal s uniquely from its transform S?

7.1.2

Square integrability of representations The first question posed above can be reformulated as follows. If the total energy of the transform S is identified with the square of its L 2 -norm: 2 E(S) = S = dµ(g) |S(g)|2 , G

then of course, we want S < ∞. In fact we require that the mapping s → S be (up to a constant) an isometry. Thus, the original question about the finiteness of energy becomes: is it possible to find ψ ∈ H such that S is finite for all s or that the map s → S be a multiple of the isometry? It will turn out that a positive answer to this question will also guarantee the possibility of reconstructing the signal from its transform, i.e., a positive answer to the second question. Interestingly enough, from a mathematical point of view, in order to obtain finiteness of energy for all transforms S, it is enough to require this of the transform of the probe ψ only. Thus, we require the existence of a nonzero vector ψ ∈ H such that I (ψ) = dµ(g) |U (g)ψ | ψ|2 < ∞ . (7.4) G

In case such a vector exists, we call it an admissible vector or, in signal analytic language, a generalized mother wavelet. The representation U (g) is then said to be square integrable. Square integrability is a property of both the representation U (g) and of the group G itself. Not all groups have square integrable representations and the same group may have representations which are square integrable as well as other ones which are not. From the general theory of square integrable group representations (see, e.g., [Ali00] for a detailed account) one knows the following: r The existence of one admissible vector guarantees the existence of a dense set of such vectors. In particular, if ψ is admissible, then so also are all the vectors U (g)ψ, g ∈ G. Let us denote the set of all admissible vectors by A. Then there exists an operator C, in general unbounded, on the Hilbert space H such that it is self-adjoint, has positive spectrum and such that its domain is precisely the set of all admissible vectors: cψ ≡ Cψ2 < ∞

⇔

ψ ∈ A.

(7.5)

Moreover, C −1 also exists as a densely defined positive (spectrum) operator. As in the 1-D wavelet case (6.17), we call C the Duflo–Moore operator of the representation. r If the group is nonunimodular (i.e., the left and right Haar measures are different) then C is an unbounded operator and A is a proper subset of H. If G is unimodular,

251

7.1 A group-adapted wavelet analysis

then C = λI, λ > 0, and A coincides with the entire Hilbert space H (every vector is an admissible vector). r For any two admissible vectors ψ, ψ and arbitrary vectors s, s ∈ H, the following orthogonality relation holds: dµ(g) U (g)ψ | s U (g)ψ | s = Cψ | Cψ s | s. (7.6) G

Note that (6.37) is just a special case of this relation, when G is the affine group. Proceeding in the same way by which we arrived at (6.38) from (6.37), we derive from (7.6) the resolution of the identity, 1 dµ(g) |U (g)ψ U (g)ψ| = I, (7.7) Cψ | Cψ G provided Cψ | Cψ = 0. Taking ψ = ψ in the above gives, 1 dµ(g) |ψg ψg | = I. cψ G

7.1.3

(7.8)

Construction of generalized wavelet transforms Given the existence of a square integrable representation U (g) of the group G, on the signal space H = L 2 (Y, dν), (generalized) wavelets and wavelet transforms can be constructed by exploiting the orthogonality relation (7.6), in much the same way as was done for the affine group in the previous chapter. We start by taking a mother wavelet ψ ∈ H and defining generalized wavelets as the vectors ψg = U (g)ψ ∈ H, g ∈ G. In the physical literature, the vectors ψg are called coherent states of the representation U (g). Our convention will be to call these vectors generalized wavelets only when the group contains some sort of a dilation transformation on the space Y and the group space itself can be identified with a phase space. Otherwise we shall use the term coherent state. The question of when the group also has the structure of a phase space will be analyzed later. Using the generalized wavelets ψg , the (generalized) wavelet transform of the signal s ∈ H is defined to be the function S(g) = ψg | s on G. Since the group G is derived by analyzing symmetry transformations of signals, its elements g are defined in terms of these very symmetry parameters (e.g., rotation angle, translation distance, zoom factor, etc.) and hence the wavelet transform also becomes a function of these parameters. Computing now the expectation value of both sides of (7.8) with respect to s ∈ H yields 2 S L 2 (G,dµ) = dµ |S(g)|2 = cψ s2H, (7.9) G

implying that the map Wψ : H → L 2 (G, dµ), given by −1/2

(Wψ s)(g) = cψ

ψg | s,

(7.10)

252

Matrix geometry of wavelet analysis. II

is a linear isometry. Similarly, acting on a signal s with the operators appearing on both sides of (7.7), we derive the general reconstruction formula 1 s= dµ(g) S(g)ψg . (7.11) Cψ | Cψ G Again, it ought to be emphasized here that the wavelet transform S(g), appearing in this expression, is computed using the wavelets ψg , while the reconstruction is done using the different set ψg . Of course, the formula is valid also if ψ = ψ . The above discussion again illustrates how in the general situation, as in the case of the affine group discussed in the last chapter, it is the square integrability of the representation which leads to the finiteness of the energy of the transform, on the one hand, and to the reconstruction formula, on the other. The orthogonality relation (7.6) acquires a more transparent physical meaning if we express it in terms of wavelet transforms. If Sψ (g) is the wavelet transform of the signal s, with respect to the mother wavelet ψ and Sψ (g) the transform of s with respect to ψ , then (7.6) can be rewritten as, dµ(g) Sψ (g) Sψ (g) = Cψ | Cψ s | s. (7.12) G

In other words, wavelet transforms of orthogonal signals are always orthogonal, independent of the mother wavelets chosen to represent them, while wavelet transforms of arbitrary signals, when computed with respect to a mother wavelet ψ, are all orthogonal to their transforms computed with respect to ψ , if Cψ is orthogonal to Cψ .

7.1.4

Reproducing kernels, partial isometries and localization operators Let ψ be an admissible vector and let Wψ [L 2 (Y, dν)] ≡ Hψ be the range of the isometry Wψ . This means that Hψ is a closed Hilbert subspace of L 2 (G, dµ), consisting of all wavelet transforms S(g) associated to the mother wavelet ψ. Let Pψ be the projection operator onto Hψ , i.e., Pψ L 2 (G, dµ) = Hψ . If ψ and ψ are two admissible vectors, chosen so that Cψ | Cψ = 0, then from (7.6) we infer that the corresponding spaces of wavelet transforms, Hψ and Hψ , are orthogonal. However, unlike in the case of the affine group (see Theorem 6.1.1), this fact cannot in general be used to obtain a complete decomposition of L 2 (G, dµ) into orthogonal subspaces of wavelet transforms – generally the space L 2 (G, dµ) contains more than just wavelet transforms. An interesting feature, connecting the spaces of wavelet transforms Hψ corresponding to different mother wavelets ψ, now emerges. All these spaces of transforms sit inside L 2 (G, dµ) and the same signal s can be mapped into different transform spaces, by choosing different mother wavelets. It is natural to ask whether it is possible to interpolate between these spaces of transforms, for that would allow us to take the transform of a signal with respect to one mother wavelet and re-express it as a transform with respect to a different mother wavelet. The following discussion, in particular Theorem 7.1.1, will show exactly how this can be done.

253

7.1 A group-adapted wavelet analysis

7.1.4.1

Partial isometries For two arbitrary mother wavelets ψ, ψ , define the function K ψ,ψ : G × G → C,

K ψ,ψ (g, g ) = (cψ cψ )−1/2 ψg | ψg .

(7.13)

This function has the easily verifiable properties, K ψ,ψ (g, g ) = K ψ ,ψ (g , g) , dµ(g

) K ψ,ψ

(g, g

) K ψ

,ψ (g

, g ) = K ψ,ψ (g, g ),

(7.14)

G

the first following from the definition of K ψ,ψ in (7.13) and the second from the resolution of the identity in (7.8). Next let us define an integral operator Vψ,ψ , with K ψ,ψ (g, g ) as its kernel: dµ(g ) K ψ,ψ (g, g ) F(g ), F ∈ L 2 (G, dµ). (7.15) (Vψ,ψ F)(g) = G

We show below that K ψ ≡ K ψ,ψ is the integral kernel of the projection operator Pψ and hence defines a reproducing kernel of the type seen in the last chapter (see (6.28)– (6.30)), while the operator Vψ,ψ , for different ψ, ψ , is a partial isometry on L 2 (G, dµ). This means that the range of Vψ,ψ is the space Hψ , of wavelet transforms corresponding to the mother wavelet ψ , while its kernel is the orthogonal complement, H⊥ ψ , of the space of transforms Hψ . Between Hψ and Hψ the operator interpolates as a linear isometry. Theorem 7.1.1 . The operator Vψ,ψ is a partial isometry on L 2 (G, dµ), which maps the subspace Hψ isometrically onto Hψ , and has the properties Vψ,ψ Vψ ,ψ

= Vψ,ψ

,

∗ Vψ,ψ

= Vψ ,ψ ,

∗ Vψ,ψ Vψ,ψ

Vψ,ψ = Pψ ,

= Pψ ,

∗ Vψ,ψ

Vψ,ψ

(7.16) = Pψ .

(7.17)

Proof . The two relations in (7.16) follow from (7.14). For any F ∈ L 2 (G, dµ), it can be shown that dµ(g ) K ψ (g, g ) F(g ), (7.18) (Pψ F)(g) = G

in exactly the same way as (6.33) was proved for the affine group. This establishes the first of the relations in (7.17). Also,

−1/2 dµ(g ) K ψ,ψ (g, g ) F(g ) = (cψ cψ ) ψg | dµ(g ) ψg F(g ) G

where s = (cψ cψ )−1/2

= ψg | s, G

dµ(g) ψg F(g) ∈ L 2 (Y, dν).

G

254

Matrix geometry of wavelet analysis. II

Thus, Vψ,ψ F is the wavelet transform of s with respect to the mother wavelet ψ, implying that the range of Vψ,ψ is contained in the subspace Hψ ⊂ L 2 (G, dµ). To see that it actually coincides with Hψ , let F be an arbitrary element of Hψ . Then there exists a signal s ∈ L 2 (Y, dν) for which F(g) = ψg | s. But, by (7.8), 1 dµ(g ) ψg | ψg ψg | s ψg | s = cψ G = dµ(g ) K ψ,ψ (g, g )F (g ) = (Vψ,ψ F )(g), G

where we have written 1/2 cψ

ψg | s . F (g) = cψ Thus every vector in Hψ is also in the range of Vψ,ψ . On the other hand, using (7.14) and (7.18), dµ(g ) K ψ,ψ (g, g ) F(g ) G dµ(g ) dµ(g

) K ψ,ψ (g, g

) K ψ (g

, g ) F(g ) = G G dµ(g

) K ψ,ψ (g, g

) (Pψ F)(g

), = G

implying that, for any vector F ∈ H⊥ ψ , Vψ,ψ F = 0.

Thus H⊥ ψ is contained in the kernel of Vψ,ψ . Finally, if F ∈ Hψ , so that F (g) = ψg | s, for some signal vector s ∈ L 2 (Y, dν), and F 2L 2 (G,dµ) = cψ s2L 2 (Y,dν) , then

−1/2 (Vψ,ψ F )(g) = (cψ cψ ) dµ(g ) ψg | ψg ψg | s

=

c cψ

ψ

1/2

G

ψg | s,

by (7.8).

But the function F(g) = ψg | s is the wavelet transform of s with respect to the mother wavelet ψ. Hence, cψ s2L 2 (Y,dν) = cψ s2L 2 (Y,dν) = F 2L 2 (G,dµ) . Vψ,ψ F 2L 2 (G,dµ) = cψ cψ Thus, Vψ,ψ maps Hψ isometrically onto Hψ and its kernel coincides with H⊥ ψ .

This theorem is useful for it shows, first of all, that the various transforms that can be obtained for the same signal but using different mother wavelets, are in a sense equivalent. Second, all these transforms have the same energy content as well as the same information content. From a theoretical point of view, therefore, it does not matter

255

7.1 A group-adapted wavelet analysis

which mother wavelet is used. However, for practical or computational reasons, it may be preferable to use mother wavelets having specific properties.

7.1.4.2

Left regular representation and localization operators We had started out by choosing an irreducible representation U (g) of the group G, on the space of signals L 2 (Y, dν). Assuming U to be square integrable and taking different admissible vectors (or mother wavelets) ψ, we defined generalized wavelets ψg , using which we were able to map the signal space isometrically into subspaces Hψ of signal transforms. The isometric mapping Wψ was defined (see (7.10)) by (Wψ s)(g) = −1/2 cψ ψg | s. As a consequence of this isometry, the unitary operators U (g) are mapped to certain unitary operators on Hψ and it is interesting to obtain these image operators. Since, for arbitrary s ∈ L 2 (Y, dν) and fixed g0 ∈ G, −1/2

ψg | U (g0 )s = cψ

−1/2

U (g0−1 g)ψ | s = cψ

(Wψ U (g0 )s)(g) = cψ = cψ

−1/2

U (g0 )∗ U (g)ψ | s

−1/2

ψg0−1 g | s,

we obtain, Wψ U (g0 )Wψ−1 = Uψ (g0 ) (the inverse of Wψ being computed on its range), where, for any S ∈ Hψ , (Uψ (g0 )S)(g) = S(g0−1 g),

g ∈ G,

(7.19)

which ought to be compared with (6.35). Being the isometric image of a unitary irreducible representation, the representation given by these operators, on Hψ , is also unitary and irreducible. In other words, U (g) and Uψ (g) are equivalent representations of the group G, however, expressed on different Hilbert spaces. What is important to note here is that the right-hand side of (7.19) is independent of ψ. In other words, one obtains the same form for the transformed operators Uψ (g) on all the subspaces Hψ ⊂ L 2 (G, dµ), irrespective of the mother wavelet chosen. Using this fact let us define the operators U (g), g ∈ G, on the whole of L 2 (G, dµ), adopting this same form: (U (g)F)(g ) = F(g −1 g ), Since

U (g)F2 =

g, g ∈ G,

F ∈ L 2 (G, dµ).

(7.20)

dµ(g )|(U (g)F)(g )|2 G

dµ(g ) |F(g −1 g )|2 = F2 ,

= G

the last step following from the invariance of the measure dµ, the operators U (g) are unitary on L 2 (G, dµ). Also, clearly, the map g → U (g) is a group homomorphism. Thus, we have obtained a unitary representation of G in terms of these operators, called the left regular representation. The situation is exactly what we also had in the case of the affine group (see (6.34) and the discussion leading up to and following it). The left regular representation is not in general irreducible, and indeed, (7.19) shows that, restricted to each one of the subspaces Hψ ⊂ L 2 (G, dµ), it admits the irreducible subrepresentation Uψ (g). Another way to express this fact is to write

256

Matrix geometry of wavelet analysis. II

Wψ U (g) = U (g)Wψ = Uψ (g).

(7.21)

Summarizing, the rˆole of the representation U (g) on the signal space is played by the left regular representation U (g) on the spaces of wavelet transforms. The construction of localization operators on the group space G follows the same pattern as laid out for the affine group in Section 6.1.3. Let ψ ∈ L 2 (Y, dν) be an admissible vector, K ψ (g, g ) the corresponding reproducing kernel and ⊂ G a measurable set (with respect to the Haar measure dµ). Associated to , we define the integral kernel, aψ (g, g ) = dµ(g

) K ψ (g, g

) K ψ (g

, g ), (7.22)

and the resulting operator aψ ( ) on L 2 (G, dµ), (aψ ( )F)(g) = dµ(g ) aψ (g, g ) F(g ), F ∈ L 2 (G, dµ).

(7.23)

From (7.18) and the definition of aψ ( ), it follows that F | aψ ( )F = dµ(g) |(Pψ F)(g)|2 ,

(7.24)

G

which shows that the operator is bounded, self-adjoint and has positive spectrum. In particular, for S ∈ Hψ , S | aψ ( )S = dµ(g) |S(g)|2 .

Thus, the quantity p S ( ) =

S | aψ ( )S , S2

(7.25)

represents the fraction of the wavelet transform of s (where S(g) = ψg | s) which is localized in the region . This also motivates the term localization operator for aψ ( ). Additionally, these operators have the measure theoretical properties as those obtained for the localization operators of the affine group (see (6.53)): aψ (G) = Pψ , aψ (∅) = 0, 8 7

j = ∅, whenever i = j, aψ ( i ), if i aψ ( i ) = i∈J

(7.26)

i∈J

where, again, the sum in (7.26) is to be understood in the sense of scalar products. These relations also imply that p S ( ) is a probability measure. Finally, the localization operators satisfy the imprimitivity or covariance condition, U (g) aψ ( ) U (g)∗ = aψ (g ), where g is the shifted set,

(7.27)

257

7.1 A group-adapted wavelet analysis

g = {gg ∈ G | g ∈ }. The above covariance relation can be derived in the same way as (6.57) and again leads to the corresponding relation, p S (g ) = pU (g)−1 S ( ),

(7.28)

for the probability measure p S ( ). We interpret this relation in the same way as in the case of the affine group namely, that the probability of localization in the transformed set is the same as the probability of localization of the transformed signal in the original set.

7.1.5

Wavelet transforms on general quotient spaces The group theoretical analysis outlined above was intended to underscore the fact that, using purely symmetry arguments, one can arrive at a general wavelet transform, which then displays all the basic properties of the standard 1-D wavelet transform. However, the power of this general group theoretical analysis lies in its applicability to a vast number of other symmetry groups, thus opening up the possibility of constructing extremely general classes of wavelet transforms. Many of these turn out to be of enormous practical value, as well. It is in this light that we will undertake in Section 7.2 a general analysis of the 2-D wavelet transform. Building the generalized wavelet transform, defined in (7.10), depended on the assumption that the underlying group representation U (g) was square integrable – in the sense that there existed a vector ψ satisfying the admissibility condition (7.4). Furthermore, since we generally wish to identify the space of signal variables with the structure of a physical phase space, the group G itself would have to possess such a structure, if wavelet transforms are to be defined as functions over it. However, already in the case of the Gabor transform, we saw that these conditions were not fulfilled in the strict sense. Indeed in that case admissibility was only defined with respect to a quotient space of the group (see (6.114) and the discussion following it); it was this quotient space which had the structure of a phase space and on which the Gabor transform was defined. Let us briefly indicate here how this sort of a construction can be put against a more general setting. As a first case, consider the situation where the analyzing wavelet ψ has a nontrivial isotropy subgroup Hψ ⊂ G, up to a phase. This means that ψ satisfies the condition U (h)ψ = eiα(h) ψ,

h ∈ Hψ ,

(7.29)

where α(h) is a (real) phase factor, generally depending on h. (In the physical literature, this is the setting for the construction of Gilmore–Perelomov type of coherent states [191,192,Per86,312].) Clearly now, the integrand in (7.4) does not depend on g, but only on the coset g Hψ ∈ G/Hψ , so that the finiteness of the integral would force the subgroup Hψ to be compact. Failing that, we assume that the quotient space X = G/Hψ

258

Matrix geometry of wavelet analysis. II

carries an invariant measure ν. (Recall that elements of the quotient space X are the cosets g Hψ , g ∈ G, and they transform under the action of an element g0 of the group in the manner, g Hψ → g0 g Hψ .) We impose the weaker admissibility condition [191,192,312], dν(x) |U (g)ψ|φ|2 < ∞, ∀ φ ∈ H (x ≡ g Hψ ), (7.30) X

on ψ. The integrand in (7.30) manifestly does not depend on individual elements g ∈ G, only on their cosets modulo Hψ , x ≡ g Hψ ∈ G/Hψ , and the condition (7.30) makes sense. This condition means that the representation U is square integrable on the coset space X = G/Hψ or, as it is called, square integrable modulo the subgroup Hψ . Notice that the latter need no longer be compact. In order to define wavelets, it is necessary to go back to the group. We do this using the notion of a section. This is a map σ : X → G, chosen so that if σ (x) = g then x = g Hψ . In these terms, the admissibility condition (7.30) may be rewritten in the slightly different, but completely equivalent form: c X (ψ, φ) = dν(x) |U (σ (x))ψ|φ|2 < ∞, ∀ φ ∈ H, (7.31) X

where σ is an arbitrary section σ : X → G. Indeed, since two different sections σ and σ are related as σ (x) = σ (x)h(x), where h(x) ∈ Hψ , it is obvious that the integrand does not depend on the choice of the section. Correspondingly, the wavelet vectors are written as ψσ (x) = U (σ (x))ψ, x ∈ X , which emphasizes that the proper index set is X = G/Hψ and not G. Under the condition (7.30) or (7.31), the whole construction may be performed exactly as before [Ali00,6]. In particular, the map Wψ : H → L 2 (X, dν) given by −1/2 (Wψ s)(x) ≡ c X ψσ (x) |s is an isometry, where c X ≡ c X (ψ, ψ); in other words, one has a resolution of the identity c−1 dν(x) |ψσ (x) ψσ (x) | = I. (7.32) X X

From this follows, as before, that the range of Wψ is a closed subspace Hψ of L 2 (X, dν), the corresponding projection Pψ = Wψ Wψ∗ is an integral operator with (reproducing) kernel K (x , x) = c−1 X ψσ (x ) |ψσ (x) , the familiar reconstruction formula holds, etc. Coming back to the subject of this book, it is true that the continuous wavelet transform, both in one and two dimensions, are examples of wavelet transforms living directly on the associated group, G (+) aff and SIM(2), respectively (see below). However, we have also seen in Section 6.3 that the Gabor transform is an example of a construction modulo a subgroup, in this case the phase subgroup + of the Weyl–Heisenberg group G WH . In the same way, in dimensions higher than 2, the CWT with respect to an axisymmetric wavelet leads to wavelet transforms defined on a quotient of the above type (see Section 9.1). Physically, this means that while the total set of signal symmetries may be large,

259

7.2 The 2-D continuous wavelet transform

because of the needs of the problem at hand, the wavelet transform is defined over a smaller set of parameters. Moreover, it is this quotient space which turns out to be the relevant phase space of the problem. Actually, one can go a step further, and extend the whole construction to the case of an arbitrary coset space X = G/H , where H is not the stability subspace of any vector ψ in the sense of (7.29). The main difference is that (i) the validity of the admissibility condition (7.31) may depend on the choice of the section σ ; and (ii) when the condition holds, it reads 0< dν(x) |U (σ (x))ψ|φ|2 = φ | Aσ φ , ∀ φ ∈ H, (7.33) X

where Aσ is a bounded positive invertible operator, sometimes called the resolution operator. Equivalently, the resolution of the identity (7.32) becomes dν(x) |ψσ (x) ψσ (x) | = Aσ . (7.34) c−1 X X

Note that A−1 σ may be unbounded in general. In the case where it is bounded, the system of wavelets {ψσ (x) , x ∈ X } is called a (continuous) frame. (The unbounded case yields a far reaching generalization of the notion of a frame discussed in Section 2.4.1.) However, this extension of the theory of wavelets will not concern us in this book, with the sole exception of wavelets on the 2-sphere, discussed at length in Section 9.2.

7.2

The 2-D continuous wavelet transform Group theoretically, we expect the 2-D continuous wavelet transform to arise from a group which should be an appropriate generalization of the affine group. This indeed is the case, and we are led to it by a relatively straightforward analysis of the symmetries which might be attributed to a two-dimensional signal. It turns out that the group in question, the 2-D similitude group, is a generalization of the affine group and in fact contains it as a subgroup.

7.2.1

The similitude group and 2-D wavelets We begin with a model of a two-dimensional image. For our purposes, a 2-D image will be a finite energy signal s ∈ L 2 (R2 , d 2 x), as discussed in Section 2.1.1. The operations we want to apply to s are translations in the image plane (b ∈ R2 ), global dilations (zooming in and out by a > 0) and rotations around the origin (θ ∈ [0, 2π )). Together these transformations constitute a four-parameter group, called the similitude group of the plane and denoted by SIM(2). The action on the plane is a, θ )y = arθ y + b, x = (b, where rθ is the 2 × 2 rotation matrix,

(7.35)

260

Matrix geometry of wavelet analysis. II

rθ =

cos θ sin θ

− sin θ cos θ

.

(7.36)

Our convention is to take a quantity such as x to be a column vector, the corresponding a, θ ) row vector being x T . A convenient representation of the joint transformation (b, is in the form of 3 × 3 matrices b ar θ a, θ ) ≡ , 0 T = (0, 0). (7.37) (b, 0 T 1 Matrix multiplication then replicates the composition of successive transformations and thus we derive the group law, a, θ )(b , a , θ ) = (b + arθ b , aa , θ + θ ) (b, 1, 0), (unit element) e = (0, a, θ)−1 = (−a −1r−θ b, a −1 , −θ). (b, From this, we deduce the following. r The set of rotations (0, 1, θ), θ ∈ [0, 2π), is a subgroup of SIM(2) and so also is the set of dilations (0, a, 0), a > 0. Moreover, these two subgroups commute, i.e., (0, 1, θ)(0, a, 0) = (0, a, 0)(0, 1, θ ) = (0, a, θ ). r The set of all translations (b, 1, 0), b ∈ R2 , is also a subgroup. Moreover, it has the a, θ) ∈ SIM(2) is structure of an invariant subgroup in the following sense: if (b, arbitrary and (b0 , 1, 0) any element of the translation subgroup, then a, θ)−1 = (arθ b0 , 1, 0), a, θ )(b0 , 1, 0)(b, (b, which again is an element of the same subgroup. Thus, the similitude group SIM(2) has the structure of a semidirect product: SIM(2) = R2 (R+ ∗ × SO(2)) where R2 is the subgroup of translations, R+ ∗ that of dilations, and SO(2) of rotations. Topologically, we can identify R2 with C, the complex plane and R+ ∗ × SO(2) with ∗ C , the complex plane with the origin removed. Thus we may write SIM(2) = C C∗ , and denoting a group element by (z, w), where z ∈ C and w ∈ C∗ , the group composition law is, very simply, (z 1 , w1 )(z 2 , w2 ) = (z 1 + w1 z 2 , w1 w2 ). In particular, if we only consider elements (z, w) for which z = b + ic, with c = 0, and w = aeiθ , with θ = 0, then these elements clearly constitute a subgroup, which is just the affine group of the line. Thus, G aff ⊂ SIM(2), meaning that the similitude group

261

7.2 The 2-D continuous wavelet transform

is a generalization of the affine group, as indicated earlier. In fact, we may consider SIM(2) as being a complexification of G aff . Let us compute next the left (invariant) Haar measure on SIM(2). If (b0 , a0 , θ0 ) is a a, θ ) arbitrary, then writing fixed element of the group and (b, a0 a, θ0 + θ ), a, θ) = (b0 + a0rθ0 b, (b , a , θ ) = (b0 , a0 , θ0 )(b, and noting that det [rθ0 ] = 1, we get d 2 b = a02 d 2 b,

da = a0 da,

dθ = dθ.

Thus, the measure 1 2 (7.38) d b da dθ, a3 is invariant under left transformations. Similarly, the right Haar measure can be computed to be

a, θ) = dµ(b,

a, θ ) = 1 d 2 b da dθ, dµr (b, a and thus like the affine group, the SIM(2) group is also nonunimodular. a, θ represent parameters in terms of which we want to analyze the signals Since b, 2 s ∈ L (R2 , d 2 x), we shall identify the Hilbert space L 2 (SIM(2), dµ) with the space of 1 all finite energy 2-D wavelet transforms. It will turn out that SIM(2) R2 × R+ ∗ ×S again has the structure of a phase space (S 1 being the unit circle). We shall later analyze the orbits of SIM(2) under the coadjoint action and we shall see that there is only one nontrivial orbit, which topologically is isomorphic to the group itself. Correspondingly, there is only one nontrivial unitary irreducible representation of SIM(2). This representation, which is a straightforward realization of the action (7.35) on the space of signals a, θ ) (see (2.13)): L 2 (R2 , d 2 x), is given by the operators U (b, a, θ)s ( x − b)), b ∈ R2 , a > 0, 0 θ < 2π. (7.39) U (b, x ) = a −1 s(a −1 r−θ ( The fact that these operators define a unitary representation is straightforward to verify. 2 , d 2 k) of Fourier-transformed signals In the space L 2 (R s, this representation acquires the form (see (2.14)) (b, a, θ] = a e−i b· k U s (k) s(ar−θ k). (7.40) This representation is also square integrable, with an admissibility condition on mother wavelets, see (2.16), which is analogous to (6.15). Indeed we have the result: a, θ) defines a unitary irreducible repTheorem 7.2.1 . The family of operators U (b, resentation of SIM(2) in the Hilbert space L 2 (R2 , d 2 x), which is unique up to unitary equivalence. This representation is square integrable, and a vector ψ ∈ L 2 (R2 , d 2 x) is admissible if, and only if, it verifies the condition:

262

Matrix geometry of wavelet analysis. II

cψ ≡ (2π)

2 R2

d 2 k 2 |ψ(k)| < ∞. 2 |k|

(7.41)

Proof . That U is a representation of SIM(2) results from explicit computation; its unitarity is obvious and its irreducibility follows from Proposition 2.1.2. As for the square integrability, it is proved by a direct calculation, using the Fourier-transformed (see above), realization U + +2 + | ψ ++ d 2 b da dθ I (ψ) ≡ + U (b, a, θ ) ψ a3 SIM(2) ∞ da 2π 2 k) ψ( r−θ (k)) d b dθ d 2 k ei b·k ψ(a = a 0 0 R2 R2 × d 2 k e−i bk ψ(ar −θ (k )) ψ(k ). R2

Integrating first over b (the permutation of integrals is allowed by Fubini’s theorem) yields a factor (2π )2 δ(k − k ) and, therefore, ∞ + +2 + +2 da 2π + k) ++ ++ ψ( ++ . I (ψ) = (2π)2 r−θ k) dθ d 2 k + ψ(a a 0 0 From this, we get, exactly as in the proof of Proposition 2.2.1, I (ψ) = cψ ψ 2 , with cψ given by (7.41), which proves the statement.

in the Fourier transformed space: Introducing the Duflo–Moore operator C, ψ 2, cψ = C we obtain, ψ)( k) = (C

2π ψ(k), |k|

(7.42)

which should be compared to (6.16). Thus every function ψ ∈ L 2 (R2 , d 2 x), such that (i.e., satisfies (7.41)), is an admissible its Fourier transform lies in the domain of C vector and can be used to build wavelets. Following our established practice, we shall call such vectors mother wavelets. Choosing a mother wavelet ψ, we define the 2-D wavelets as: 1 1 ψb,a,θ r−θ ( ( x ) = U (b, a, θ)ψ ( x) = ψ x − b) , (7.43) a a

263

7.2 The 2-D continuous wavelet transform

and the 2-D continuous wavelet transform as the inner product of the signal s with the wavelet ψb,a,θ : a, θ) = ψb,a,θ | s, S(b,

(7.44)

which is a function on SIM(2) (see (2.18)–(2.20)). All the general properties of wavelets, as outlined in Section 7.1 and in particular the relations (7.6)–(7.11) and (7.21)–(7.27), follow in a straightforward manner. Some of these were worked out in detail in Section 2.2. Specific examples of 2-D wavelets with special symmetry properties have also been worked out in Chapter 2. For the sake of illustration, we display here the general resolution of the identity and reconstruction formula for signals. Following (7.7) we may write, ∞ ∞ ∞ 2π 1 da

db1 db2 3 dθ |ψb,a,θ ψb,a,θ | = I, (7.45)

Cψ | Cψ −∞ −∞ 0 0 a provided, (b1 , b2 being the components of the vector b), Cψ | Cψ = (2π )2

R2

d 2 k

k) (k) ψ ψ( = 0. 2 |k|

From this we obtain the reconstruction formula for a signal, ∞ ∞ ∞ 2π 1 da a, θ)ψ , db1 db2 3 dθ Sψ (b, s= b,a,θ Cψ | Cψ −∞ −∞ 0 0 a

(7.46)

a, θ) = ψb,a,θ where Sψ (b, | s is the wavelet transform of s in terms of the mother wavelet ψ. One ought to comment here on the freedom that one has in designing the 2-D wavelet transform. On the one hand, one may ignore the rotation variable θ, for instance, if directions are irrelevant. This is achieved by choosing an isotropic or rotation invariant wavelet, ψ(rθ ( x )) = ψ( x ). Equivalently, one may consider as transformations of the plane only translations and dilations, with the corresponding group R2 R+ ∗ . In this case, however, the representation structure is much more complicated, since every a, θ ), subspace of the form L 2 (C, d 2 k) is invariant under the action of the operators U (b, where C is a cone with apex at the origin in the k-plane. More interesting is the opposite move. If, besides the similitude operations, one considers also certain types of deformations, one gets a larger group, namely, the group obtained by replacing in (7.37) the matrix arθ ∈ R+ ∗ × SO(2) by an arbitrary nonsingular 2 × 2 real matrix. This group is much more complicated and so also is its representation structure. In fact, it is unlikely that any of its representations would be of much use for our purposes. Thus, the similitude group SIM(2) seems to occupy a privileged position in the construction, although later, in Section 7.4, we shall study a second group in which the spatial rotations rθ are replaced by hyperbolic rotations.

264

Matrix geometry of wavelet analysis. II

7.2.2

The group as the primary object We started out by defining a two-dimensional image as a function on R2 and then obtained the group SIM(2) by considering a set of transformations on R2 (see (7.35)), which would lead to the physically desirable transformations (2.7)–(2.9) on the signal space. It is possible to reverse the argument, i.e., to start with the group SIM(2) as the primary object and then to obtain the space R2 , over which the images are to be defined, as intrinsic to the group and on which it has the natural action given by (7.35). To see this, note first that the set of matrices in SIM(2), arθ 0 , a > 0, θ ∈ [0, 2π ), 0 T 1 constitute the subgroup H of rotations and dilations and since, I2 b arθ b arθ 0 , = 0 T 1 0 T 1 0 T 1

(7.47)

the set of matrices I2 y , y ∈ R2 , 0 T 1 is identifiable with the quotient space SIM(2)/H . Acting on such a matrix from the left, by an element of SIM(2), we see that, I2 y arθ b arθ arθ y + b = 1 0 T 1 0 T 1 0 T =

I2 0 T

arθ y + b 1

arθ

0

0 T

1

,

implying the transformation y → arθ y + b on R2 . Thus, the action of SIM(2) on the quotient space SIM(2)/H is the same as its action (7.35) on R2 . The situation here is the same as that encountered in the case of the affine group (see (6.10) and the discussion following). One can just as well adopt the point of view that the group is the basic geometrical object, from which signals, their transformation properties and their representations in various spaces of functions, all follow as mathematical consequences.

7.2.3

Decomposition theory of 2-D wavelet transforms As in the case of the affine group, we would like to identify the space L 2 (SIM(2), dµ) with the set of all finite energy 2-D wavelet transforms. In order to do this, we have to be able to decompose any vector S ∈ L 2 (SIM(2), dµ) into a sum (possibly infinite) of

265

7.2 The 2-D continuous wavelet transform

wavelet transforms of appropriate signals with respect to appropriate mother wavelets. For the affine group this was achieved in (6.41) and we would like to do the same in the 2-D case. Since the Duflo–Moore operator C has an inverse, we can choose an orthonormal 2 2 2 basis, {φn }∞ ), such that each φn is in the domain of n=1 , in the signal space L (R , d x −1 −1 C . Thus, the vectors ψn = C φn are admissible. We recall next that the orthogonality condition (7.12) implies that all wavelet transforms are elements of L 2 (SIM(2), dµ). a, θ) denotes the wavelet transform of a signal s with respect to Specifically, if Sn (b, a, θ) that of a signal s with respect to the mother the mother wavelet ψn and Sm (b, wavelet ψm , then a, θ ) Sn (b, a, θ) Sm (b, a, θ) = δnm s | s . dµ(b, (7.48) SIM(2) For each n = 1, 2, . . . , ∞, denoting by Hψn the space of all wavelet transforms with respect to the mother wavelet ψn , we infer from the above equation that these spaces are 2 mutually orthogonal. Furthermore, ⊕∞ n=1 Hψn ⊆ L (SIM(2), dµ), and since the group has only one irreducible unitary representation, one can in fact show that ∞

Hψn L 2 (SIM(2), dµ),

(7.49)

n=1

where denotes (unitary) equivalence. (A general discussion of such decompositions, and related results, may be found in [Ali00].) Thus, we have justified the expansion S=

∞

Sn ,

S ∈ L 2 (SIM(2), dµ),

(7.50)

n=1

of an arbitrary element of L 2 (SIM(2), dµ) in terms of wavelet transforms. Note that, a, θ) = U (b, a, θ)ψn | s, for some signal vector s ∈ L 2 (R2 , d 2 x), in the above, Sn (b, where in general, s is different for different n. Note also that the sum in (7.50) holds a, θ ) are continuous in the sense of the L 2 -norm, so that, although the functions Sn (b,

∞ in all variables, we only have S(b, a, θ) = i=1 Sn (b, a, θ ) almost everywhere (with respect to dµ). We know, from the general theory outlined in Section 7.1.4, that each one of the subspaces Hψn is a reproducing kernel Hilbert space. Let a, θ; b , a , θ ) = U (b, a, θ)ψn | U (b , a , θ )ψn K ψn (b, be the reproducing kernel for Hψn . Then, given S ∈ L 2 (SIM(2), dµ), the component Sn appearing in (7.50) can be computed using (7.18): a, θ; b , a , θ ) S(b , a , θ ), dµ(b , a , θ ) K ψn (b, (7.51) Sn (b, a, θ) = SIM(2)

a, θ) = U (b, a, θ )ψn | s, the signal vector s may be computed using and, writing Sn (b, (7.46):

266

Matrix geometry of wavelet analysis. II

s= SIM(2)

a, θ ) Sn (b, a, θ) ψb,a,θ dµ(b, .

(7.52)

If we introduce the basic wavelet transforms, a, θ ) = U (b, a, θ )ψn | φm , Snm (b,

φm = Cψm , n, m = 1, 2, . . . , ∞,

(7.53)

then, by (7.48) and the orthonormality of the φn , these functions are seen to satisfy a, θ ) Snm (b, a, θ) Sk (b, a, θ) = δnk δm . dµ(b, (7.54) SIM(2)

Hence, any S ∈ L 2 (SIM(2), dµ) has the orthogonal decomposition ∞

a, θ ) = S(b,

a, θ), cnm Snm (b,

(7.55)

n,m=1

with

a, θ) Snm (b, a, θ) S(b, a, θ), dµ(b,

cnm =

(7.56)

SIM(2)

and

a, θ) |S(b, a, θ)|2 = dµ(b,

S2 = SIM(2)

7.2.3.1

∞ ∞

|cnm |2 .

n=0 m=−∞

A concrete example Finally, we give a concrete example of the decomposition (7.50), in terms of mother wavelets built out of the well-known trigonometric functions and Laguerre polynomials. 2 , d 2 k), we choose the basis vectors In the Fourier-transformed signal space L 2 (R , 1 ,−1/2 e− 2 L n (,) eimϑ , (2π )1/2 n = 0, 1, 2, . . . , ∞ , m = 0, ±1, ±2, . . . , ±∞ ,

nm (k) = φ

(7.57)

where ,, ϑ are the polar coordinates of k and the L n (,) are the Laguerre polynomials: n n (−,)k L n (,) = . k k! k=1 These satisfy the orthogonality relations, ∞ L m (,) L n (,) e−, d, = δmn , 0

nm are orthonormal, implying that the φ k = δnk δm . nm | φ φ

(7.58)

267

7.2 The 2-D continuous wavelet transform

2 , d 2 k), follows from well-known properties of The fact that they form a basis of L 2 (R Laguerre polynomials and trigonometric functions. Moreover, it is clear that the vectors −1 φ nm = C nm , where ψ nm (k) nm (k) = , φ ψ 2π 1 1/2 − ,2 = e L n (,) eimϑ , 3 , (2π ) 2 n = 0, 1, 2, . . . , ∞, m = 0, ±1, ±2, . . . , ±∞,

(7.59)

2 , d 2 k) and hence legitimate mother are also elements of the Hilbert space L 2 (R wavelets. nm , we can construct the spaces Hψnm of wavelet Using the mother wavelets ψ 2 , d 2 k). (b, nm | The total space of a, θ) = U a, θ)ψ transforms Snm (b, s , s ∈ L 2 (R all transforms would then decompose as: L 2 (SIM(2), dµ)

∞ ∞

Hψnm .

n=0 m=−∞

7.2.3.2

Decomposition into orthogonal angular channels We saw, at the end of Section 6.1.2, how a 1-D wavelet transform can be analyzed into wavelet transforms in orthogonal channels. Here we carry out a similar decomposition of a 2-D wavelet transform into orthogonal angular channels, again following a suggestion nm defined above. Let ψ be an arbitrary mother in [105]. We use the mother wavelets ψ wavelet in the Fourier domain. We may then write, k) = ψ(

∞ ∞

mn (k) = amn ψ

n=1 m=1

∞

am (,) eimϑ ,

(7.60)

m=1

where am (,) =

∞

1 (2π )

3 2

,

anm ,1/2 e− 2 L n (,),

mn | C ψ. amn = φ

(7.61)

n=1

The sum in (7.60) may be looked upon as a decomposition of the mother wavelet into angular channels. Next, writing m (k) = am (,) eimϑ , ψ

(7.62)

m is a vector which is in the domain of the Duflo–Moore operator C we easily see that ψ a, θ ) is the 2-D wavelet transform and hence it can be used as a mother wavelet. If Sψ (b, 2 2 2 of a signal s ∈ L (R , d k), then it is straightforward to verify that a, θ) = Sψ (b,

∞ m=1

a, θ), Sm (b,

(b, m | a, θ ) = U a, θ )ψ Sm (b, s.

(7.63)

268

Matrix geometry of wavelet analysis. II

It is clear that, for m = n, the wavelet transforms Sm and Sn are orthogonal functions m and ψ n are themselves orthogonal in L 2 (SIM(2), dµ), while the mother wavelets ψ 2 2 2 vectors in L (R , d k). Thus we call (7.63) a decomposition of the wavelet transform a, θ ) into orthogonal angular channels, the transform Sm (b, a, θ ) being the comSψ (b, ponent along the m-th channel.

7.3

2-D wavelets on phase space In Section 2.3.2 we pointed out how the 2-D wavelet transform could also be looked upon as a function on a physical phase space. Here we take up this point again and give a more exhaustive mathematical treatment of it. The SIM(2) group has only one nontrivial coadjoint orbit and hence only one phase space. Moreover, this phase space is topologically homeomorphic to the group itself, meaning that wavelet transforms may also be viewed upon as transforms built on this phase space. In order to analyze these features, it will first be necessary to study the matrix structure of the generators of the various transformations constituting the group.

7.3.1

Lie algebra and orbits The four basic sets of operations of dilation, rotation and the two translations, each constitute one-parameter subgroups of SIM(2). More precisely, these subgroups are generated by group elements of the type et , 0), a, θ ) = (0, (b, t ∈ R, or (b, a, θ ) = (0, 1, t), t ∈ [0, 2π ), a, θ) = (ei t, 1, 0), t ∈ R, i = 1, 2, or (b, where 0 0 = , 0

1 , e1 = 0

0 e2 = , 1

constitute one-parameter subgroups. A general element of SIM(2) can be written as a product of elements of these subgroups: 1, θ ) (0, a, 0) , where b = b1 . a, θ ) = (e1 b1 , 1, 0) (e2 b2 , 1, 0) (0, (b, b2 Generically, writing elements in any one of these subgroups as g(t) and computing the derivative at the identity: + + d , g(t)++ dt t=0

as was done for the affine group in (6.77), we obtain the four 3 × 3 matrices

269

7.3 2-D wavelets on phase space

1 0 0 D = 0 1 0, 0 0 0 0 0 1 P1 = 0 0 0 , 0 0 0

0 −1 0 J = 1 0 0, 0 0 0 0 0 0 P2 = 0 0 1 . 0 0 0

(7.64)

et , 0), J that of rotations, et J = (0, 1, t), Here D is the generator of dilations, et D = (0, t Pi and P1 , P2 those of translations, e = (tei , 1, 0), i = 1, 2. The four generators satisfy the commutation relations [D, J ] = 0,

[D, Pi ] = Pi , i = 1, 2,

[J, P1 ] = P2 ,

[J, P2 ] = −P1 ,

[P1 , P2 ] = 0, and together they generate the Lie algebra of SIM(2) which, in this case, is a fourdimensional real vector space. We denote this Lie algebra by sim(2) and its dual space by sim(2)∗ . A general element of the Lie algebra sim(2) has the form λ −θ β1 = λD + θ J + β1 P1 + β2 P2 = (7.65) X ( α , β) θ λ β2 , 0 0 0 with θ, λ, β1 , β2 ∈ R and λ β1 . α = , β = θ β2

(7.66)

v ), Corresponding to a 2-vector v, consider again, as in (2.52), the 2 × 2 matrix s( v1 −v2 v1 s(v ) = , v = . (7.67) v2 v1 v2 Then, for any two vectors v and w,

s(v )s(w) = s(w) s( v ),

s(v )w = s(w) v.

Using s, the general Lie algebra element (7.65) may be rewritten as s ( α ) β = X ( α , β) . 0 0 yields the group element Exponentiating the matrix X ( α , β)

(7.68)

(7.69)

270

Matrix geometry of wavelet analysis. II

g=e

X ( α ,β)

=

eλ rθ

F(s( α ))β

0

1

,

(7.70)

α )) is defined as the sum of an infinite where now 0 θ < 2π and the 2 × 2 matrix F(s( series: α )) = I2 + F(s(

[s( s( α ) [s( α )]2 α )]3 + + + ... . 2! 3! 4!

(7.71)

We shall see below that every group element can be so obtained, by exponentiating an appropriate Lie algebra element. It will be useful for the sequel to express the operator F(s( α )) in a somewhat different form. Let us define a function, sinch, of a real variable u as sinch u =

sinh u , u

sinch 0 = 1.

(7.72)

This is a positive, infinitely differentiable function and so also is the related function u u F(u) = e 2 sinch ( ). 2

(7.73)

This latter function has the Taylor expansion u u u u2 u3 + + + ... . F(u) = e 2 sinch ( ) = 1 + 2 2! 3! 4!

(7.74)

For u = 0, we may also write u u F(u) = e 2 sinch ( ) = u −1 (eu − 1). 2

Using the function F we now define the 2 × 2 matrix valued function A A A2 A3 A = F(A) = I2 + + + + . . . , F(O2 ) = I2 , e 2 sinch 2 2! 3! 4!

(7.75)

(7.76)

for any 2 × 2 real matrix A and where O2 and I2 are, respectively, the 2 × 2 null and identity matrices. If detA = 0, then A A F(A) = e 2 sinch (7.77) = A−1 [e A − I2 ], 2 and if, det[e A − I2 ] = 0, we shall also write A

F(A)

−1

e− 2 = [e A − I2 ]−1 A. = sinch A2

Hence, for | α | = 0,

(7.78)

271

7.3 2-D wavelets on phase space

−eλ sin θ λ θ eλ cos θ − 1 eλ sin θ −θ λ eλ cos θ − 1 s( α) sinch , 2

1 F(s( α )) = 2 λ + θ2 =e

s( α)

2

and α ))] [F(s(

−1

1 = 2(cosh λ − cos θ ) =

e− sinch

cos θ − e−λ − sin θ

sin θ cos θ − e−λ

(7.79) λ θ

−θ λ

s( α)

2

s( α)

.

(7.80)

2

Going back to (7.70), we rewrite it as s( α) α) λ 2 sinch s( r e β e θ 2 , g = e X (α,β) = T 0 1

(7.81)

θ, a) in the form given in (7.37), we find the relations and writing (b, λ = log a,

α ))]−1 b , β = [F(s(

(7.82)

between the group parameters and those of the Lie algebra. This also shows that any θ, a) can be written as the exponential of some element X ( in group element (b, α , β) the Lie algebra. The group SIM(2) acts on its Lie algebra sim(2) via the adjoint action: a, θ) X ( a, θ)−1 = X ( = (b, (b, X ( α , β) α , β) α , β) Ad(b,a,θ) α + a rθ β s( α ) −s(b) , = 0 T 0

(7.83)

(in computing the above, we have made use of the fact that rθ s( α )r−θ = s( α ).) Writing

X ( α , β) = X ( α , β ), we get the transformation rules for the components: α = α α + a rθ β , β = −s(b)

(7.84)

θ, a) of the adjoint action in the {D, J, P1 , P2 } basis becomes so that the matrix M(b, I2 O2

α α a, θ ) = , (7.85) M(b, = M(b, a, θ) β . β −s(b) a rθ The adjoint action induces an action on the dual sim(2)∗ of the Lie algebra, which we now proceed to determine. Let {D ∗ , J ∗ , P1∗ , P2∗ } denote the elements of the basis in sim(2)∗ which is dual to the basis {D, J, P1 , P2 } of sim(2). We write a general element X ∗ ∈ sim(2)∗ as

272

Matrix geometry of wavelet analysis. II

X ∗ = α1∗ D ∗ + α2∗ J ∗ + β1∗ P1∗ + β2∗ P2∗ , We also set α1∗ α ∗ = , α2∗

β∗ =

β1∗ β2∗

α1∗ , α2∗ , β1∗ , β2∗ ∈ R.

and

γ =

α ∗ β∗

(7.86)

,

(7.87)

so that that the dual pairing between X and X ∗ is given by X ∗ ; X = α ∗ · α + β∗ · β.

(7.88)

a, θ) on sim(2)∗ , in the The matrix of the coadjoint action of a group element (b, a, θ), is the transposed inverse of the matrix (7.85) of the above basis, denoted M * (b, adjoint action: −1 T I a s ( b) r 2 θ * −1 T a, θ) )] = . (7.89) M (b, a, θ ) = [M((b, O2 a −1 rθ Writing γ =

α ∗ ∗

β

a, θ)γ = M * (b, a, θ) = M * (b,

α ∗ β∗

,

(7.90)

we obtain the transformation rules for the components of the dual vectors X ∗ : T β∗ , α ∗ = α ∗ + a −1 rθ s(b) β∗ = a −1 rθ β∗ , |β∗ | = a −1 |β∗ | .

(7.91)

Since orbits under the coadjoint action are the sets a, θ)γ0 | (b, a, θ ) ∈ SIM(2)}, O∗ = {M * (b,

(7.92)

for fixed vectors γ0 ∈ R4 , it is easy to see that there are exactly two types of orbits. (i) Trivial orbits: these are degenerate orbits, which are single point sets {γ0 }, obtained by choosing ∗ α1 α∗ 2 γ0 = α1∗ , α2∗ ∈ R. , 0 0 The isotropy subgroup of any such point γ0 , i.e., the subgroup which leaves it invariant, is of course the entire group SIM(2). (ii) The open free orbit: this is the only nontrivial orbit of SIM(2) and is obtained by choosing

273

7.3 2-D wavelets on phase space

0 0 γ0 = , 1

(7.93)

0 or by taking for γ0 any other vector such that at least one of its last two components is nonvanishing. This is the only orbit which concerns us here and we denote it by O∗ . Note that in this case the isotropy subgroup of γ0 (i.e., the subgroup of SIM(2) elements for which this vector is a fixed point) is just the trivial subgroup consisting of the identity element of SIM(2). This also means that topologically, the orbit is a, θ) → γ = M * (b, a, θ)γ0 , homeomorphic to the group space itself (the map, (b, from the group to the orbit, is open and free).

7.3.2

The coadjoint orbit O ∗ as a phase space Since the orbit O∗ is homeomorphic to the group SIM(2) itself, wavelet transforms a, θ ) may be considered as being transforms defined on this space. In other words, S(b, a, θ in terms of which the signal is being analyzed, can be looked upon the parameters b, as phase space parameters. This is completely in line with the situation encountered earlier, for 1-D wavelets and Gabor transforms. In order to understand better, the structure of the coadjoint orbit O∗ as a physical phase space, let us first note that points on the orbit are obtained from (7.91) upon setting 1 α ∗ = 0 and taking for β∗ the two dimensional unit vector e1 = (see (7.93)). Thus 0 a generic point γ ∈ O∗ is given as σ3 rθ b α ∗ 0 1 a, θ ) , = M * (b, γ = = a r θ e1 β∗ e1 1 0 , σ3 r−θ = rθ σ3 . (7.94) σ3 = 0 −1 a, θ) written as in (7.37), we find for the general Explicitly, with the group element (b, ∗ phase space point γ ∈ O , ∗ b1 cos θ + b2 sin θ α1 −b sin θ + b cos θ α∗ 1 2 2 1 . (7.95) γ = ∗ = a β1 cos θ β2∗

sin θ

274

Matrix geometry of wavelet analysis. II

a, θ) in terms of The above relations can be solved to express the group parameters (b, ∗ ∗ the phase space variables α , β : ∗ 1 ∗ −1 β2 b = a rθ σ3 α , , θ = tan . (7.96) a= β1∗ |β∗ | This also reflects the fact that the orbit O∗ is topologically homeomorphic to the group space, and moreover, from the form of (7.95) one infers that, geometrically, O∗ is also the cotangent bundle of R2∗ (2-D plane with the origin removed), i.e., O∗ SIM(2) R2 × R2∗ = T ∗ R2∗ .

(7.97)

Indeed, consider the point cos θ 1 ∈ R2∗ . β∗ = a sin θ Differentiating with respect to a and θ, we get the two tangent vectors at β∗ : 1 cos θ 1 − sin θ . , tθ = ta = − 2 a a cos θ sin θ We may thus take, as basis for the tangent space at β∗ , the columns of the matrix, 1 cos θ − sin θ . T= a sin θ cos θ The columns of the transposed matrix: cos θ sin θ 1 1 , , tθ∗ = ta∗ = a a − sin θ cos θ then form a basis for the cotangent (i.e., dual of the tangent) space at β∗ . An arbitrary element of this dual space has, therefore, the form b1 cos θ + b2 sin θ 1 , b1 , b2 ∈ R . α ∗ = b1 ta∗ + b2 tθ∗ = a −b sin θ + b cos θ 1

2

Thus, we shall interpret β∗ in (7.95) as representing a point in the manifold R2∗ and α ∗ as a vector in its cotangent space at this point. Physically, one calls the α ∗ configuration space vectors , while the β∗ are momentum vectors. Coadjoint orbits carry natural invariant measures under the group action (see, for example, [Kir76]). Using the transformation rules (7.91) under the coadjoint action, it is straightforward to compute the invariant measure on O∗ . Expressed in terms of the α ∗ , β∗ , it is

275

7.3 2-D wavelets on phase space

d( α ∗ , β∗ ) =

d 2 α ∗ d 2 β∗ dα1∗ dα2∗ dβ1∗ dβ2∗ = . β1∗ 2 + β2∗ 2 |β∗ |2

(7.98)

(In the physical literature, this would be called the Liouville measure for this phase space.) If we express the position vector β∗ in polar coordinates (ρ = |β∗ |, θ ), the coadjointinvariant measure (7.98) transforms to d( α ∗ , ρ, θ) = d 2 α ∗

dρ dθ. ρ

(7.99)

In these coordinates it is easy to verify that the differential 2-form ω( α ∗ , ρ, θ) =

1 dα1∗ ∧ dρ + dα2∗ ∧ dθ, ρ

(7.100)

is invariant under the coadjoint action. Moreover, the coordinate transformations on phase space, α ∗ → − α∗,

ρ →

1 , ρ

θ → −θ,

(7.101)

leave this 2-form invariant, meaning that they constitute a canonical transformation of the phase space. (We might point out that the 2-form (7.100) is just the well-known Kirillov–Kostant–Souriau symplectic structure [Kir76], carried by coadjoint orbits.) a, θ ) On the other hand, if we parametrize O∗ by means of the group parameters (b, using (7.96), the coadjoint action (7.91) transforms to group multiplication from the left. In other words, the coadjoint action Ad*g0 , corresponding to the group element a, θ ) as follows: g0 = (b0 , a0 , θ0 ) transforms the point in O∗ represented by (b, a0 a, θ0 + θ), a, θ) = (b0 + a0rθ b, a, θ ) → (b , a , θ ) = (b0 , a0 , θ0 )(b, (b,

(7.102)

and thus the invariant measure (7.98) on O∗ , changes to precisely the left Haar measure dµ in (7.38) under this transformation [compare (2.51)]: a, θ) = dµ(b, a, θ) = d(b,

1 2 d b da dθ. a3

(7.103)

There is yet another parametrization of the points of the orbit O∗ , which in a sense a, θ) parametrizations. This other is more natural than either the ( α ∗ , β∗ ) or the (b, parametrization is given in terms of the so-called Darboux coordinates, which we denote by ( q , p) and which are related to the other two sets of coordinates as q1 1 q = = rθ σ3 α ∗ = b ρ q2 cos θ cos θ 1 p1 = . (7.104) = β∗ = ρ p = p2 a sin θ sin θ

276

Matrix geometry of wavelet analysis. II

The transformation properties of these coordinates under the coadjoint action are also easily obtained. Once again, if ( q , p ) is the transform of ( q , p) under Ad*g0 , g0 = (b0 , a0 , θ0 ), then q = b0 + a0 rθ0 q,

p = a0−1 rθ0 p.

(7.105)

One can verify that the differential 2-form, ω( q , p) = dq1 ∧ dp1 + dq2 ∧ dp2

(7.106)

is invariant under the above coadjoint action and hence the corresponding Liouville measure on O∗ d( q , p) = d 2 q d 2 p = dq1 dq2 dp1 dp2 ,

(7.107)

is also invariant under this action. It is not hard to see that (7.106) and (7.107) are precisely the transforms of the 2-form (7.100) and the measure (7.103), respectively, under the coordinate change (7.104). This last choice of coordinates, and in particular the differential form (7.106) makes evident the phase space structure of the coadjoint orbit O∗ . The components of q refer to the position of the system on the configuration space R2 , while at each such point the vector p ∈ R2∗ denotes its canonical momentum. On the other hand, the fact that the a, θ) can also be used as coordinates for the orbit O∗ , shows that group parameters (b, the group itself can be identified with the phase space as well. In this case, b denotes a point in configuration space and (a −1 , θ ) are the polar coordinates of a momentum vector.

7.4

The affine Poincare´ group While the SIM(2) group is the most natural generalization of the affine group for building 2-D wavelets, it is by no means the only group which could be used. As a matter of fact, any group of the type G = R2 H , where H is a group consisting of 2 × 2 matrices, which acting on R2 generates an open free orbit, can be used to build wavelets. In other words, if the group H is such that for some fixed 2-vector x, the set, Ox = {y = hT x | h ∈ H}, h x = x if and only if h is the identity matrix, is an open set in R2 and for all x = 0, then such a group G has square integrable representations and hence can be used to build wavelets. As an example, we briefly look at the affine Poincar´e group. This group, which we denote by Paff , is a semidirect product of the above type. It differs from the SIM(2) group in that the spatial rotations rθ are replaced by hyperbolic rotations #ϑ :

277

7.4 The affine Poincare´ group

#ϑ =

cosh ϑ sinh ϑ

sinh ϑ cosh ϑ

.

(7.108)

The set of matrices {#ϑ | ϑ ∈ R} (note that det #ϑ = 1 and #−1 ϑ = #−ϑ ) constitutes a group, denoted SO0 (1, 1). (In physics, this is the group of relativistic transformations of a space–time having only one spatial dimension.) In dealing with this group, we shall use the physicists’ convention of writing the components of a vector x as x0 x = , x 0 , x ∈ R, x and use the Minkowski inner product between two such vectors: 1 0

T

x ; x = x0 x0 − xx = x g x , g= , g 2 = I2 . 0 −1

(7.109)

2 will also be defined using this inner product. If x = Duality between R2 and R

#ϑ x, then x ; x = x ; x , so that hyperbolas x02 − x2 = const are mapped into themselves by SO0 (1, 1).

7.4.1

Group structure and representations A general element of Paff has the matrix representation, a#ϑ b b0 a, ϑ) = , a > 0 , ϑ ∈ R , b = ∈ R2 , (b, b 1 0 T

(7.110)

giving the group the structure of the semidirect product, Paff = R2 (R+ ∗ × SO0 (1, 1)). Topologically, Paff R2 × C, where C is any one of the four open cones: ↑

x ∈ R2 | x02 > x2 , ±x0 > 0}, C± = { ↓

x ∈ R2 | x02 < x2 , ±x0 > 0}. C± = {

(7.111)

1, 0), b ∈ R2 , is a commutative subgroup of Paff The set of elements of the type (b, a, ϑ), a > 0, ϑ ∈ R . and so also is the set of elements, (0, The affine Poincar´e group is nonunimodular; the left and right Haar measures can be computed in exactly the same way that we computed them for the similitude group in Section 7.2.1. This time the two measures are a, ϑ) = dµ(b,

1 2 d b da dϑ, a3

and

a, ϑ) = dµr (b,

1 2 d b da dϑ, a

(7.112)

278

Matrix geometry of wavelet analysis. II

which look exactly the same as those for the SIM(2) group (to be expected, since the group SO0 (1, 1) is unimodular). The group Paff acts on the plane in the manner, y → a#ϑ y + b and therefore, we can again look for its representations in the signal space L 2 (R2 , d 2 x). As already noted, the situation here is largely similar to that of the similitude group. The signal symmetries again include translations and dilations; however, we have hyperbolic rotations now and not rigid rotations of space. Such signal symmetries could be expected in problems involving the detection of extremely fast moving objects (such as occurs, for example, in high energy physical experiments). Unlike the rotations rθ , the action of #ϑ actually has the effect of deforming the shapes of objects: the disc, x02 + x2 r 2 , is transformed into the interior of the rotated ellipse, x02 + x2 + tanh(2ϑ) x0 x0 sech(2ϑ) r 2 . Thus, if images are scanned using instruments which distort them in this manner, a group such as this could be more appropriate for their analysis than the similitude group. The natural unitary representation of Paff on the signal space L 2 (R2 , d 2 x), which a, ϑ): reflects its action y → a#ϑ y + b on R2 , is carried by the unitary operators U (b, a, ϑ)s]( x − b)), [U (b, x ) = a −1 s(a −1 #ϑ (

(7.113)

an expression which should be compared to that for the similitude group in (2.13). The unitarity of these operators is straightforward to prove; however, in contrast with the a, ϑ) do not carry an irreducible representation of operators (2.13), the operators U (b, Paff . For isolating the irreducible sectors, it is best to work in the Fourier domain. In order to do this, it will be convenient to adopt the physicists’ convention for defining the Fourier transform, using the Minkowski inner product. Accordingly, we define 2 , d 2 k), F : L 2 (R2 , d 2 x) → L 2 (R = = 1 (Fs)(k) s(k) d 2 x eik ; x s( x ), s ∈ L 2 (R2 , d 2 x), 2π R2 1 2 , d 2 k). = s( s )(k) x) = s(k), s ∈ L 2 (R (7.114) d 2 k e−ik ; x (F −1 2 2π R a, ϑ) are Using this Fourier transform, and the matrix identity, g#ϑ g = #−ϑ , the U (b, 2 2 2 seen to transform into the operators U (b, a, ϑ) on L (R , d k): (b, a, ϑ) = a eik ; b U s (k) s(a#−ϑ k). (7.115) Let C be any one of the four open cones defined in (7.111). A quick computation shows (b, a, ϑ) in (7.115) that if k ∈ C, then also a#−ϑ k ∈ C. From the nature of the operator U we see that if s has support inside this cone, then so also does the transformed function 2 , d 2 k), (b, a, ϑ) which is a subspace of L 2 (R U s. Thus, the Hilbert space L 2 (C, d 2 k), carries a subrepresentation of U (b, a, ϑ), i.e., restricted to this subspace the operators (b, a, ϑ) again define a unitary representation of Paff . This representation can also be U shown to be irreducible (see, for example, [Ali00] for a detailed proof of such results). Also, we have the obvious Hilbert space decomposition,

279

7.4 The affine Poincare´ group

2 , d 2 k) = L 2 (C+↑ , d 2 k) ⊕ L 2 (C−↑ , d 2 k) ⊕ L 2 (C+↓ , d 2 k) ⊕ L 2 (C−↓ , d 2 k), L 2 (R (b, a, ϑ) to these four suband using a self-evident notation for the restrictions of U spaces, we may write (b, +↑ (b, −↑ (b, +↓ (b, −↓ (b, a, ϑ) = U a, ϑ) ⊕ U a, ϑ) ⊕ U a, ϑ) ⊕ U a, ϑ). U

(7.116)

(b, a, ϑ) is a direct sum of four irreducible repThis shows that the representation U 2 , d 2 k). Each one of these resentations, carried by four orthogonal subspaces of L 2 (R subspaces consists of signals whose supports are contained in a cone. Returning to the inverse Fourier domain, the signal space decomposes as: ↑

↑

↓

↓

L 2 (R2 , d 2 x) = H+ ⊕ H− ⊕ H+ ⊕ H− , ↓

where, for example, H+ consists of all functions in L 2 (R2 , d 2 x) whose Fourier trans↓ forms have supports contained in C+ . Correspondingly, the representation decomposes as a, ϑ) = U+↑ (b, a, ϑ) ⊕ U−↑ (b, a, ϑ) ⊕ U+↓ (b, a, ϑ) ⊕ U−↓ (b, a, ϑ) . U (b,

(7.117)

a, ϑ) and in Let us generically represent any one of these subrepresentations by UC (b, the Fourier domain by UC (b, a, ϑ). This latter representation acts on the Hilbert space of all square integrable functions supported in the cone C. L 2 (C, d 2 k)

7.4.2

Affine Poincare´ wavelets C (b, a, ϑ) is known to be square integrable [Ali00]. Indeed, if we The representation U compute the integral C (b, | ψ| 2, a, ϑ) |U a, ϑ)ψ I (ψ) = dµ(b, Paff

∈ L 2 (C, d 2 k), we easily find for some ψ d 2 k 2 = (2π )2 ψ k)| 2. I (ψ) |ψ( 2 2 C |k0 − k |

(7.118)

From this follows the admissibility condition for an affine Poincar´e wavelet: a vec ∈ L 2 (C, d 2 k) is admissible if and only if it satisfies the integrability condition tor ψ (the < ∞, i.e., if and only if it is in the domain of the unbounded operator C I (ψ) Duflo–Moore operator), where 2π ψ)( k) k). = (C ψ)( (7.119) 2 |k0 − k2 |1/2 2 , we may also write down the resolution of the identity Setting cψ = C ψ 1 ψ | = I, a, ϑ) |ψ dµ(b, ψ b,a,ϑ b,a,ϑ b,a,ϑ = UC (b, a, ϑ)ψ, cψ Paff or, more generally,

(7.120)

280

Matrix geometry of wavelet analysis. II

1 ψ ψ | C C

Paff

ψ | = I, a, ϑ) |ψ dµ(b, b,a,ϑ b,a,ϑ

(7.121)

ψ ψ | C = 0. These equations should ψ such that C for two admissible vectors ψ, be compared to (7.41), (7.42), and (7.45). We can now go ahead and define the affine Poincar´e wavelet transform of an arbitrary by the quantity signal s ∈ L 2 (C, d 2 k) s(k). (7.122) S(b, a, ϑ) = ψb,a,ϑ | s=a d 2 k e−ik ; b ψ(a# −ϑ k) C

All the analysis carried out for the 2-D wavelet transform (obtained using the similitude group), including the phase space considerations, can again be repeated in the present case. In particular, any two-dimensional signal can be decomposed using affine Poincar´e wavelets. If the signal s is, for example, the quantum mechanical wave function of a fastmoving elementary particle (in a space–time of one time and one spatial dimension), a, and ϑ could represent its position (in space–time), its the analyzing parameters, b, mass and its rapidity.

8

Minimal uncertainty and Wigner transforms

This chapter is devoted to a brief examination of two topics. The first concerns a certain minimality property of gaborettes and how it generalizes to wavelets in one and two dimensions. The second is an analysis of the Wigner transform, as an alternative to the wavelet transform. This latter transform is extensively used in certain physical computations and in the analysis of radar signals. Notice that neither of these topics is a prerequisite for the study of the more general wavelets described in Chapters 9 and 10.

8.1

Phase space distributions and minimal uncertainty gaborettes The generalized gaborettes defined in (6.123), which give rise to holomorphic Gabor transforms, have a well-known minimal uncertainty property, related to localization in phase space. In (7.23) we had introduced the localization operators aψ ( ). As discussed in Chapter 7, these operators can be used to measure the proportion of the signal transform S which is concentrated in the (phase space) region . Consider the case of Gabor wavelets and let ψq, p ∈ L 2 (R, d x) be the family of gaborettes defined in (6.109), using the window function ψ. These vectors satisfy the resolution of the identity (6.112). Assuming the normalization ψ2 = 1/2π, we see that the operators aψ ( ) = dq d p |ψq, p ψq, p |,

give rise to the probability measure s|aψ ( )s 1 p S ( ) = = dq d p |S(q, p)|2, s2 2π s2

(8.1)

for any signal s ∈ L 2 (R, d x) with Gabor transform S. In Section 6.3 it was noted that the Gabor transform S(q, p) is a time–frequency transform, with q being the time and p the frequency parameter. On the other hand, if the signal s = s(q) is given as a function over time, then its Fourier transform, s = s( p) is a function over frequency. However, as is well known, the density distribution in time, |s(q)|2 /s2 , gives no information on the frequency content of the signal, while the frequency distribution, | s( p)|2 /s2 , gives no information on the variation of the signal with time. The phase 281

282

Minimal uncertainty and Wigner transforms

space density, |S(q, p)|2 , does however, carry information on both time and frequency, but this information is not expected to be sharp, i.e., we do not expect |s(q)|2 to be the marginal density of |S(q, p)|2 in time or | s( p)|2 to be its marginal in frequency. Indeed, computing these marginal densities, we find ∞ ∞ S1 (q) = dp |S(q, p)|2 = d x χq (x) |s(x)|2 , −∞ −∞ ∞ ∞ 2 dq |S(q, p)| = dx χ q (ξ ) | s(ξ )|2 , (8.2) S2 ( p) = −∞

−∞

where χq (x) = 2π |ψ(x − q)|2

− p)|2 , and χ p (ξ ) = 2π |ψ(ξ

(8.3)

are the shifted weight functions in time and frequency, respectively, generated by the window ψ. Thus, the marginal density S1 (q) appears as a weighted average, at each point, over the sharp density |s(x)|2 in time, while S2 ( p) appears as a similarly weighted average over the sharp density | s(ξ )|2 in frequency. It is in this sense that we should 2 think of |S(q, p)| as being an averaged phase space density, the variables q and p representing averages of their “sharp” values, computed with respect to the probability )|2 , respectively. It then becomes pertinent to ask for densities 2π |ψ(x)|2 and 2π|ψ(ξ what choice of window function, this averaging would be optimal, i.e., entail a minimum of unsharpness, for it is clear from (8.2) and (8.3) that there is no L 2 -function ψ for which S1 (q) = |s(q)|2 and S2 ( p) = | s( p)|2 . Consider the standard deviations of the (sharp) time and frequency variables, measured with respect to the probability distributions )|2 , respectively: 2π |ψ(x)|2 and 2π |ψ(ξ " σψ = 2π " σψ = 2π

∞

−∞

∞

−∞

d x x |ψ(x)| − 2π 2

2

)|2 − 2π dξ ξ 2 |ψ(ξ

∞

2 # 12 2

d x x|ψ(x)| −∞

∞

−∞

)|2 dξ ξ |ψ(ξ

,

2 # 12 .

(8.4)

It is well known, from the theory of Fourier transforms, that their product satisfies the inequality σψ .σψ 12 and that equality is attained when ψ(x) is a Gaussian as in (6.116) (or one of the other modified gaborettes ψ(0,q, p) , constructed using such a Gaussian). Hence the choice of a Gaussian for the window function ψ would, in the light of the present analysis, lead to a Gabor transform S(q, p) which measures the variables q and p with optimal accuracy. In order to put the above discussion in operator terms and to make the connection with group theory again, let us assume that the chosen window function ψ satisfies the symmetry property |ψ(x)|2 = |ψ(−x)|2 , almost everywhere. A fairly straightforward computation (see, for example, [5] for details) then leads to the following interesting average values:

283

8.1 Phase space distributions and minimal uncertainty gaborettes

q=

R

s|Qs d p S (q, p) q = , 2π s2

p=

R

d p S (q, p) p =

s|Ps . 2π s2

(8.5)

where Q and P are (unbounded) self-adjoint operators, defined by the integral relations Q= dq d p q|ψq, p ψq, p |, P= dq d p p|ψq, p ψq, p |. (8.6) R2

R2

By (8.1), the phase space probability distribution p S has the density ρ(q, p) =

1 |S(q, p)|2 2π s2

and hence it follows from (8.5) that the operator Q (respectively, P) gives the mean value of the phase space position parameter q (respectively, momentum parameter p), computed using the probability distribution determined by the Gabor transform S(q, p) of the signal s. Indeed, for an arbitrary signal vector s, having Gabor transform S corresponding to the window ψ, we get s|Qs = dq d p q|S(q, p)|2 , s|Ps = dq d p p|S(q, p)|2 . (8.7) R2

R2

It is remarkable that the values of the two integrals in (8.7) depend only on the signal vector and not on the window ψ. To see this, we note that the actions of these operators on the Hilbert space L 2 (R, d x) are easily calculated. One obtains, (Qs)(x) = xs(x),

(Ps)(x) = −

i d s(x), λ dx

(8.8)

on vectors s taken from appropriate domains. Thus, for λ−1 = , Q and P are the well-known position and momentum operators of quantum mechanics, satisfying the commutation relations, [Q, P] = Q P − P Q =

i I, λ

(8.9)

but now appearing explicitly as measures of average phase space position and momentum localization. This shows that Q and P are independent of the window function ψ, although their expressions in (8.6) appear in terms of it. Once again, this brings out a point we noted earlier: although the Gabor transform (like the wavelet transform) depends on the window, intrinsic quantities, such as mean values of the parameters of the transform, turn out to be independent of its choice. It is also possible to rewrite the unitary operators U λ in (6.108), realizing an irreducible representation of the Weyl– Heisenberg group, in terms of the operators Q and P [Ali00]. We get U λ (θ, q, p) = eiλ(θ+ p Q−q P) .

(8.10)

This means that Q, P (together with the identity operator I ) also constitute a Hilbert space representation of the generators (6.106) of the Lie algebra of the Weyl–Heisenberg group.

284

Minimal uncertainty and Wigner transforms

Given any self-adjoint operator A on a Hilbert space H, we define its mean value A in the state (i.e., normalized vector) φ ∈ H by A/≡ Aφ = φ|Aφ and its standard deviation A (in the state φ) by A ≡ φ A = A2 − A2 . In standard quantum mechanical lore, it well-known [Coh89,Got66], that given two self-adjoint operators A and B, the product of their standard deviations obeys the uncertainty relation

A. B

1 |[A, B], 2

[A, B] = AB − B A.

(8.11)

The state φ is said to have minimum uncertainty if equality holds in (8.11), which happens if and only if (A − A)φ = −iλo (B − B)φ,

(8.12)

for some λo > 0. By (8.9), for the operators Q and P, the uncertainty relation (8.11) assumes the form:

Q. P

1 , 2λ

(8.13)

which in this case is exactly the same relation as (8.4). Thus, minimal uncertainty is attained for vectors of the type: λ2 φ(x) = π

! 14

q

eiλ(x− 2 ) p e−

λ2 (x−q)2 2

,

(8.14)

for fixed λ = 0 and q, p ∈ R. Of course, these vectors are precisely the generalized gaborettes defined in (6.123), which minimize the product of the standard deviations (8.4) and which give rise to holomorphic Gabor transforms. Thus, referring to our previous discussion, for a signal vector s, the absolute square of its Gabor transform |S(q, p)|2 can be interpreted as giving a sort of “unsharp” joint probability distribution of the position (which now appears as the spectrum of the operator Q) and momentum (the spectrum of P) variables, the uncertainty relation (8.13) forbidding the existence of a sharp joint distribution. However, since the uncertainty is the smallest when the window is a vector of the type (8.13), the corresponding Gabor transform is in a sense optimal. In the following section, we shall use the ideas developed here to construct minimal uncertainty wavelets. We shall proceed group theoretically and isolate two generators of the corresponding representation and use their commutation relation to compute and minimize the uncertainty.

8.2

Minimal uncertainty wavelets We turn our attention now to determining minimal uncertainty wavelets in one and two dimensions. As we know, the relevant groups are G (+) aff and SIM(2) and we have to work

285

8.2 Minimal uncertainty wavelets

with the representations of their Lie algebras on the Hilbert spaces of irreducible representations [24,116]. Consider first the one-dimensional case. We know from Section 6.2.2 that the Lie algebra of G (+) aff is two-dimensional. The two matrix generators of this + (R ) algebra, X 1 and X 2 , computed in (6.77) are represented on the Hilbert space H + (b, a) (see (6.59)–(6.61)) by the generators D and of the irreducible representation U P, of dilation and translation, respectively. They act on Hilbert space vectors in the manner 1 ξ (D s )(ξ ) = −i ([ + ξ ] s )(ξ ), 2 dξ

(P s )(ξ ) = ξ s(ξ ),

(8.15)

+ (R ) are chosen from the appropriate domains of the unbounded operators. where s∈H (Note, this is the “momentum space representation” of these operators.) The operators and P satisfy the commutation relations (see (6.78)) D P] = i P. [ D,

(8.16)

The minimal uncertainty vectors in this case are found to be [246,Pau85,305] the 1-D m (ξ ) = ξ m e−ξ , for ξ 0 (m > 0) and 0, otherwise. Note Cauchy wavelets, namely, ψ that these are also the wavelets which lead to holomorphic wavelet transforms. A similar analysis applies for 2-D wavelets. The Lie algebra is now four-dimensional. The four generators, denoted by P1 , P2 for translations, D for dilations and J for rotations, may be derived explicitly from the transformation (7.35), or its action (2.13) on signals or its equivalent (2.14) in k-space. (Note that we are using the same notation as for the corresponding matrix generators of the Lie algebra (see (7.64)).) Among these four operators, there are four nonzero commutators, namely [D, P1 ] = i P1 ;

[J, P2 ] = −i P1 ;

[D, P2 ] = i P2 ;

[J, P1 ] = i P2 ,

(8.17)

but the first two transform into the last two under a rotation by π/2. More generally, defining Pγ = P1 cos γ + P2 sin γ , we replace both pairs of commutators by the relations: [D, Pγ ] = i Pγ ;

[J, Pγ +π/2 ] = −i Pγ ,

(8.18)

and we look for wavelets which are minimal with respect to this pair. Thus, minimality has to be defined with respect to a fixed direction eγ , and it is impossible to do it for two directions at the same time, for instance, for all four relations (8.17) simultaneously [116]. For fixing ideas, we consider the uncertainty relations for the first pair in (8.17), corresponding to γ = 0:

D. P1

1 |P1 |; 2

J. P2

1 |P1 |. 2

(8.19)

saturates these inequalities iff it satisfies the Then, according to (8.12), a vector ψ following system of equations

286

Minimal uncertainty and Wigner transforms

k) k) = (D + iλ1 P1 )ψ( (D + iλ1 P1 )ψ( k) k) = (J + iλ2 P1 )ψ( (J + iλ2 P2 )ψ(

(λ1 , λ2 > 0) .

(8.20)

This system of partial differential equations may be solved, in polar coordinates k = (ρ, φ), imposing successively five conditions: (i) Integrability of the system requires λ1 = λ2 = λ > 0. (ii) 2π -periodicity in φ implies that P2 = 0 and J = m ∈ Z. (iii) ∈ L 2 , implies that the support of ψ is restricted to a convex Square integrability, ψ cone in the right half-plane. (iv) Admissibility of ψ implies λP1 > 1. (v) Finally, k) be real implies that J = D = 0. imposing the condition that ψ( is minimal with respect to the first pair of the The result is that a real wavelet ψ commutation relations (8.17) iff it vanishes outside some convex cone C in the halfplane k x > 0 and is exponentially decreasing inside: κ −λkx (κ > 0, λ > 0), k ∈ C, k) = c |k| e (8.21) ψ( 0, otherwise. k) |k| κ e−λ kx = c χC (k) ψ(

(κ > 0, λ > 0),

(8.22)

where χC is the characteristic function of C, or a smoothened version thereof. More generally, if one chooses the commutation relations in (8.18), one obtains a similar result, rotated by γ , that is, a wavelet supported in a convex cone Cγ with axis in the direction eγ , and exponentially decreasing inside. We may now impose some degree of regularity (vanishing moments) at the boundary of the cone, by taking an appropriate linear superposition of such minimal wavelets ψ. Thus we obtain finally: C (k) = c χC (k) F(k) e−λkx , ψ

(λ > 0)

(8.23)

is a polynomial in k x , k y , vanishing at the boundaries of the cone C, where F(k) including the origin. Clearly a Cauchy wavelet is of this type but, of course, one uses in practice a narrow support cone C, in order to obtain good directional selectivity, as discussed in Section 3.3.4. Other minimal wavelets may be obtained if one includes commutators with elements of the enveloping algebra, i.e., polynomials in the generators. For instance, if one requires the wavelet to be rotation invariant, one may start from the commutator between D and the Laplacian − = P12 + P22 . Then one finds a whole family of minimal isotropic wavelets, among them all powers of the Laplacian, n , acting on a Gaussian, i.e., the wavelets (3.7) [22]. For n = 2, this gives the 2-D isotropic Mexican hat (3.6) [116]. There exist more general solutions of the minimizing equations, but most of them are not square integrable. We ought to emphasize at this point, that the property of minimality for wavelets is a mathematical one, and it is not clear whether it implies an operational meaning in the same way as was discussed for gaborettes. Cauchy wavelets are linear combinations of

287

8.3 Wigner functions

minimal wavelets, but they are not the most efficient conical wavelets for directional analysis. This is not new: in 1-D too, the Cauchy–Paul wavelet ( [Pau85]) is minimal, but many others are as least as useful in practice, for instance the derivatives of the Gaussian or the Morlet wavelet. As a last remark, it may be interesting to note that a concept closely related to minimality has been developed by Simoncelli et al. [341] under the name of jointly shiftable filters. First, shiftable filters are the natural generalization of steerable filters to other variables than rotations, such as translation or scaling. Then a filter is jointly shiftable in two variables simultaneously if and only if the corresponding operations commute (i.e., “are independent”). Thus strict joint shiftability is impossible for position and spatial frequency; only approximately shiftable filters exist and the optimal ones, that is, those that minimize the “joint aliasing”, are the same as our minimal wavelets.

8.3

Wigner functions In Chapter 1, we mentioned the Wigner–Ville transform as an example of a signal transform which is quadratic (or generally, sesquilinear) in the signal vector(s). We now take a closer look at this transform. Wigner functions have long been used in signal analysis as phase space transforms of, generally, a pair of signal vectors. The original transform, due to Wigner [373], was introduced as a phase space quasi-probability density for computations in atomic physics. We begin by introducing the cross-Wigner function, W λ (ψ, φ|q, p), of two signal vectors ψ, φ ∈ L 2 (R, d x): λ x x λ W (ψ, φ|q, p) = d x ψ(q − ) e−iλ x p φ(q + ), λ > 0, (8.24) 2π R 2 2 adopting a somewhat different notational convention than used in (1.5). The phase space here is R2 , with the variables q and p identified either as time and frequency or position and momentum. The superscript λ will eventually be identified with the parameter labeling the representation of the Weyl–Heisenberg group (see (6.108)). The form of the expression for the cross-Wigner function makes it plausible to look upon it as a mapping of the rank-one operator ρ = |φψ| to a function of the phase space variables q, p. (Recall that the operator ρ acts on an arbitrary vector χ ∈ L 2 (R, d x) by ρχ = ψ|χ φ.) Indeed, the integral on the right-hand side of (8.24) can be manipulated to be brought into the form W λ (ψ, φ|q, p) ≡ W λ (ρ|q, p) λ

= dq dp eiλ(qp − pq ) Tr[e−iλ(Qp −Pq ) ρ], 2π R2

(8.25)

where Q and P are the operators defined in (8.8), and “Tr” denotes the trace of an operator:

288

Minimal uncertainty and Wigner transforms

Trρ =

∞

φk |ρφk ,

i=k 2 {φk }∞ k=1 being an orthonormal basis of L (R, d x). A fairly straightforward computation also shows that the unitary operator appearing within the square brackets on the right-hand side of (8.25) is none other than the representation operator of the Weyl–Heisenberg group, appearing in (6.108). Indeed,

eiλ(Qp−Pq) = U λ (0, q, p), and hence

λ W (ρ|q, p) = 2π λ

R2

dq dp eiλ(qp − pq ) Tr[U λ (0, q , p )∗ ρ],

(8.26)

which clarifies the relationship of the superscript λ, in the definition of the Wigner function, to representations of the Weyl–Heisenberg group. Let us introduce the symplectic Fourier transform, f of a function f ∈ L 2 (R2 , dq d p): λ

f (q, p) = dq dp eiλ(qp − pq ) f (q , p ), 2π R2 with inverse, λ f (q, p) = 2π

R2

f (q , p ). dq dp eiλ(qp − pq )

Clearly, the symplectic Fourier transform is a Hilbert space isometry. Then (assuming for the moment that W λ (ρ|q, p) is an L 2 -function), we may write λ (ρ|q, p). Tr[U λ (0, q, p)∗ ρ] = W

(8.27)

Comparing with (6.125), we see that the symplectic Fourier transform of the crossWigner function W λ (ψ, s|q, p) is just the generalized Gabor transform of the signal s, computed using the window ψ: λ (ψ, s|q, p). S(q, p) = W

(8.28)

However, it ought to be emphasized that, while the Gabor transform S(q, p) is an L 2 transform, on phase space, of the signal s, the cross-Wigner function W λ (ψ, s|q, p) is to be looked upon as an L 2 -transform, on the same phase space, but of the rank-one operator ρ = |sψ|. The original Wigner function [373] was defined with ψ = φ and ρ = |ψψ|. For this case, we shall use the simpler notation W λ (ψ|q, p), and call it the Wigner function for the wave function ψ (or more accurately, for the operator ρ). This function has a number of well-known properties, which make it resemble a probability distribution. However, as is clear from its definition, for a general ψ, its Wigner function W λ (ψ|q, p) is not positive for all q, p. In fact, it is only when ψ is a Gaussian of the type in (6.122), that W λ (ψ|q, p) is everywhere positive. It is for this reason that the Wigner function

289

8.3 Wigner functions

is also called a quasi-probability distribution. Nevertheless, it is still possible to think of it as a phase space transform, which in a certain sense, is the signature of the signal vector. A few properties of the Wigner function W λ (s|q, p) of a signal vector s are now in order. (i) Reality: The Wigner function is real-valued, W λ (s|q, p) = W λ (s|q, p).

(8.29)

This is not generally true of the cross-Wigner function, for which one has the hermiticity condition, W λ (ψ, φ|q, p) = W λ (φ, ψ|q, p).

(8.30)

(ii) Trace condition: dq d p W λ (s|q, p) = Trρ = s2 R2

ρ = |ss|.

(8.31)

This condition is reminiscent of the fact that for a probability distribution, the total probability equals one (as would be the case when s2 = 1). (iii) Marginality: λ 2 d p W (s|q, p) = |s(q)| , dq W λ (s|q, p) = | s( p)|2 . (8.32) R

R

If we think of |s(q)|2 as giving the distribution of the signal in time (or position) and | s(q)|2 its distribution in frequency (or momentum), then the above relations make W λ (s|q, p) formally look like a joint time–frequency (or position–momentum) distribution. Again, the fact that this “joint distribution” is not everywhere positive is a reflection of the uncertainty relations (8.13). (iv) Covariance: An important property, which the Wigner function inherits from the Weyl–Heisenberg group, is reflected in the covariance relation: W λ (U λ (0, q0 , p0 )s|q, p) = W λ (s|q − q0 , p − p0 ).

(8.33)

Identifying the phase space variables with coordinates on a coadjoint orbit of the Weyl–Heisenberg group, the transformation q → q − q0 , p → p − p0 is the coadjoint action discussed in Section 6.3.2. Thus, the above relation expresses symmetry under the natural phase space transformations. We come now to the problem of reconstructibility. In general it is not possible to reconstruct the signal s itself from its Wigner function; however, the operator |ss| can be recovered from it. More generally, the operator |φψ| can be reconstructed from the cross-Wigner function W λ (ψ, φ|q, p). In order to obtain an inversion formula, we go back to the general orthogonality relation (7.12) and see that in the present case it leads to the relation

290

Minimal uncertainty and Wigner transforms

λ 2π

R2

Sψ (q, p) dq d p Sψ (q, p) = ψ|ψ s |s,

(8.34)

where Sψ (q, p) is the generalized Gabor transform of the signal s, computed using the window ψ and Sψ (q, p) that of s computed using ψ. We note, however, that in this case, the above relation holds for any four vectors ψ, ψ , s, and s . Moreover, we immediately get from it the resolution of the identity λ

dq d p |ψ(0,q, (8.35) p) ψ(0,q, p) | = ψ|ψ I, 2π R2 again for arbitrary ψ, ψ ∈ L 2 (R, d x), and where, of course, ψ(0,q, p) = U λ (0, q, p)ψ,

λ

ψ(0,q, p) = U (0, q, p)ψ .

λ (ρ|q, p), the symplectic Fourier transform of its crossLet ρ = |ψφ| and consider W Wigner function. Then, for any ψ ∈ L 2 (R, d x) we have, ! λ λ (ρ|q, p)U λ (0, q, p) ψ dq d p W 2π R2 λ

= dq d p|ψ(0,q, p) φ(0,q, p) |ψ = φ|ψ ψ 2 2π R = ρψ . Since ψ is arbitrary in the above expression, we obtain the reconstruction formula, λ λ (ρ|q, p)U λ (0, q, p), dq d p W (8.36) ρ= 2 2π R with λ (ρ|q, p) = λ W 2π

R2

dq dp eiλ(qp − pq ) W λ (ρ|q , p ).

We can exploit the orthogonality relations (8.34) to extend the definition of the crossWigner function to arbitrary Hilbert–Schmidt operators on L 2 (R, d x). Recall that these operators form a Hilbert space with respect to the scalar product ρ1 |ρ2 2 = Tr[ρ1∗ ρ2 ]. Let us denote this Hilbert space by B2 (R). It is well-known that finite linear combinations of rank-one operators form a dense set in B2 (R). Next note that using (8.28) and the fact that the symplectic Fourier transform is an isometry, we may transform (8.34) into λ dq d p W λ (ρ |q, p) W λ (ρ|q, p) = Tr[ρ ∗ ρ], (8.37) 2 2π R where ρ = |sψ|,

ρ = |s ψ |.

The relation (8.37) remains valid if we replace ρ, ρ by finite linear combinations of rank-one operators, meaning that the map given by the integral on the right-hand

291

8.4 Wigner functions for the wavelet groups

side of (8.25), associating an operator ρ ∈ B2 (R), taken from this dense set, to the function W λ (ρ|q, p), is linear and an isometry (up to a factor of λ/2π). Using this fact we can extend the map to a unitary transformation between the Hilbert spaces B2 (R) and L 2 (R, dq d p). Thus, to any Hilbert–Schmidt operator ρ, we can associate a general Wigner function W λ (ρ|q, p). However, the explicit expression for this function is given by the integral in (8.25) only for operators with a well-defined trace. Otherwise, it has to be obtained as an L 2 -limit of such functions. Conversely, every function f in L 2 (R, dq d p) is the general Wigner function of a unique Hilbert–Schmidt operator. If f is also L 1 -integrable, then this operator is given by (see (8.36)) λ ρ= dq d p f (q, p)U λ (0, q, p), (8.38) 2π R2 where again f is the symplectic Fourier transform of f . If f is not L 1 -integrable, then the corresponding Hilbert–Schmidt operator is obtained as a Hilbert space limit (in B2 (R)) of operators obtained using (8.38). The Wigner function of a general Hilbert– Schmidt operator probably does not have a natural meaning in signal analysis. It is only Wigner functions of rank-one operators that have been used directly, as signature functions of signals. In quantum mechanics, however, a Hilbert–Schmidt operator, which is trace-class and of unit trace, represents a mixed state and its Wigner function again has the interpretation of a phase space quasi-probability distribution for this state. We end this section by repeating what we said already in Chapter 1, that the Wigner function W λ (s|q, p) of a signal is in a sense more intrinsic than its Gabor transform, since the former does not depend on an arbitrarily chosen window. On the other hand, the reconstruction formula (8.36) only gives back the operator |ss|, and not the function s itself. Still, both are transforms carrying information about the signal in terms of phase space variables. A second point to be borne in mind is that while the cross-Wigner function is sesquilinear (see (8.30)), and the Wigner function is quadratic, when looked upon as a transform on signal vectors, it is in fact linear when looked upon as a transform on the space of Hilbert–Schmidt operators.

8.4

Wigner functions for the wavelet groups In view of the fact that the Wigner function has proved itself to be an extremely useful tool, both in signal analysis and in atomic and quantum optical computations [63,64,376], it makes sense to look for similar signal transforms related to groups other than the Weyl–Heisenberg group. In particular, one would like to construct such transforms for the one- and two-dimensional wavelet groups. These could then provide one with alternatives to the wavelet transforms discussed in the previous two chapters. We now proceed to obtain such transforms, with the proviso that these generalized Wigner functions should also be phase space functions (i.e., functions defined

292

Minimal uncertainty and Wigner transforms

on coadjoint orbits of the relevant groups). Also, we would like to preserve as many of the properties (8.29)–(8.33) as possible. A general procedure for constructing such functions has been proposed in [7– 9,248]. (See also [72– 74].) Again, the objective is to find a one-to-one linear correspondence between Hilbert–Schmidt operators on the carrier space of an irreducible representation of the group and L 2 -functions on phase space. As with wavelet transforms, the square integrability of the group representations, used in the construction, will turn out to be of crucial importance and as before, the orthogonality relations (7.12) will guarantee a reconstructibility condition. We shall keep the discussion here mainly descriptive, without venturing into too many mathematical details. Just as while constructing two- (or higher) dimensional wavelet transforms, we exploited the symmetry groups of signals, so also for constructing generalized Wigner transforms we look at semi-direct product groups of a particular type. Suppose that our signal vectors are elements of L 2 (Rn , d n x), where for the rest of this discussion n = 1 or 2. Following our discussion in Section 7.1.1, we assume that the allowable transformation symmetries of our signals are of the following types: r Translations: s( x ) → s( x + b), r

Dilations: s( x ) → s(a x),

r

b ∈ Rn .

a > 0.

Matrix transformations of Rn : s( x ) → s(h x ), where h is an n × n nonsingular matrix. We shall assume that the set of all such admissible matrices form a group H and, furthermore, that the following technical condition is satisfied: there exists a vector x ∈ Rn such that the set, Ox = {y = ahT x | a > 0 , h ∈ H },

is true only when h is the identity matrix. is open in Rn and y = hy , for any y = 0, In this case Ox is called an open free orbit. (This condition is satisfied by all the wavelet groups used in the current literature, including of course, the affine group, the 2-D wavelets group and the affine Poincar´e group, studied in the previous two chapters.) Thus, we are assuming that the full symmetry group of allowed transformations on our signals is G = Rn (R+ ∗ × H ). Elements of this group can be conveniently represented by the matrices ah b a, h) = , a > 0 , h ∈ H, b ∈ Rn. g ≡ (b, (8.39) 0 T 1

293

8.4 Wigner functions for the wavelet groups

A point x ∈ Rn undergoes the transformation x → ah x + b under its action. This group is nonunimodular and its left invariant (Haar) measure is, a, h) = dµG (b,

1 a n+1 det h

d n b da dµ H (h),

(8.40)

where d n b is the Lebesgue measure on Rn and dµ H (h) the left invariant measure of the group H . The Hilbert space of signals L 2 (Rn , d n x) then carries a natural unitary repre a, h), reflecting the transformation sentation of the group G by unitary operators U (b, properties of the signals under the group action 1 −1 −1 a, h)s ( U (b, x) = h ( x − b)). (8.41) 1 s(a n (a det h) 2 n , d n k), the corresponding operators act in the manner In the Fourier space L 2 (R (b, a, h) = (a n det h) 12 e−i b· k s(a −1 hT k). (8.42) U s (k) (b, a, h) is in general not irreducible. However, if O is an open The representation U n 2 , and L (O, d n k) n , d n k), the subspace of L 2 (R consisting of functions free orbit in R supported on this orbit, then the representation U (b, a, h), restricted to this subspace, is irreducible and square-integrable; hence it is appropriate for constructing generalized wavelet transforms. It can also be shown [69] that all irreducible subrepresentations of (b, a, h) are of this type and that there is only a finite number of them (the connected U affine group, G + aff , has two, while the SIM(2) group has only one, etc.). Generalized Wigner functions, which bear a strong resemblance to the ones defined in (8.24) and (8.26), can also be constructed now using these same irreducible representations. The phase space on which the Wigner functions are defined is O∗ = Rn × O, which can be identified with a coadjoint orbit of the group G and is associated to an irreducible representation. The dimension of the phase space is 2n. The exact construction relies rather heavily on the properties of the Lie algebra g of G and its dual g∗ , which we prefer to omit, displaying and discussing only the final expressions for the two wavelet groups G + aff and SIM(2). For details, the reader may refer to [7,8,248]. Let X 1 , X 2 , . . . , X 2n be a (vector space) basis of the Lie algebra and X 1 ∗ , X 2 ∗ , . . . , X 2n ∗ the dual basis for the vector space g∗ . The basis elements are (n + 1) × (n + 1) matrices and we assume that the last n elements, X n+1 , X n+2 , . . . , X 2n correspond to translations of Rn (i.e., they are the generators of the one-parameter translation subgroups). A general element in the Lie algebra is a linear combination

2n i X = i=1 x X i = x · X , in an obvious notation. Similarly, a general element in the

2n dual is a linear combination, X ∗ = i=1 γi X i ∗ = γ · X ∗ . The identification of the 2n coadjoint orbit O∗ with Rn × O ⊂ R2n is then done with respect to the basis {X i ∗ }i=1 T and we denote a point in the orbit by a column vector γ = (γ1 , γ2 , . . . , γ2n ) . Let U (g), g ∈ G, be an irreducible (square-integrable) representation of G on the Hilbert space H. Denote by C the Duflo–Moore operator defined in (7.5) (see also (6.16) and

294

Minimal uncertainty and Wigner transforms

(7.42)) and by X i the self-adjoint operator which represents X i on the Hilbert space H will denote a vector operator with components, (via the representation U (g)). Also, X X i . The generalized Wigner function, corresponding to the irreducible representation U (g) is then explicitly given, for any Hilbert–Schmidt operator ρ on H, such that the operator ρC −1 is of trace class, by the expression 1 1 2n i( X −γ )· x −1 W (ρ | γ ) = d x Tr e ρC (8.43) [σ (γ ) m( x )] 2 , n (2π ) N0 0 is an appropriate subset of R2n , depending on certain properties of the Lie where N 0 = R2n . The function m( algebra. In the two examples given below, N x ) is needed to transform the Haar measure on the group to a measure on the Lie algebra parameters xi , while the function σ (γ ) is a density which converts the invariant measure d(γ ) on the phase space O∗ to the Lebesgue measure on R2n . The orthogonality relations (7.6) play a crucial role in the derivation of the above expression for the Wigner function, which can in fact be extended (by taking limits) to all Hilbert–Schmidt operators on H. In other words, the generalized Wigner function is again a transform of a Hilbert–Schmidt operator (on the Hilbert space of signals) to a function on phase space. If ρ = |φψ|, we obtain the cross-Wigner function, 1 1 W (ψ, φ | γ ) = d 2n x C −1 ψ | ei( X −γ )·x φ[σ (γ ) m( x )] 2 , (8.44) n (2π ) N0 provided the vector ψ is chosen from the domain of the operator C −1 , implying thereby an admissibility condition on the operator ρ. The inversion formula is then, 1 1 ρ= d(γ ) d 2n x e−i( X −γ )·x C −1 W (ρ|γ ) [σ (γ ) m( x )] 2 . (8.45) n (2π) O∗ 0 N

8.4.1

Wigner functions for the affine group + (b, a) of the connected affine group G + We take the irreducible representation U aff , defined in Section 6.2 (see (6.61)–(6.62)). The Hilbert space of this representation is + (R ) of L 2 -functions supported on the positive real axis (see (6.62)). The open free H ∗ orbit is O = R∗+ and the phase space is the coadjoint orbit O+ = R × R∗+ , identified in (6.89). This orbit is equipped with the measure (6.92), which is invariant under + (R ), φ ∈ H the coadjoint action spelled out in (6.87). If (γ1 , γ2 ) ∈ R × R∗+ , and ψ, inserting into (8.44) the cross-Wigner function becomes, ∞ x x 2 γ2 e−iγ1 x γ2 e− 2 γ2 e φ|γ 1 , γ2 ) = 1 1 φ dx ψ , (8.46) W (ψ, sinch x2 sinch x2 sinch x2 (2π ) 2 −∞ where (see (7.72)), sinch u =

sinh u . u

295

8.4 Wigner functions for the wavelet groups

For the purposes of signal analysis, the parameters γ1 and γ2 would be identified with time and frequency, respectively. The following facts should be noted about this function. r W is sesquilinear in φ, ψ and indeed, it is again a transform of the rank-one operator ψ| to a phase space function. Moreover, the support of this function is ρ = |φ contained entirely in the coadjoint orbit O∗ . r Writing W (ψ|γ 1 , γ2 ) for the case when φ = ψ, this function is seen to be real, although in general not everywhere positive, and again, the trace condition is satisfied: ∞ ∞ 1 dγ2 1 , γ2 ) = Tr [ρ], ψ|. dγ W (ψ|γ ρ = |ψ (8.47) 1 1 γ2 2 (2π) −∞ 0 r

As anticipated, the covariance condition assumes the form: + (b, a)ψ| γ ) = W (ψ|M * (b, a)−1 γ ) = W (ψ|γ 1 − bγ2 , aγ2 ), W (U

(8.48)

with M * (b, a) being the matrix of the coadjoint action of the group on the phase space, obtained in (6.85)–(6.87). r The definition of the Wigner function can be extended, using the trace condition, to + (b, a) all Hilbert–Schmidt operators ρ, on the Hilbert space of the representation U and then, the orthogonality relation ∞ ∞ 1 dγ2 dγ W (ρ1 |γ1 , γ2 ) W (ρ2 |γ1 , γ2 ) = Tr [ρ1∗ ρ2 ], (8.49) 1 1 γ 2 2 (2π) −∞ 0 holds. From this, or using (8.45) it is easy to write down a reconstruction formula for ρ, given its Wigner function. (For details, the reader is referred to [7].) r Unlike the original Wigner function (8.24), only one of the two marginality conditions (8.32) is satisfied in this case. One gets, ∞ 1 dγ1 1 , γ2 ) = |ψ(γ 2 )|2 . W (ψ|γ (8.50) 1 (2π) 2 −∞ γ2 The nonexistence of a simple form for the second marginality condition is related to the fact that there does not seem to be a “natural” choice of coordinates on the phase space, in terms of which the condition could be stated.

8.4.2

Wigner functions for the similitude group To construct Wigner functions for the SIM(2) group, we use its only irreducible a, θ), carried by the Hilbert space L 2 (R2 , d 2 x) (see (2.13)) or representation U (b, (b, a, θ ), carried by the Hilbert space equivalently, its Fourier-transformed version U 2 2 2 L (R , d k) (see (2.14)). The Duflo–Moore operator is the one given in (7.42) and we adhere to the notation and terminology introduced in Sections 7.3.1 and 7.3.2. The computation of the cross-Wigner function follows in several steps, which we now summarize.

296

Minimal uncertainty and Wigner transforms

(i) Taking a ψ in the domain of C −1 and an arbitrary vector φ in the Hilbert space we may write

C −1 ψ|ei( X −γ )·x φ = e−i[α·α

∗

β∗ ] +β·

U (e X )C −1 ψ|φ,

(8.51)

T . (Recall that in the above equation, e X indicates an where we have set x = ( α , β) element of the group SIM(2), while e−i X ·x is a unitary operator giving its Hilbert space representation.) (ii) Using the explicit form of the representation, the expression for the Duflo–Moore operator and equations (7.70) and (7.82), we obtain, = (U (e X )C −1 ψ)(k)

e2λ −i[F(s(α))β]· k |k| e ψ(eλ r−θ k). 2π

(8.52)

(iii) Next we note that, if V is a 2 × 2 real matrix of the type in (7.67) and u a 2-vector, 1 then |V u| = [detV ] 2 | u |. Thus, ! 12 λ2 + θ 2 −1 α )) k| = |k|. (8.53) |F(s( 2eλ (cosh λ − cos θ ) Using this fact to effect a change of variables in the integration involved in the scalar product, we obtain ! 32 e2λ (λ2 + θ 2 ) X −1 U (e )C φ|ψ = d 2 k ei β·k |k| 2 2π 2eλ (cosh λ − cos θ ) R ! T T /2 −s( s ( α ) e α) /2 e k ψ k . (8.54) ×φ sinch(s( sinch(s( α )T /2) α )T /2) (iv) Inserting the above expression into the formula for the Wigner function in (8.44) and using the expressions (see [8] for details of derivation), cosh λ − cos θ , λ2 + θ 2 σ (γ ) = σ ( α ∗ , β∗ ) = |β∗ |2 = (β1∗ )2 + (β2∗ )2 ,

= 2e−λ m( x ) = m( α , β)

leads to W (ψ, φ| α ∗ , β∗ ) =

1 2(2π )3

d 2 k

R2

∞

−∞

2π

dλ

(8.55)

dϑ 0

R2

d 2 β

λ2 + ϑ 2 cosh λ − cos ϑ ! T −s( α )T /2 es(α) /2 e k φ k , ×ψ sinch(s( α )T /2) sinch(s( α )T /2) λ ϑ λ s( α )T = , α = . (8.56) −ϑ λ ϑ e

β∗ )·β −i α · α i(k− ∗

e

|β∗ | |k|

297

8.4 Wigner functions for the wavelet groups

The integration over β represents a delta measure in k − β∗ . Using this fact to perform the k-integration, we obtain finally ∞ 2π λ2 + ϑ 2 1 ∗ ∗ W (ψ, φ| α ∗ , β∗ ) = dλ dϑ e−i(α1 λ+α2 ϑ) |β∗ |2 4π −∞ cosh λ − cos ϑ 0 ! T e−s(α) /2 es(α)T /2 ∗ ∗ ×ψ β φ β . sinch(s( α )T /2) sinch(s( α )T /2) (8.57) This then is the final expression of the cross-Wigner function for SIM(2) at a phase space point, written in the ( α ∗ , β∗ ) coordinates. More compactly, using the matrices defined in (7.79)–(7.80), we may write |β∗ |2 ∗ ∗ ∗ W (ψ, φ| α ,β ) = d 2 α e−i α ·α ψ(eλ r−ϑ F(s( α )T )−1 β∗ ) 2π R2 α )T )]−1 φ(F(s( α )T )−1 β∗ ). (8.58) × eλ [detF(s( Also, it is straightforward to express this function in terms of the other coordinates discussed at the end of Section 7.3.2. For example, in the Darboux (or canonical) coordinates, introduced in (7.104), the cross-Wigner function assumes the form: ρ2 W (ψ, φ| q , p) = d 2 α e−i[ ρ rϕ σ3 q ]·α ψ(eλ r−ϑ F(s( α )T )−1 p) 2π R2 α )T )]−1 φ(F(s( α )T )−1 p), × eλ [detF(s(

(8.59)

where we have written p = (ρ, ϕ) in polar coordinates while, as before, α has the Cartesian coordinates λ, ϑ. By its very construction, for each pair of wave functions ψ, φ, the cross-Wigner function W is a function on phase space. The exact variables of phase space may however be chosen to suit our convenience. Moreover, it is clear, that even when φ = ψ, the Wigner function may not be an everywhere positive function. The covariance property, in the present context, can be expressed as W (U (g0 )ψ, U (g0 )φ|γ ) = W (ψ, φ|M * (g0 )−1 γ ), g0 = (a0 , θ0 , b0 ).

(8.60)

Using (7.85), this can be written more explicitly as W (U (a0 , θ0 , b0 )ψ, U (a0 , θ0 , b0 )φ | α ∗ , β∗ ) = W (ψ, φ| α ∗ , β∗ ) α ∗ = α ∗ − s(b0 )T β∗ , β∗ = a0 r−θ0 β∗ .

(8.61)

The orthogonality condition is reflected in the cross-Wigner function through the following relation: d 2 α ∗ d 2 β∗ W (ψ1 , φ1 | α ∗ , β∗ ) W (ψ2 , φ2 | α ∗ , β∗ ) = ψ2 | ψ1 φ1 | φ2 . (8.62) ∗ 2 ∗ |β | O

298

Minimal uncertainty and Wigner transforms

Although this relation is guaranteed by the very way the Wigner function is constructed, it is still instructive to demonstrate it directly, as we now proceed to do. Denoting the left-hand member of (8.62) by I (ψ1 , φ1 , ψ2 , φ2 ) and substituting from (8.58) we get ∗

I (ψ1 , φ1 , ψ2 , φ2 ) = d 2 α ∗ d 2 β∗ |β∗ |2 d 2 α d 2 α ei α ·(α −α) O∗

R 2 ×R 2

× ψ1 (eλ r−ϑ F(s( α )T )−1 β∗ ) φ1 (F(s( α )T )−1 β∗ )

× e(λ +λ) [detF(s( α ))T ]−1 [detF(s( α ))T ]−1

! × ψ2 (eλ r−ϑ F(s( α )T )−1 β∗ ) φ2 (F(s( α )T )−1 β∗ ) .

(8.63)

The integral in α ∗ represents a delta measure in α − α which allows us to perform the α -integration. Next changing variables, β∗ → F(s( α )T )−1 β∗ , we obtain I (ψ1 , φ1 , ψ2 , φ2 ) = d 2 α ∗ d 2 β φ1 (β∗ ) φ2 (β∗ ) |β∗ |2 e2λ O∗

λ

× ψ1 (e r−ϑ

! ∗ λ ∗ β ) ψ2 (e r−ϑ β ) .

(8.64)

Noting that λ, ϑ are the components of α , we change variables as α → eλ r−ϑ β∗ to rewrite the α ∗ -integration. The determinant of this transformation is precisely |β∗ |2 e2λ , implying that I (ψ1 , φ1 , ψ2 , φ2 ) = ψ2 |ψ1 φ1 |φ2 ,

(8.65)

as asserted in (8.62). Finally, we again obtain in this case only one marginality condition for the Wigner function W (ψ| α ∗ , β∗ ) (obtained by taking φ = ψ in the cross-Wigner function): d 2 α ∗ 1 W (ψ | α ∗ , β∗ ) = |ψ(β∗ )|2 . (8.66) ∗ |2 2π R2 |β Note also the trace condition, d 2 α ∗ d 2 β∗ 1 W (ψ | α ∗ , β∗ ) = |ψ|2 . 2π O∗ β∗ 2

8.4.3

(8.67)

The Wigner function and the wavelet transform As already noted, there is a close connection between the Wigner function and the wavelet transform arising from a general Lie group. In this section we explicitly demonstrate this relationship for the SIM(2) group. The starting point is the resolution of the identity (7.45): 1 U (a, θ, b)|ψψ|U ∗ = I, dµ(a, θ, b) (a, θ, b) cψ = Cψ2 . (8.68) cψ SIM(2)

299

8.4 Wigner functions for the wavelet groups

Let s be a signal in the Fourier domain, i.e., an element of the Hilbert space L 2 (R2∗ , d 2 k). Its wavelet transform, corresponding to the mother wavelet ψ, is a, θ) = ψa,θ,b | s = U (a, θ, b)ψ | s. Sψ (b,

(8.69)

Using the group parameters as the phase space variables for the Wigner function (see (7.96)), we easily deduce √ 2π 2 i a, θ ) = W (Cψ, s|b, dλ dϑ d 2 β e− a [α·(rθ σ3 b) + β·eθ ] 2 (2π )2 a R 0 R ! 1 (cosh λ − cos ϑ) 2 λ × (8.70) Sψ (F(s( α ))β, e , ϑ) , | α| T T with eθ = cos θ, sin θ , α = ϑ, λ and F(s( α )) given, as before, by (7.71). Thus, the relationship between the wavelet transform and the Wigner function is an integral transform on the Hilbert space L 2 (SIM(2), dµ) of all square integrable (with respect to the Haar measure) functions of the group, and it is then straightforward to check that a, θ ) |Sψ (b, a, θ )|2 = a, θ) |W (Cψ, s | b, a, θ)|2 . (8.71) dµ(b, dµ(b, SIM(2)

SIM(2)

If the signal s is a Hilbert space vector which is in the domain of the inverse operator C −1 , then taking s = Cψ we also obtain √ 2π 2 i a, θ ) = W (s|b, dλ dϑ d 2 β e− a [α·(rθ σ3 b) + β·eθ ] 2 (2π ) a R 0 R2 ! 1 (cosh λ − cos ϑ) 2 λ e , ϑ) . SC −1 s (F(s( × α ))β, (8.72) | α| The above results show that any information about the signal which can be obtained using the wavelet transform, can also be derived using the Wigner function. The use of one rather than the other is therefore more a matter of practical convenience or computational ease, in any given situation.

9

Higher-dimensional wavelets

In the previous chapters, we have thoroughly discussed the 2-D CWT and some of its applications. Then we have made the connection with the group theoretical origins of the method, thus establishing a general framework, based on the coherent state formalism. In the present chapter, we will apply the same technique to a number of different situations involving higher dimensions: wavelets in 3-D space R3 , wavelets in Rn (n > 3), and wavelets on the 2-sphere S 2 . Then, in the next chapter, we will treat time-dependent wavelets, that is, wavelets on space–time, designed for motion analysis. In all cases, the technique is the same. First one identifies the manifold on which the signals are defined and the appropriate group of transformations acting on the latter. Next one chooses a square integrable representation of that group, possibly modulo some subgroup. Then one constructs wavelets as admissible vectors and derives the corresponding wavelet transform.

9.1

Three-dimensional wavelets Some physical phenomena are intrinsically multiscale and three-dimensional. Typical examples may be found in fluid dynamics, for instance the appearance of coherent structures in turbulent flows, or the disentangling of a wave train in (mostly underwater) acoustics, as discussed above. In such cases, a 3-D wavelet analysis is clearly more adequate and likely to yield a deeper understanding [56]. The same is true for many problems in astrophysics, such as galaxy/void counting or grouping, or cluster structure analysis. Hence we will also describe briefly the 3-D CWT, following the general pattern of the previous section.

9.1.1

Constructing 3-D wavelets Given a 3-D signal s ∈ L 2 (R3 , d 3 x), with finite energy, one may act on it by translation, dilation and rotation: a, ,)s ( sb,a,, ( x ) ≡ U (b, x ) = a −3/2 s(a −1 ,−1 ( x − b)), (9.1)

300

301

9.1 Three-dimensional wavelets

where b ∈ R3 , a > 0 and , ∈ SO(3). The 3 × 3 rotation matrix , ∈ SO(3) may be parametrized, for instance, in terms of three Euler angles. These three operations generate the 3-D Euclidean group with dilations, i.e., the similitude group of R3 , SIM(3) = R3 (R+ ∗ × SO(3)). Then (9.1) is a unitary representation of SIM(3) in L 2 (R3 , d 3 x), which is irreducible and square integrable, hence it generates a CWT exactly as before. Wavelets are taken in L 2 (R3 , d 3 x) and the admissibility condition is now d 3 k 2 |ψ(k)| < ∞. (9.2) 3 R3 |k| As in 2-D, a necessary, and almost sufficient, condition for admissibility is the familiar zero mean condition: d 3 x ψ( x ) = 0. (9.3) R3

The two standard wavelets have a 3-D realization. r The 3-D Mexican hat is given by ψH ( x ) = (3 − |A x|2 ) exp(− 12 |A x|2 ).

(9.4)

where A = diag[$1 −1/2 , $2 −1/2 , 1], $1 1, $2 1, is a 3 × 3 anisotropy matrix. We distinguish three cases: (1) If $1 = $2 = 1, one has the fully anisotropic 3-D Mexican hat (the stability subgroup Hψ is trivial); (2) If $1 = $2 = 1, one has the isotropic, SO(3)-invariant, 3-D Mexican hat (Hψ = SO(3)); (3) If $1 = $2 ≡ $ = 1, the wavelet is axisymmetric, i.e., SO(2)-invariant, but not isotropic (Hψ = SO(2)). Thus, in this case, wavelets are coherent states of Gilmore–Perelomov type, whose parameter space is indeed the quotient SO(3)/SO(2) S 2 (see the discussion in Section 7.1.5). r

The 3-D Morlet wavelet is given by ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) + corr.,

(9.5)

where A is the same 3 × 3 anisotropy matrix as in the first example. Here again, for $1 = $2 ≡ $ = 1 and ko along the z-axis, the wavelet ψ is invariant under SO(2) and we obtain coherent states of Gilmore–Perelomov type. r A 3-D Cauchy wavelet is defined by a straightforward generalization of the 2-D case, as follows. Consider the convex simplicial (or pyramidal) cone C(α, β, γ ) defined by the three unit vectors eα , eβ , eγ , the angle between any two of them being smaller ˜ γ˜ ), where eα˜ = eβ ∧ eγ than π. The dual cone is also simplicial, namely C = C(α, ˜ β, is orthogonal to the β-γ face, etc. With these notations, given a vector η ∈ C and l, m, n ∈ N∗ , we define a 3-D Cauchy wavelet in spatial frequency space as: (k · eα˜ )l (k · eβ˜ )m (k · eγ˜ )n e−k·η , k ∈ C(α, β, γ ), (C,η) (k) (9.6) = ψ lmn 0, otherwise.

302

Higher-dimensional wavelets

As in the 2-D case, the factors (k · eα˜ )l , etc. represent vanishing moments on the faces of the cone, and thus determine the regularity of the wavelet in k-space. Here too the expression for the 3-D wavelet in position space may be obtained explicitly: l+m+n vol [eα , eβ , eγ ] i l+m+n+3 (C,η) ψlmn ( x) = , (9.7) l! m! n! · det A · 2π (z · eα )l+1 (z · eβ )m+1 (z · eγ )n+1 where A is the matrix that transforms the unit vectors e1 , e2 , e3 into the triple eα , eβ , eγ , vol [·, ·, ·] denotes the volume of the parallelepiped generated by the three vectors, and we have written z = x + i η. Note that the direct calculation, following the pattern of the 2-D case (Section 3.3.4) yields for the numerator in (9.7) the expression (eα˜ · eα )l (eβ˜ · eβ )m (eγ˜ · eγ )n , but then one has eα˜ · eα = eβ ∧ eγ · eα = eβ · eγ ∧ eα = eβ · eβ˜ = eγ · eγ˜ = vol [eα , eβ , eγ ], which proves (9.7). From the expressions (9.6) and (9.7), one may then obtain other 3-D Cauchy wavelets, for instance, one supported in a circular cone. Take a circular convex cone, aligned on the positive k z -axis, with total opening angle 2θo (0 < θo < π/2). In spherical polar coordinates k = (|k|, θ, φ), the interior of the cone is simply C(θo ) = {k ∈ R3 | θ θo }. Then an axisymmetric 3-D Cauchy wavelet supported in this cone may be defined, for instance, by |k|l (tan2 θo − tan2 θ)m e−|k| cos θ , 0 θ θo ; (θo ) ψm (k) = (9.8) 0, otherwise. Again m ∈ N defines the number of vanishing moments on the surface of the cone, that is, the regularity of the wavelet. For θo very small, this wavelet lives inside a narrow pencil: it clearly evokes the beam of a searchlight – a vivid illustration of the wavelet as a directional probe! If we note that the expression on the right-hand side of (9.8) may be written as

|k|l (tan2 θo k z2 − k x2 + k 2y )m e−kz , we see that all these wavelets are built on the same = 0 is the equation of the cone. m e−kz , where F(k) model, namely F(k)

9.1.2

The 3-D continuous wavelet transform Then, given a signal s ∈ L 2 (R3 ), its CWT with respect to the admissible wavelet ψ is given as a, ,) = a −3/2 s( S(b, d 3 x ψ(a −1 ,−1 ( x − b)) x) . (9.9) R3

As compared with (2.19), the only differences are in the normalization factors and the rotation matrices. Since the structure of the formulas is the same as before, so are the

303

9.1 Three-dimensional wavelets

(a)

(b)

Fig. 9.1. The 3-D cube and its Fourier transform.

interpretation and the consequences (local filtering, reproducing kernel, reconstruction formula, etc.). Thus the CWT (9.9) may be interpreted as a mathematical camera with magnification 1/a, position b and directional selectivity given, in the axisymmetric case, a, ,) by the rotation parameters ζ ≡ (θ, ϕ). As for the visualization, the full CWT S(b, is a function of seven variables. However, if the wavelet ψ is chosen axisymmetric, i.e., SO(2)-invariant, S depends on six variables only, b ∈ R3 , a > 0, and ζ ∈ S 2 SO(3)/SO(2), the unit sphere in R3 . In this case again, (a −1 , ζ ) may be interpreted as polar coordinates in spatial frequency space. This is in fact true in any number of dimensions. It follows that, here too, there are two natural representations for the visualization of the WT: the position representation (a, ζ fixed) and the scale-orientation (or spatial frequency) representation (b fixed). Of course, there are many other possible representations that may be useful. As an example, we present in Figure 9.1 (the characteristic function of) a 3-D cube and its Fourier transform (note the occurrence of sinc functions among all three coordinate axes), then in Figure 9.2 the wavelet transform of the cube signal, with a Morlet wavelet, at four successively finer scales, a = 1, 0.5, 0.25, 0.125. As in 2-D, the net result, for a sufficiently fine scale, is the contour of the cube. The latter has become totally transparent, only the edges survive! A similar, slightly more complicated example is that of a cube with a small box removed, see Figure 9.3. Here one faces the well-known visual ambiguity: depending of the angle of view, the small cube appears either as removed (concave) or added (convex). As shown in Figure 9.4, the ambiguity is resolved with the WT, the part that has been removed yields a negative WT, whereas an added, convex, portion would yield a positive contribution to the WT. This is true, of course, if one uses a real wavelet (here the 3-D Mexican hat) and plots the WT itself, not its modulus. We recover exactly the 2-D situation, discussed in Section 4.1.2.

304

Higher-dimensional wavelets

(a)

(c)

(b)

(d)

Fig. 9.2. Wavelet transform of the 3-D cube at four successively finer scales.

(a) Fig. 9.3. The 3-D cube minus a box and its Fourier transform.

(b)

305

9.1 Three-dimensional wavelets

(a)

(c)

(b)

(d)

Fig. 9.4. Wavelet transform of the 3-D cube minus a box at four successively finer scales (a = 1,

0.5, 0.25, 0.125).

A practical application of this 3-D CWT is the detection of 3-D objects in a cluttered medium. We consider a scene with 3-D objects (targets) immersed in a cluttered medium, modeled by the signal: s( x) =

L

sl ( x ) + n( x ),

(9.10)

l=1

where sl ( x ) denotes the density of the target l, and n( x ) the density of the medium. Since the density of the targets is very different from that of the medium, there will be a high density gradient at the boundary between the objects and the medium. In this a, ζ ) may be used for extracting the 3-D objects and situation, the wavelet transform S(b, determining their characteristics, position (range and orientation) and spatial frequency. This is, of course, nothing but a 3-D version of the ATR problem discussed already in Section 4.2 (see also [16]). In particular, 3-D directional wavelets (e.g. Morlet or Cauchy) will behave as their 2-D counterparts. If a 3-D image s( x ) contains elongated

306

Higher-dimensional wavelets

a, ζ ) shows “sausage” images for all the objects which objects, a visualization of S(b, have the same direction as the wavelet, and small spheres at the tips of all the objects which are misaligned. With an appropriate thresholding, we may filter out all the objects which are not in the desired direction. Another domain of application is in computational fluid dynamics. For instance, the technique has been used in [56] for analyzing two types of simulated 3-D flows: (i) detection of elongated vortex structures in homogeneous 3-D turbulence, and (ii) interaction of a shock wave with a 3-D supersonic mixing layer between two parallel streams. In both situations, the use of scale as a fundamental variable offers distinct advantages over traditional methods based on sophisticated graphic tools. In the first case, the selection of structures is based on characteristic scales rather than on vorticity values. For the shear/shock interaction, the simultaneous detection of main steady flow features with unsteady multiscale mixing features is a practical way to the study of compressibility controlled mixing phenomena. Clearly this opens interesting perspectives for wavelet analysis in fluid dynamics. As a final remark, we may note that the 3-D discrete WT has also been used successfully in several applications. One of them is the analysis of the substructure of galaxy clusters [300]. Another one is the lossless compression of 3-D images produced by computer tomography or radar scanning [81]. The lossless character of the transform is based, as in 1-D or 2-D, on the use of an integer WT (see Section 2.5.2.4).

9.1.3

Extension to higher dimensions From the analysis of the previous section, it should be clear that the extension to higher dimensions is straightforward – although it has probably an academic interest only (except perhaps in quantum mechanics, since an N -particle wave function belongs to L 2 (R3N , d 3N x)). Given a finite energy signal s ∈ L 2 (Rn , d n x), the transformation group to consider is again the similitude group, now SIM(n) = Rn (R+ ∗ × SO(n)). This group has a (unique) natural unitary irreducible representation in L 2 (Rn , d n x), given by sb,a,, ( x ) ≡ U ( b, a, ,)s ( x ) = a −n/2 s(a −1 ,−1 ( x − b)), (9.11) where b ∈ Rn is a translation, a > 0 is a global dilation and , ∈ SO(n) is a rotation. This representation is square integrable, with admissibility condition d n k 2 |ψ(k)| < ∞ , (9.12) n Rn |k| reducing as usual to the necessary condition Rn d n x ψ( x ) = 0. The n-D wavelet transform is, as before, s( a, ,) = a −n/2 d n x ψ(a −1 ,−1 ( x − b)) x) . (9.13) S(b, Rn

307

9.1 Three-dimensional wavelets

Standard wavelets are the same, namely, r The n-D Mexican hat: ψH ( x ) = (n − |A x|2 ) exp(− 12 |A x|2 ). r

The n-D Morlet wavelet: ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) + corr.

Once again, the anisotropy matrix A = diag[$1 −1/2 , $2 −1/2 , . . . , $n−1 −1/2 , 1], $1 1, $2 1, . . . , $n−1 1, leads to various situations, depending on the number of different values taken by the parameters $ j . The interesting case is, of course, the axisymmetric or SO(n − 1)-invariant case, where all the $ j coincide and $ j = $ = 1. n−1 Then, the parameter space reduces to Rn × R+ Rn × Rn∗ and the WT reads ∗ ×S a, ζ ), ζ ∈ S n−1 . As before, (a −1 , ζ ) may be interpreted as polar coordinates in S(b, spatial frequency space, and thus the n-D CWT yields a phase space representation of signals. The rest is as before. Cauchy wavelets also extend to Rn , as follows. We start with the n-simplex C(e1 , e2 , . . . , en ), generated by the unit vectors e1 , e2 , . . . , en . Define successively e1˜ = e2 ∧ e3 ∧ . . . ∧ en−1 ∧ en e2˜ = e3 ∧ e4 ∧ . . . ∧ en ∧ e1 ... en˜ = e1 ∧ e2 ∧ . . . ∧ en−2 ∧ en−1 . Then the only nonzero inner products are e1˜ · e1 = e2˜ · e2 = . . . = en˜ · en = vol [e1 , e2 , . . . , en ]. ˜ we define the n-dimensional Cauchy wavelet as Thus, for η ∈ C, 6n η j˜ )l j e−k· , k ∈ C(e1 , e2 , . . . , en ), j=1 (k · e (C,η) ψ ( k) = l1 l2 ...ln 0, otherwise,

(9.14)

and, in position space,

(C,η) ψl1 l2 ...ln ( x)

l +l +...+ln vol [e1 , e2 , . . . , en ] 1 2 6n = const . z · e j )l j +1 j=1 (

(9.15)

In these formulas, l1 , l2 , . . . , ln ∈ N∗ and the factors (k · e j˜ )l j represent vanishing moments on the faces of the cone. Exactly as in three dimensions, one may also construct Cauchy or conical wavelets = 0 and, in particular, n-D axisymmetric adapted to a general cone of equation F(k) wavelets in the case of a circular cone.

308

Higher-dimensional wavelets

9.2

Wavelets on the 2-sphere and other manifolds

9.2.1

The problem In most cases of physical interest, experimental data are given on the line (signal processing), on the plane (image analysis), or occasionally in R3 (e.g., in fluid dynamics). However, there are situations where data are given on a sphere, for instance, in geophysics or astronomy, of course, but also in statistical problems [Fis87], computer vision or medical imaging (see [216] for precise references). If one is interested only in very local features, one may forget the curvature and work on the tangent plane with standard methods, that is, Fourier or time–frequency analysis, in particular the CWT. However, when global aspects become important (description of plate tectonics on the Earth, for instance, or structure of the Universe as a whole), curvature can no longer be ignored, so that one needs a genuine generalization of wavelet analysis to the sphere (or more general manifolds). Let us first make that statement precise. We may speak of a genuine spherical CWT if (i) the signals and the wavelets live on the sphere; (ii) the transform involves (local) dilations of some kind; and (iii) for small scales, the spherical CWT reduces to the usual CWT on the (tangent) plane (Euclidean limit). Several authors have studied this problem, with various techniques, mostly discrete. For instance: r One may extend to S 2 the discrete wavelet scheme based on a multiresolution analysis, but this approach leads often to numerical difficulties around the poles [114,197,318]. r A different technique is to use the lifting scheme and second generation wavelets, as described in Chapter 2, Section 2.5.2. An efficient solution has been obtained in this way by Schr¨oder and Sweldens [336], but this obviously misses the particular symmetry of the sphere. r One may exploit the geometry of the sphere, as encoded in the system of spherical harmonics [Fre97,Fre99,169,294,319], but the resulting analyzing functions are poorly localized (in fact they do not really resemble wavelets). On the other hand, this approach leads to good approximation methods for spherical functions; we shall come back to this problem in Section 9.3. However, to fully preserve the rotational invariance of the sphere, a continuous approach is clearly necessary. Here too, several authors have proposed solutions. r Considering the fact that the sphere does not admit global dilations, since it is compact, one resorts to a wavelet transform on the tangent bundle of the sphere [116] or, instead, to a Gabor transform on the sphere itself [355,356]. r The most satisfactory approach is that of Holschneider [225], who produces a CWT on the sphere that satisfies the three criteria above. However, the rˆole of dilation is

309

9.2 Wavelets on the 2-sphere and other manifolds

played by an abstract parameter that satisfies a number of ad hoc assumptions. The correct Euclidean limit is obtained, but it is essentially put by hand. As can be seen from this brief description, none of the proposed solutions fully qualifies for a genuine CWT on S 2 . It turns out that the general formalism developed in [Ali00,6] and sketched in Chapter 7 yields an elegant solution to the problem, entirely derived from group theory, and in particular allows one to derive all the assumptions of [225]. Although the discussion is too technical to be given here in detail, it is interesting to outline the main ideas, because they lead to significant generalizations. A detailed treatment may be found in [23,28,29].

9.2.2

The continuous wavelet transform on S 2

9.2.2.1

Affine transformations on the sphere S2 We consider the 2-sphere S 2 , with polar spherical coordinates ζ = (θ, ϕ). As usual, finite energy signals are taken as square integrable functions on the 2-sphere, s ∈ H = L 2 (S 2 , dµ), where dµ(ζ ) ≡ dµ(θ, ϕ) = sin θ dθ dϕ is the usual (rotation invariant) measure on S 2 . The first step for constructing a CWT on S 2 is to identify the natural operations on such signals. These are of two types: (i) Motions or displacements, given by elements of the rotation group SO(3), which indeed acts transitively on S 2 , and S 2 SO(3)/SO(2). (ii) Dilations, that may be derived in two steps. First, dilations around the North Pole are obtained by considering usual dilations in the tangent plane and lifting them to S 2 by inverse stereographic projection from the South Pole. This gives: Da(N ) (θ, ϕ) = (θa , ϕ),

with

tan

θ θa = a · tan . 2 2

(9.16)

Then a dilation around any other point ζ ∈ S 2 is obtained by moving ζ to the North Pole by a rotation , ∈ SO(3), performing a dilation Da(N ) as before and going back (ζ ) by the inverse rotation: Da = ,−1 Da(N ) ,. Clearly the dilations act also transitively on S 2 . Next we have to identify a group of affine transformations on S 2 . First we note that motions , ∈ SO(3) and dilations by a ∈ R+ ∗ do not commute. Also it is impossible to build a semidirect product from SO(3) and R+ ∗ , and therefore the only extension of + SO(3) by R∗ is their direct product. A way out is to embed the two factors into the Lorentz group S Oo (3, 1), by the so-called Iwasawa decomposition: SOo (3, 1) = SO(3) · A · N ,

(9.17)

where A ∼ SOo (1, 1) ∼ R ∼ R+ ∗ is the subgroup of Lorentz boosts in the z-direction and N ∼ C is two-dimensional and abelian (under the stereographic projections, N corresponds to translations in the tangent plane). The appearance of the Lorentz group

310

Higher-dimensional wavelets

SOo (3, 1) in this context is not fortuitous, since it is the conformal group of the sphere S 2 (and of the plane R2 as well). It turns out that the stability subgroup of the North Pole is the so-called minimal parabolic subgroup P = M = SO(2) · A · N , where M = SO(2) is the subgroup of rotations around the z-axis. Thus we get S 2 SOo (3, 1)/P SO(3)/SO(2).

(9.18)

This shows that SOo (3,1) acts transitively on S 2 as well. This action may be computed explicitly using the Iwasawa decomposition (9.17). For a pure dilation by a, the result is precisely the usual dilation lifted on S 2 by inverse stereographic projection, given in (9.16).

9.2.2.2

Spherical wavelets The next step towards constructing wavelets (affine coherent states) on S 2 is to find a suitable unitary irreducible representation (UIR) of the Lorentz group SOo (3, 1) acting in the Hilbert space L 2 (S 2 , dµ). Natural candidates are the representations of the continuous principal series, also called class I representations [Kna96,351]. The simplest one, that we shall use, is given by the operators: (9.19) [U (g) f ] (ζ ) = λ(g, ζ )1/2 f g −1 ζ , g ∈ SOo (3, 1), f ∈ L 2 (S 2 , dµ), where g = ,an by the Iwasawa decomposition and the multiplier λ (g, ζ ) is a Radon– Nikodym derivative (or a 1-cocycle), expressing the fact that the measure dµ is not invariant under the full group SOo (3, 1): dµ g −1 ζ λ (g, ζ ) = (9.20) , g ∈ SOo (3, 1). dµ (ζ ) This representation U of SOo (3, 1) is unitary and irreducible. Since we are only interested in the action of dilations and motions, we quotient out the subgroup N . In other words, the parameter space of the spherical wavelets is X = SOo (3, 1)/N SO(3) · R+ ∗ . Then, introducing a suitable section σ : X = SOo (3, 1)/N → SOo (3, 1), we concentrate on the reduced expression (9.21) [U (σ (x)) f ] (ζ ) = λ(σ (x), ζ )1/2 f σ (x)−1 ζ , x ≡ (,, a). We choose the natural (Iwasawa) section σ (,, a) = , a, , ∈ SO(3), a ∈ A. Using the action (9.16) of dilations, one gets easily 4a 2 λ(σ (,, a), ζ ) ≡ λ(a, θ) = $ %2 , a 2 − 1 cos θ + a 2 + 1

ζ = (θ, ϕ).

(9.22)

The function λ(a, θ) satisfies the so-called cocycle relation (which guarantees that U is indeed a representation): λ(a −1 , θ ) λ(a, θa ) = λ(1, θ) = 1.

(9.23)

311

9.2 Wavelets on the 2-sphere and other manifolds

In addition, from the choice of the section, we have U (σ (,, a)) = U (, a) = U (,)U (a), and therefore the representation (9.21) factorizes as U (σ (,, a)) = R, Da .

(9.24)

In this relation, R, ≡ Uqr (,), , ∈ SO(3), where Uqr is the quasi-regular representation of S O(3) in L 2 (S 2 , dµ), and Da , a ∈ R+ ∗ , is an operator of pure dilation, that is, (Da f ) (ζ ) = λ(a, θ)1/2 f (ζ1/a ), with ζa ≡ (θa , ϕ). The quasi-regular representation of SO(3), (Uqr (,) f )(ζ ) = f (,−1 ζ ), is infinite dimensional and decomposes into the direct sum of all the familiar (2l + 1)-dimensional representations, l = 0, 1, 2, . . . . Following the general approach of [Ali00,6], we build now a system of spherical wavelets, realized as coherent states for the Lorentz group, indexed by points of X = SOo (3, 1)/N . Since N is not the isotropy subgroup of a particular vector in the representation Hilbert space, the resulting coherent states are not of the Gilmore– Perelomov type [Per86] (see Section 7.1.5). First we show that the UIR (9.21) is indeed square integrable on X , that is, we check that there exists a nonzero vector ψ ∈ L 2 (S 2 , dµ) such that dν(,, a) |U (σ (,, a))ψ|φ|2 < ∞, ∀ φ ∈ L 2 (S 2 , dµ), X

where dν(,, a) = a −3 d, da, and d, is the invariant (Haar) measure on SO(3). Proposition 9.2.1 The representation U given in (9.21) is square integrable modulo the subgroup N and the section σ . A nonzero vector ψ ∈ L 2 (S 2 , dµ) is admissible mod(N , σ ) iff there exists c > 0, independent of l, such that, for all l 0: l ∞ da ( 8π 2 |ψa (l, m)|ψa |2 < c . Gl ≡ 2l + 1 m=−l 0 a 3

(9.25)

Here f (l, m) ≡ Ylm | f denotes a Fourier coefficient of f ∈ L 2 (S 2 ), with Ylm the usual spherical harmonic, and ψa = Da ψ = U (σ (e, a))ψ corresponds to a pure dilation. The proof, as usual, consists in an explicit calculation, using the properties of Fourier analysis on the sphere. Thus any admissible ψ generates a continuous family {ψa,, ≡ U (σ (,, a))ψ, (,, a) ∈ X } of spherical wavelets, but in fact we have more. 2π Proposition 9.2.2 For any admissible vector ψ such that 0 dϕ ψ(θ, ϕ) ≡ 0 (for instance, axisymmetric), the family {ψa,, , (,, a) ∈ X } is a continuous frame, that is, there exist constants A > 0 and B < ∞ such that 2 A φ dν(,, a) |ψa,, |φ|2 B φ2 , ∀ φ ∈ L 2 (S 2 , dµ). (9.26) X

312

Higher-dimensional wavelets

Thus, for most admissible vectors ψ, we get a continuous frame, but not necessarily a tight frame. We conjecture that the resulting frame is never tight, that is, A = B.

9.2.2.3

The spherical wavelet transform Proposition 9.2.1 yields the basic ingredient for writing the CWT on S 2 . Given an admissible vector ψ ∈ L 2 (S 2 , dµ), our wavelets on the sphere are the functions ψa,, = U (σ (,, a))ψ = R, Da ψ = R, ψa .† Then, the spherical CWT of a signal s ∈ L 2 (S 2 ) is defined as S(,, a) = ψa,, |s = dµ(ζ ) ψa,, (ζ ) s(ζ ) 2 S dµ(ζ ) ψa (,−1 ζ ) s(ζ ). =

(9.27)

S2

This relation gives the spherical CWT as a convolution on the sphere S 2 : ψa (ζ ) and s(ζ ) are two functions on S 2 , and their convolution ψa % s is a function on SO(3). Such a formulation leads both to mathematical subtleties and to numerical difficulties. The former will be treated in Section 9.3 (see also Appendix A) and the latter in Section 9.2.5, where we will discuss the numerical implementation of the spherical CWT. According to the general theory, the admissibility of the wavelet ψ is sufficient to guarantee the invertibility of the spherical CWT on its range, that is, we can reconstruct the signal s(ζ ) from its wavelet transform S(,, a): s(ζ ) = dν(,, a) S(,, a) A−1 (9.28) σ ψa,, (ζ ) . X

In this relation, Aσ denotes the resolution operator (7.34), whose action is a multiplication in Fourier space (as it is the case with most Duflo–Moore operators, for instance, (7.42)), A σ f (l, m) = G l f (l, m), with G l defined in the admissibility condition (9.25). As usual, the integral in (9.28) is to be taken in the weak sense. Of course, as in the flat case, one may consider more general reconstruction formulas, with two different wavelets for the analysis and the synthesis (see (2.31) in Section 2.2). These formulas gets simpler if the wavelet ψ is axisymmetric, i.e., ψ(ζ ) ≡ χ(θ), for, then, we may exploit the fact that S 2 SO(3)/SO(2). First, since the dilation is purely radial, ψa is also axisymmetric and ψa (ζ ) = χa (θ ). Then the action of R, on ψa has for sole effect to transport its center from the North Pole ζo to some point ζ = , ζo . †

As the notation ψa,, suggests, all the operations involved in the CWT consist in manipulating the function ψa at a fixed scale a. This is consistent with [Fre97,169] and [35], but not with [29], where the same wavelets were denoted ψ,,a .

313

9.2 Wavelets on the 2-sphere and other manifolds

We can thus characterize R, ψa by its center point ζ , which is independent of a, and we may write ψa,, ≡ ψa,ζ . A more precise way of achieving this is to split , ∈ SO(3) into , = (χ , [ζ ]) with χ ∈ SO(2) and ζ ∈ S 2 . This is formally done through a projection , → ζ (,) in the fiber bundle SO(3) → S 2 SO(3)/SO(2), followed by an arbitrary choice of section ζ → [ζ ] in S O(3). The splitting corresponds to decomposing the motion R, of the wavelet ψa into an initial rotation of angle χ around the North Pole ζ0 followed by a transport to the point ζ = , ζo on the sphere. In other words, R, ψa (ζ ) = Rχ ψa ([ζ ]−1 ζ ) where Rχ is a rotation around the North Pole. Accordingly, the spherical wavelet transform will also be denoted by S(χ, ζ , a). This amounts simply to decomposing the parameter space SO(3) × R+ ∗ into a more appropriate form, including “translations” ζ , dilations a, and “rotations” χ , exactly as in the Euclidean case. Of course, when the wavelet ψ is axisymmetric, the dependence on χ can be dropped and the spherical wavelet transform will be written simply as S(ζ , a): dµ(ζ ) ψa,ζ (ζ ) s(ζ ). (9.29) S(ζ , a) = S2

(A related statement is given in Proposition 9.3.1, in Section 9.3.) In that case, the −3 parameter space of the spherical CWT reduces to S 2 × R+ dµ(ζ ) da. ∗ , with measure a Hence, the integral over SO(3) in the reconstruction formula is replaced by an integral over S 2 : dµ(ζ ) da s(ζ ) = S(ζ , a) A−1 (9.30) σ ψa,ζ (ζ ). 3 + 2 a S R∗ According to the general coherent state formalism, the reconstruction formulas (9.28) and (9.30) are valid only in the weak sense. In the flat case, however, we have seen in Section 2.6 that the corresponding formula holds in the strong L 2 sense. This guarantees that it can be used for approximating functions on the plane through an approximate identity. That means, the approximating function is obtained by convolution with a smoothing kernel, which tends to the identity (δ function) as the parameter goes to 0. We will show in Section 9.3 that exactly the same situation prevails on the sphere. In order to prove these results, we will have to switch to an L 1 formalism (as already mentioned in [29]), by introducing a modified dilation operator D a that preserves the L 1 norm of functions. As a consequence, we will have at our disposal two types of spherical CWT. An important aspect of the flat space CWT is its covariance property (Proposition 2.2.3). In the present case, an explicit calculation [29] shows that the spherical CWT (9.27) is covariant under motions on S 2 , but not covariant under dilations. r It is covariant under motions on S 2 , namely, for any , ∈ SO(3), the transform of the o rotated signal s(,o−1 ζ ) is the function S(,o−1 ,, a).

314

Higher-dimensional wavelets r

But it is not covariant under dilations. Indeed the wavelet transform of the dilated signal λ(ao , ζ )1/2 s(ao−1 ζ ) is U (g)ψ|s, with g = ao−1 ,a, and the latter, while a welldefined element of SOo (3, 1), is not of the form σ (, , a ). For applications, of course, it is the covariance under motions that is essential, since it reduces to translation covariance in the Euclidean limit, as we shall see in Section 9.2.3. As for dilations, the negative result reflects the fact that the parameter space X SO(3) × R+ ∗ of the spherical CWT is not a group. The condition (9.25), which was derived in [225] in a different way, is necessary and sufficient for the admissibility of ψ, but it is somewhat complicated to use in practice, since it requires the evaluation of nontrivial Fourier coefficients. Instead, there is a simpler, although only necessary, condition. Proposition 9.2.3 A function ψ ∈ L 2 (S 2 , dµ) is admissible only if it satisfies the condition ψ(θ, ϕ) dµ(ζ ) = 0, ζ ≡ (θ, ϕ). (9.31) 2 1 + cos θ S This necessary condition is the exact equivalent of the usual necessary condition for wavelets in the plane, d 2 x ψ( x ) = 0, and it reduces to the latter in the Euclidean limit (see Section 9.2.3). The interesting point is that (9.31) is a zero mean condition, as in the flat case. As such it ensures that the CWT on S 2 given in (9.27) acts as a local filter. This is crucial for applications and it is one of the main reasons of the efficiency of the CWT, and the same holds here. Using Proposition 9.2.3, it is easy to build explicit wavelets on the sphere, namely “Difference wavelets”, similar to the ones described for the flat case in Chapter 3, Section 3.2.2. For that purpose, we notice the following easy result [28]. Proposition 9.2.4 Let φ ∈ L 2 (S 2 , dµ). Then 1 Da φ(θ, ϕ) φ(θ, ϕ) = , dµ(ζ ) dµ(ζ ) a S2 1 + cos θ 1 + cos θ S2 where Da = U (σ (e, a)) is again a pure (covariant) dilation. Given a square integrable (smoothing) function φ, we define ψφ(α) (θ, ϕ) = φ(θ, ϕ) −

1 Dα φ(θ, ϕ) α

(α > 1).

(9.32)

By Proposition 9.2.4, ψφ(α) satisfies the admissibility condition (9.31), that is, it is a spherical wavelet, and it is fully admissible if φ is sufficiently regular at the poles. The simplest difference wavelet is obtained with the choice φG (θ, ϕ) = exp(− tan2 θ2 ), which is essentially the inverse stereographic projection of a Gaussian in the tangent plane. The corresponding spherical wavelet, called the spherical DOG wavelet and

9.2 Wavelets on the 2-sphere and other manifolds

1 1

0.5

0.5

Z

Z 0

0

−0.5

−0.5

−1

−1 −1

−0.5

−0.5

−0.5

0 0.5

−1 0

0

−0.5 0.5

0.5

0 1

0.5

X

Y

X

Y

1

0.5

Z

315

0

−0.5

−1 −1 −0.5

−0.5 0

0 0.5

0.5 1

X Y

(α)

Fig. 9.5. The spherical DOG ψG wavelet, for α = 1.25. (Top line) The wavelet at scale a = 0.125

and positioned at the North Pole θ = 0◦ (left) and on the equator θ = 90◦ , ϕ = 90◦ (right); (Bottom) The same at scale a = 0.0625 and position θ = 90◦ , ϕ = 0◦ . As mentioned in the text, “at scale a” means the function Da ψφ(α) .

denoted ψG(α) in the sequel, is an axisymmetric wavelet, which is shown in Figure 9.5, for different values of the scale a and in various positions (θ, ϕ) on the sphere. Note that here “ψ at scale a” means that the function being plotted is Da ψ, i.e., one must always use the covariant dilation operator Da . This wavelet yields an efficient detection of discontinuities on the sphere [28]. Explicit examples of spherical wavelet transforms based on it will be given in Section 9.2.5.3. Now, the construction of the scaling function φG by inverse stereographic projection from the tangent plane suggests a general procedure for generating spherical wavelets.

316

Higher-dimensional wavelets

We will implement it in Section 9.2.4, where we will construct directional wavelets on the sphere by the same method. Before going into that, let us discuss the Euclidean limit of our spherical wavelet transform.

9.2.3

The Euclidean limit As said above, a good wavelet transform on the sphere should be asymptotically Euclidean, that is, the spherical WT should match the usual CWT in the plane (in this case, the tangent plane at the North Pole) at small scales or, what amounts to the same, for large values of the radius of curvature. This statement may be given a precise mathematical meaning, using the technique of group contractions (or deformations). Without entering into technical details, we sketch the successive steps. First, we reformulate the theory on a sphere of radius R and let R → ∞. Then SR2 becomes the plane R2 , the group SO(3) contracts into the Euclidean group of R2 and the Lorentz group SOo (3, 1) into the (semidirect) product G E = R2 SIM(2), where 2 SIM(2) = R2 (R+ ∗ × SO(2)) is the similitude group of R , that is, the invariance group of the Euclidean CWT, discussed in Chapter 7, Section 7.1.2. Notice that the contraction preserves the minimal parabolic subgroup P = M AN ∼ SIM(2), and, in particular, the subgroup SO(2) of rotations around the z-axis. The next step is to transfer the contraction process to the relevant homogeneous spaces. On one hand, the manifolds S 2 = SOo (3, 1)/M AN and R2 = G E /M AN , that carry the respective CWT, are related through contraction. On the other hand, since the abelian subgroup N is preserved under the contraction, the parameter space X = SOo (3, 1)/N SO(3) × A of the spherical CWT goes into that of the Euclidean CWT, namely SIM(2) = G E /N . Notice that the former is not a group (and this forces us to use the general formalism of [Ali00,6], described in Chapter 7), whereas, after contraction, we get SIM(2), that is, the missing group structure is restored by the contraction! In this geometrical context, the Euclidean limit itself can be formulated as a contraction at the level of group representations. Whereas contractions of Lie algebras and Lie groups are relatively ancient and well-known [232,331], the extension of the procedure to group representations is rather recent [272]. A rigorous version has been given by Dooley [150–152], that was followed in [23,29]. The additional difficulty here is that the representation space itself varies during the procedure. Let H R = L 2 (S R2 , dµ R ) be the Hilbert space of square integrable functions on a sphere of radius R (where dµ R = R 2 dµ) and H = L 2 (R2 , d 2 x). The two spaces are related by the unitary map I R : H R → H, given as (I R f ) (r, ϕ) =

4R 2 r , ϕ , f 2 arctan 4R 2 + r 2 2R

(9.33)

317

9.2 Wavelets on the 2-sphere and other manifolds

where (r, ϕ) are polar coordinates in the plane. Clearly, the map I R just describes the stereographic projection of the sphere S R2 onto its tangent plane at the North Pole. The inverse map, that is, the inverse stereographic projection, reads as (I R−1 f )(θ, ϕ) =

θ 2 f (2R tan , ϕ), 1 + cos θ 2

(9.34)

(in the case R = 1, this map was given already in (5.1)). Now let U be the usual wavelet representation (2.13) of SIM(2) in H and U R the representation (9.21) of SOo (3, 1) realized in H R . For each R, we choose the corresponding representation U R . Then the precise statement is that the representation U of SIM(2) is a contraction, in the sense of Dooley [150–152] of the family of representations U R of SOo (3, 1) as R → ∞. This means that, for every g ∈ SIM(2), the following strong limit holds in H: ˜ R (g) I R−1 φ − U (g)φH = 0, lim I R U R (9.35) R→∞

˜ R : SIM(2) → X is the so-called reduced contraction map (see [29] for details). where This theorem yields the expected result that local wavelet analysis on the sphere as defined here is equivalent to local wavelet analysis in flat space. Indeed the whole structure on the sphere S R2 goes into the corresponding one in R2 as R → ∞. Since U R → U , the corresponding matrix elements converge to one another, and so the square integrability condition (9.25) converges into the corresponding one for the CWT in R2 , namely k)| 2 |ψ( d 2 k < ∞. 2 |k| R2 Admissible wavelets on S 2 converge to admissible wavelets on R2 (Proposition 9.2.5), and the necessary condition (9.31) also goes into the usual one in the plane, namely d 2 x ψ( x ) = 0. R2

9.2.4

Directional wavelets on the sphere The general mathematical setting for designing directional wavelets on S 2 has been introduced in the previous sections, but the construction and the properties of these particular spherical wavelets have not been discussed. Yet these wavelets are quite important in practice, since directional features (roads, streams, geological faults, . . . ) abound on the spherical Earth! Thus one really needs the additional degree of freedom they offer for characterizing signals.

9.2.4.1

General remarks Whenever the wavelet ψ is not axisymmetric, the continuous spherical wavelet transform depends on the additional parameter χ , the rotation angle around the North Pole, thus (9.27) may be rewritten as

318

Higher-dimensional wavelets

S(χ , ζ , a) =

S2

dµ(ζ ) Rχ ψa ([ζ ]−1 ζ ) s(ζ ).

(9.36)

In this formula, there is an arbitrariness in the way the rotation [ζ ] of S O(3) is associated to the point ζ on the sphere. The section [·] : S 2 → S O(3) can be depicted as mapping the sphere to a tangent vector field of unit length defined on it. Intuitively, one would like this mapping to be smooth to correspond to the idea of direction defined on the sphere, and the arbitrariness in the choice of the section may be exploited to that effect, at least locally. When this is the case, we expect that the values of the wavelet transform correspond to filtering in a given direction χ and at a given scale a as in the case of the 2-D wavelet transform in the plane. Some caution must be exercised, however, when dealing with directions on the sphere. Indeed, it is a classical result in topology that there exists no differentiable vector field of constant norm on S 2 , which means there is no global way of defining directions. There will always be some singular point where the definition fails.† In other words, one cannot comb a perfectly spherical porcupine! Nevertheless, testing orientations using directional wavelets is a small-scale operation, that is, a local procedure [24]. Then, around any point on S 2 , the support of the wavelet at small scale defines a neighborhood in which the following reasoning holds. This ability to perform local analysis is definitely one the most important properties of wavelet analysis. We will see more about this in the examples below. From now on, we will make use of the classical parametrization of S O(3) in terms of Euler angles, , ≡ (χ, θ , ϕ ), which corresponds to the choice of section (θ , ϕ ) → (0, θ , ϕ ), which in turn defines a direction on the sphere. The singular points are the North and South Poles: it makes no sense to define cardinal points at the poles! For this choice of parametrization, we may write Rχ ψa ([ζ ]−1 ζ ) = ψa,χ,ζ (ζ ) ≡ ψa,χ,θ ,ϕ (θ, ϕ),

(9.37)

which implies ψa,χ ,θ ,ϕ (θ, ϕ) = ψa,χ,θ ,0 (θ, ϕ − ϕ ).

(9.38)

Therefore, (9.36) becomes a convolution in ϕ which, by means of the convolution theorem, takes the form π 2π S(χ , θ , ϕ , a) = ψa,χ,θ ,0 (θ, ϕ − ϕ ) s(θ, ϕ) sin θ dθ dϕ (9.39) 0

= 2π

0

∞ k=−∞

†

e

i kϕ

π

ψˇ a,χ ,θ ,0 (θ )[k] sˇ (θ)[k] sin θ dθ,

(9.40)

0

This statement is valid for S 2 , but not in the case of the circle S 1 and the higher dimensional spheres S 3 and S 7 .

319

9.2 Wavelets on the 2-sphere and other manifolds

ˇ )[k] denotes its Fourier coefficient with where, for any function h on the sphere, h(θ respect to the longitudinal coordinate ϕ: 2π ˇ )[k] = h(θ dϕ h(θ, ϕ) e−i kϕ . (9.41) 0

In the discretization method of Section 9.2.5, the relations (9.39)–(9.40) will give us a tool for reducing the computational time of the spherical CWT. Indeed, they will allow us to use the fast Fourier transform (FFT).

9.2.4.2

Estimating the angular selectivity of a wavelet Given a wavelet ψ, it is very important in practice to know how well it will discriminate between two close directions. In other words, we would like to quantify the angular resolving power (ARP) of a spherical wavelet (see Section 3.4.1). A tempting definition is simply to look at the correlation between ψ and its rotated version: ψ|Rχ ψ K ψ (χ ) = = ψ−2 ψ|Ylm Ylm |Rχ ψ ψ|ψ l 0 |m|l −2 = ψ ψ|Ylm Rχ Ylm |ψ l 0 |m|l

= ψ =

−2

l 0 |m|l

ψ−2

=

m)|2 eimχ |ψ(l,

l |m|

m∈Z

ψ|Ylm eimχ Ylm |ψ

am e

imχ

.

m∈Z

As a function of χ, this expression reduces to a Fourier series where the coefficients are the same as in the spherical harmonic expansion of ψ. If the wavelet is highly sensitive to changes in the orientation, K ψ should be peaked around χ = 0. As we shall see later, the problem with this definition is that it does not depend on the scaling parameter a. This is not a problem for wavelets in R2 , but is rather counterintuitive on the sphere, because we know that a direction cannot be defined at large scales. This motivates the study of a more general indicator. Let us then introduce the following operator-valued function: Rψ,a (χ ) = dµ(ζ ) |ψa,χ,ζ ψa,χ,ζ |, (9.42) S2

where we choose again for [·] the Euler angles section. A good candidate for the ARP would then be the mean value: Rψ,a (χ )ψa =

ψa |Rψ,a (χ )ψa . ψa |ψa

(9.43)

320

Higher-dimensional wavelets

This time the inspection of Rψ,a (χ )ψa for different values of a should reveal a lack of angular precision at large scales. This is precisely what is observed in the case of the spherical Morlet wavelet, as shown later in Figure 9.8.

9.2.4.3

Designing directional spherical wavelets We have not yet addressed the problem of constructing good directional wavelets on S 2 . We will show now that this job is very naturally handled in our framework. First of all, we recall that the very definition of a direction on S 2 forces us to work at small scales. As is well known, the geometry of S 2 at small scales, or for large radii of the sphere, is closer and closer to that of R2 . As discussed in Section 9.2.3, the spherical wavelet transform respects one’s intuition by nicely approximating the Euclidean wavelet transform at small scales (the Euclidean limit property). We may remark that the notation used in (9.37) is consistent with it: roughly speaking, as the radius of the sphere goes to infinity, ψa,χ ,ζ (ζ ) goes to ψa,χ,b ( x ), where b ∈ R2 is the translation parameter. Moreover, it is a simple application of the Euclidean limit that small-scale Euclidean wavelets can be mapped to the sphere and yield small-scale admissible spherical wavelets. These can then be dilated to larger scales using the spherical dilation. This is neatly summarized by the following result, where we repeat for convenience the definition (5.1) of the inverse stereographic projection, namely, in polar coordinates, (I −1 f )(θ, ϕ) =

θ 2 f (2 tan , ϕ). 1 + cos θ 2

(9.44)

Proposition 9.2.5 The inverse stereographic projection (9.44) is a unitary map I −1 : L 2 (R2 , d x) → L 2 (S 2 , sin θ dθ dϕ) between the respective Hilbert spaces. Moreover, if ψ ∈ L 2 (R2 ) is an admissible 2-D Euclidean wavelet, then the function I −1 ψ is an admissible spherical wavelet. This results tells us that we can construct a spherical wavelet starting from any Euclidean wavelet. Now what does this tell us about directional wavelets? Since directional sensitivity is a local or small-scale attribute, it should intuitively survive this process. Yet there is more than intuition in this result. The stereographic projection and both spherical and Euclidean dilations are conformal mappings. Thus Proposition 9.2.5 defines a conformal application that, by definition, preserves angles. The directional sensitivity of the Euclidean wavelet is thus transported to the spherical wavelet. A natural candidate for building a directional spherical wavelet is to start with the (truncated) Euclidean Morlet or Gabor wavelet (3.19):

ψG ( x ) = ei k0 ·x e−|x | . 2

(9.45)

Using Proposition 9.2.5, we find the following spherical wavelet: θ

2 θ 2

eik0 tan 2 cos (φ0 −φ) e− 2 tan ψG (θ, φ) = 1 + cos θ 1

.

(9.46)

321

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.6. Real part of the spherical Morlet wavelet (9.46) at scale (a) a = 0.03 and (b) a = 0.3

(from [35]).

Fig. 9.7. Real part of the spherical Morlet wavelet (9.46) at scale a = 0.03 and centered at

(π/3, π/3): (a) θ = 0 and (b) θ = π/2 (from [35]).

This function is represented in Figures 9.6 and 9.7 for various values of the scale and rotation parameters. An example illustrating the directional sensitivity of this wavelet will be presented in Section 9.2.5.3 below (Figure 9.14), where we will compare it with that of the spherical DOG wavelet function. Of course, a spherical wavelet analysis should always be performed at small scales (see Section 9.2.5.2). As a confirmation, we show in Figure 9.8 the ARP (9.43) of

322

Higher-dimensional wavelets

1 0.9 0.8

a

〈 Rψ, (χ) 〉ψ

a

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.2

0.4

0.6

0.8

χ

1

1.2

1.4

1.6

1.8

Fig. 9.8. Angular resolving power (9.43) of the spherical Morlet wavelet (9.46) for various values of

the scale parameter: a = 0.03 (plain), a = 0.1 (dotted), a = 0.3 (dashed-dotted) and a = 1 (dashed).

the spherical Morlet wavelet for different values of the scale parameter. One clearly notices that, as the scale gets bigger, angles that are far apart become more and more correlated.

9.2.4.4

Representation of object surfaces The spherical wavelets just described are the basis of a novel representation of object surfaces [89] – a longstanding problem in computer graphics. The idea is the following. The surface of a 3-D (star-shaped) object is treated as a function on the 2-sphere (a human head is a good example, with a lot of neurophysiological interest). In a first step, the coarse structure of the surface is described by a truncated expansion of that function into spherical harmonics. Thus we get a band-limited function: f c (θ, ϕ) =

L

f (l, m) Ylm (θ, ϕ),

(9.47)

l=0 |m|l

for some small value of L (the authors choose L = 5). The justification is that the small l terms in the Fourier expansion are low-frequency components, which are well represented by functions supported on the whole sphere, such as spherical harmonics. Then the remainder f res = f − f c , which represents the fine structure of the surface, is described by well-localized spherical wavelets, in particular, spherical Morlet wavelets, with the help of an appropriate optimization procedure. The crucial step here is to

323

9.2 Wavelets on the 2-sphere and other manifolds

treat the low- and high-frequency components separately. We will see in Section 9.3 a mathematically precise version of this procedure. Namely, in the L 1 formalism, the reconstruction formula for the spherical CWT is based on precisely the same idea, with the scale variable a as the relevant parameter to distinguish between low and high frequencies (Theorems 9.3.8 and 9.3.9, which yield indeed approximation schemes for functions on the sphere). The same technique may be used also [90] to define smoothing on S 2 by a spherical Gaussian kernel, obtained again by inverse stereographical projection from a plane Gaussian (essentially our function φG of Section 9.2.2.3). Actually this Gaussian is nothing but a heat kernel, which generates a diffusion process on S 2 . This approach opens interesting new perspectives in object deformation (“morphing”), along the lines of the work of Sweldens et al. on digital geometry processing [242].

9.2.5

Implementation of the spherical wavelet transform The spherical CWT (9.27) is given as a convolution over the sphere S 2 . This creates numerical difficulties, for no really fast algorithm exists today. In particular, it is difficult to find an appropriate discretization of the sphere. Several methods have been proposed in the literature, mostly based on Fourier and spherical harmonics techniques [Moh97,154,169,216,274], but none of them is fully satisfactory. A possible exception is the method introduced by Wandelt and G´orski [367] and based on the use of the FFT. Interestingly enough, this approach was motivated by the analysis of cosmic microwave background (CMB) data, that we have mentioned in Section 5.1.2 – and for which spherical wavelets have been proposed! The new algorithm presented here, however, seems to answer the question rather well [35]. For a practical implementation of the spherical CWT, the first step is that of discretization. This means finding a suitable grid in the parameter space, so as to allow a fast calculation and a good approximation of the continuous theory. As we shall see, the key to the algorithm presented below is to use an FFT in the (periodic) longitude angle ϕ. Actually we also need some sort of criterion on the grid density for controlling aliasing problems, as indicated already in [13]. More precisely, we have to specify the scale interval in which the spherical wavelet transform makes sense. A possible answer will be suggested in Section 9.2.5.2. Then several examples will be discussed, both academic and real-life.

9.2.5.1

Discretization and algorithm Following an approach similar to that in [Win95], the first step is to discretize the integral (9.39) on a regular spherical grid M × N , π 2π t, ϕ p = p) | 0 t M − 1, 0 p N − 1}, M N by a weighted sum (χ and a are fixed throughout) G = {(θt =

(9.48)

324

Higher-dimensional wavelets

S(χ , θt , ϕ p , a) ≡ S[χ, t , p , a] = ψa,χ,t [t, p − p ] s[t, p] wt ,

(9.49) (9.50)

0 t M−1 0 p N −1

where: r s[t, p] ≡ s(θ , ϕ ); t p

r ψ

[t, p − p ] ≡ ψa,χ,θt ,0 (θt , ϕ p− p ); a,χ ,t r the index of ϕ is extended to Z by angular periodicity with the rule ϕ p+N = ϕ p ; r w = (2π 2 /M N ) sin θ are the weights suggested in [Win95] for the discretization t t of the Lebesgue measure on the particular grid G. Evaluating the sums in (9.50) requires M N additions and multiplications for each (t , p ), that is, M 2 N 2 operations altogether. However, an easy simplification can be obtained for the longitudinal coordinates by the use of a Fourier series and the Plancherel formula. Indeed, denoting by 2π ˇ k] = h[t, h[t, p] exp(−i kp ), (9.51) N 0 p N −1 the longitudinal Fourier coefficients of a given discrete function h, we obtain S[χ , t , p , a] = 2π wt F[χ, t , p , a, t]

(9.52)

0t M−1

with F[χ , t , p , a, t] =

2π ψˇ a,χ,t [t, k] sˇ [t, k] exp(i kp ). N 0k N −1

(9.53)

The quantity F may be computed with the inverse fast Fourier transform (IFFT), which leads to a reduction of the computational time from O(M 2 N 2 ) to O(M 2 N log N ). On a grid G of 256 × 256, the gain is a factor of 46. Notice that other discretization techniques than a plain Riemann sum, as used in (9.50), would be beneficial only if one imposes additional regularity conditions on the signal s. Furthermore, other weights wt could be chosen to achieve a better approximation of (9.49). An example of a different choice, both for the weights and for the discretization technique, is that of a band-limited spherical function, as considered in [216] and [367], that is, a function f (θ, ϕ) whose expansion in spherical harmonics has only finitely many nonzero terms, f (l, m) = 0, ∀ l > L. One can also ask whether a better algorithm could be designed on another type of grid. For instance, what is the use in the continuous case of an icosahedral grid, as introduced in [336,337] (see Figure 9.9) for the starting point of a lifting procedure? The great advantage of such a grid (or one built on a fullerene type of grid) is its better isotropy. The point is that the usual spherical coordinates necessarily introduce a preferred direction, the polar axis, and this is at the origin of the various tentatives reported in the literature

325

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.9. The geodesic sphere construction, starting with the icosahedron on the left (subdivision

level 0) and the next two subdivision levels (from [336]).

of designing an equidistribution of points on the sphere [113,169,330]. The last paper, in particular, contains many references to related works, from pure mathematics to physics (electrostatics).

9.2.5.2

Numerical criterion for the scale range The discretization of the continuous spherical wavelet transform gives rise to a sampling problem. Since the grid G is fixed, if we contract or dilate our wavelet too much, we obtain a function which is very different from the original ψ. In other words, aliasing occurs and the wavelet is not numerically admissible. This problem has been noted also in the flat space CWT (see Section 4.1.1), but only an empirical criterion was given [13]. We have seen in Proposition 9.2.3 that a function ψ ∈ L 2 (S 2 , dµ) is admissible only if it satisfies the zero mean condition (9.31). Approximating the integral by its Riemann sum, we get the quantity C[ψ] =

1t M−1 1 p N −1

ψ(θt , ϕ p ) µ(θt , ϕ p ). 1 + cos θt

(9.54)

Because of the discretization, even if ψ verifies (9.31), it is not necessarily true that C[ψ] vanishes. However, we may suppose that this quantity is very close to zero when ψ is sampled sufficiently, that is, if the grid G is fine enough. Nevertheless, it is difficult to give a quantitative meaning to the value of C[ψ]. How small is “very close to zero”? Here is a possible solution to this problem. Since the spherical measure µ and the function 1 + cos θ are positive, it is clear that C[ψ] C[|ψ|]

(9.55)

for any ψ ∈ L 2 (S 2 , dµ). So we can define a numerical normalized admissibility by C[ψ] , C[ψ] = C[|ψ|] a quantity always contained in the interval [-1,1]. We can now give a precise definition of numerical admissibility.

(9.56)

Higher-dimensional wavelets

Definition 9.2.6 A spherical wavelet of L 2 (S 2 , dµ) is numerically admissible with threshold p% or p% - admissible, if the numerical normalized admissibility (9.56) is smaller than (100 − p)/100 in absolute value: |C[ψ]|

100 − p . 100

(9.57)

As an example, we present in Figure 9.10 the behavior of the dilated spherical DOG wavelet, Da ψG(α) (α = 1.25), as a function of a > 0, discretized on a 128 × 128 grid (notice that, in the flat case, α = 1.6 is the value for which the DOG wavelet is almost indistinguishable from the Mexican hat). According to this plot, the wavelet Da ψG(α) is 99 % - admissible on the scale interval a ∈ [0.072, 24.71]. The lower limit is due to the fact that, for small a, Da ψ is not sampled enough. The upper limit comes from the subsampling of the area far from the North Pole which, according to the spherical dilation, gets more and more contracted. Figure 9.11 presents three typical behaviors of Da ψ discretized on a 22 point θ sampling. For a = 0.5, the sampling is correct. For a = 0.05, that is, below the lower admissibility bound, subsampling occurs, so that negative parts of Da ψ are completely missed. Clearly, this discretized wavelet is no longer admissible. Exactly the same effect was observed long ago in the flat case [13]. The third case, with a = 3.5, thus a=0.072

a=24.71

1

0.8

0.6

0.4

Num. Norm. Admiss.

326

0.2

0

−0.2

−0.4

−0.6

−0.8

−1 −5

−4

−3

−2

−1

0

1

2

log a

a ψG(α) ] as a function of log a for α = 1.25 (from [35]). Fig. 9.10. C[D

3

4

5

327

9.2 Wavelets on the 2-sphere and other manifolds

0.8

(a)

(α) 0.7

DaψG

0.6 0.5 0.4 0.3 0.2 0.1 0 −0.1 −0.2 −4

−3

−2

−1

0

1

8

DaψG(α)

2

3

4

θ 0.2

(b)

7

(c)

(α) 0.15

D a ψG

6

0.1

5

0.05

4

0 −0.05

3 2

−0.1

1

−0.15 −0.2

0 −1 −4

−3

−2

−1

0

1

2

3

4

θ

−0.25 −4

−3

−2

−1

0

1

2

3

4

θ

(α)

Fig. 9.11. Three typical behaviors of Da ψG discretized on a 22 × 22 grid G. (a) For a = 0.5, the

sampling is correct; (b) for a = 0.05, subsampling occurs, negative parts of Da ψG(α) are completely missed; (c) subsampling on the negative parts of Da ψG(α) for a = 3.5. Notice the minimum at θ = 0 (from [35]).

beyond the upper bound, is less intuitive. Here the subsampling takes place for large values of θ, that is, close to the South Pole, but the result is the same; the discretized wavelet does not have a zero mean, it is not admissible. In addition, the curve presents a minimum at θ = 0. This somewhat unexpected effect is in fact due to the cocycle, as is the dependence of the height on a. Indeed, if one performs the same calculation without the cocycle, all curves show a maximum at θ = 0, with the same height. Here again we see that curvature, which requires the presence of the cocycle, has a nontrivial effect. As a further illustration of this behavior, it is instructive to consider the function ι identically equal to 1. In the flat case, this function has a vanishing WT, by the admissibility condition R2 d 2 x ψ( x ) = 0 on the wavelet, but it is not square integrable and thus cannot be reconstructed. In the present case, however, the situation is different. The function ι is square integrable, since the sphere S 2 is compact, but its WT does not vanish, because of the presence of the cocycle. Indeed, the function ι is invariant under rotation, but not under dilation:

328

Higher-dimensional wavelets

(Da ι)(θ, φ) = λ(a, θ )1/2 ≡ 1,

(9.58)

and, therefore, I (,, a) = R, Da ψ|ι = ψ|Da ι ≡ I (a) = dµ(ζ ) ψ(ζ ) λ(a, θ)1/2 = 0.

(9.59)

S2

Thus, for fixed a, the WT I (a) of the unit function is constant, and essentially negligible for a 1. Significant values appear only for a > 2, and these scales are irrelevant for the analysis of signals such as contours. This behavior has been checked numerically [35], with the familiar DOG wavelet ψG(α) , discretized on a 128 × 128 grid. Because of the discretization, however, the function I (,, a) ≡ I (θ, a) does depend on a, but very little, and its value remains extremely small. As a matter of fact, for a < 0.1, the WT of ι is numerically negligible over the whole sphere, and may be taken as zero to a very good approximation. As a consequence, the spherical CWT does have the familiar local filtering effect, provided small scales are considered. This will be confirmed by the examples below. Once again, we see that the CWT is useful only as a local analysis.

9.2.5.3

Examples of spherical wavelet transforms As a first example, we analyze in Figure 9.12 an academic picture, namely, (the characteristic function of) a spherical triangle on S 2 , with one of the corners sitting at the North Pole. The triangle is given by 0◦ θ 70◦ , 0◦ ϕ 90◦ and is discretized on a 128 × 128 grid in (θ, ϕ). We use again the spherical Gaussian wavelet ψG(α) , for α = 1.25, discretized on the same grid. According to the admissibility analysis presented above (Figure 9.10), the wavelet is 95 % - admissible on the scale interval a ∈ [0.033, 29.27]. Thus we can evaluate the spherical CWT of this picture for various scales in the allowed range, and we have chosen four successive scales from a = 0.5 to a = 0.035. Figure 9.12 shows that the spherical WT behaves here exactly as, in the flat case, the WT of the characteristic function of a square, as shown in Figure 4.1, in Section 4.1.1. For large a, the WT sees only the object as a whole, thus allowing us to determine its position on the sphere. When a decreases, increasingly finer details appear; in this simple case, eventually only the contour remains, and it is perfectly seen at a = 0.035. The transform vanishes in the interior of the triangle, as it should, only the “walls” remain, with a negative value (black) just outside, a zero-crossing right on the boundary and a sharp positive maximum (white) just inside. In addition, each corner gives a neat peak, which is positive, since the corner is convex [13]. Notice that the three corners are alike, so that indeed the poles play no special rˆole in our spherical WT, contrary to what occurs often in the classical spherical analysis based on spherical harmonics [Fre97,Fre99,169,294,319]. On the other hand, the scales are different in the four cases, since the amplitude of the transform diminishes for decreasing values of a.

1

0.5

Z

0

−0.5

−1 −1 −1

−0.5

−0.5

0

0 0.5

0.5 1

1

X

Y

(a) 0.3

0.2

0.25 1

1 0.2

0.5

0.15

0.5

0

Z

Z

0.1

0.15

0.1

−0.5

− 0.5

0.05

−1 −1

0

−1 −1

−1

−0.5

−0.5

0

0.05

−0.5

0

0 0.5 Y

X

0

−1

0 0.5

0.5 1 1

−0.5

0

Y

0.5 1 1

−0.05 X

(c)

(b) 0.07

0.02

0.06 1

1

0.05 0.04

0.5

0.015

0.5

0.01

0.02 0.01

−0.5

Z

Z

0.03 0

0

0.005

−0.5

0

0 −1 −1

−0.01 −0.5

−0.5

0

0 0.5 Y

0.5 1 1

(d)

X

−1

−0.02 −0.03

−1 −1

−0.005 −1

−0.5

−0.5

0

0

0.5 Y

0.01

0.5 1 1

X

(e)

Fig. 9.12. Spherical wavelet transform of the characteristic function of a spherical triangle with apex at the North Pole, 0◦ θ 70◦ , 0◦ ϕ 90◦ , obtained with the difference wavelet (9.32), at scale a = 0.25. (a) Original image. The transform is shown at four successive scales, (b) a = 0.5; (c) a = 0.2; (d) a = 0.1; and (e) a = 0.035. As expected, it vanishes inside the triangle, and presents a “wall” along the contour, with sharp peaks at each summit. Notice that the amplitudes are different in the four cases (from [35]).

330

Higher-dimensional wavelets

(a)

(b)

(c)

(d)

Fig. 9.13. Spherical wavelet transform of the spherical map of the European area, computed with the spherical DOG wavelet for α = 1.25. (a) The original picture; (b) wavelet transform at a = 0.032; (c) the same at a = 0.016; (d) the same at a = 0.0082 (from [35]).

As a second, real life example, we present in Figure 9.13 the wavelet transform of a significant piece of the terrestrial globe, covering Europe, Greenland and North Africa. As before, we use the spherical DOG wavelet ψG(α) for α = 1.25. The transforms are shown again at three successive scales, a = 0.032, 0.016, 0.0082 (the grid used here is finer than the one used in the previous examples, so that smaller values of a are admissible). As expected, the resolution improves with diminishing a. However, at a = 0.0082, the discretization grid used for the computation of the transform coincides with that of the original picture, so that one sees exactly the same artifacts, such as a closed strait of Gibraltar, an unresolved complex Corsica–Sardinia, ragged coastlines, etc. Of course, we cannot hope to improve on the resolution of the original!

331

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.14. Spherical CWT of the spherical triangle at scale a = 0.03, using spherical Morlet wavelet

(9.46): (a) θ = 0 and (b) θ = π/2 (from [35]).

As a final example, we take again the spherical triangle of Figure 9.12(a) and analyze it with the spherical Morlet wavelet (9.46). The results confirm our analysis: the spherical CWT is able to detect the local orientation of the edges of the triangle (Figure 9.14).

9.2.6

Extension to other manifolds The whole construction made so far extends almost verbatim to the (n − 1)-dimensional sphere S n−1 = SO(n)/SO(n − 1), with help of a similar class I representation of the generalized Lorentz group SOo (n, 1) [23]. Although the spheres are the manifolds on which a CWT is most desirable for applications, the mathematical analysis made here invites to consider other manifolds with similar geometrical properties. We take first n = 3. The sphere S 2 = SO(3)/SO(2) is a compact Riemannian symmetric space of constant curvature κ = 1. It has a noncompact dual, H 2 = SOo (2, 1)/SO(2), of constant curvature κ = −1 [Hel78]. H 2 is a two-sheeted hyperboloid, symmetric around the x3 -axis. Duality corresponds to the fact that SO(3) and SOo (2, 1) are the two real forms of the complex group SO(3)C ∼ S L(2, C). Exactly as in the case of the sphere, we can perform a stereographic projection from the South Pole onto the equatorial plane x3 = 0 (or, equivalently, to the plane tangent at the other pole). Then maps the upper sheet H+2 onto the interior D+ of the unit disk, and the lower sheet H−2 onto the exterior. The domain D+ , called the Lobachewskian disk, is conformally equivalent to H+2 , and both manifolds have SOo (2, 1) as isometry group.

332

Higher-dimensional wavelets

As we did for the sphere, dilations on H+2 may be obtained by lifting dilations in the equatorial plane by inverse stereographic projection. The resulting map has all the required properties for a dilation, but does not come directly from a linear group action. Thus it can only be used for constructing wavelets on H+2 if one puts it by hand. It remains to obtain a suitable representation of the resulting set SOo (2, 1) · R+ ∗ in L 2 (H+2 , dµ), where dµ is the SOo (2, 1)-invariant measure, and to show that it is square integrable in a suitable sense. Now this suggests a further generalization. In both cases, S 2 as well as H 2 , the unit disk, image of one sheet or one hemisphere, is a classical domain. Also the stereographic projection has a group-theoretical origin [Per86]. This paves the way to the generalization of the CWT to a whole class of homogeneous spaces (Riemannian symmetric spaces). For instance, S n−1 = SO(n)/SO(n − 1)

and

H n−1 = SOo (n − 1, 1)/SO(n − 1)

are dual Riemannian symmetric spaces, with constant curvature κ = ±1, respectively. Again SO(n) and SOo (n − 1, 1) are two real forms of the complexified group SO(n)C . Their isometry groups are SO(n) and SOo (n − 1, 1), respectively, so that suitable representations of the generalized Lorentz group SOo (n, 1) should provide the corresponding wavelets. While this has been obtained explicitly in the case of the (n − 1)sphere S n−1 in [23], the problem is still open in the noncompact case, i.e., the hyperboloid H n−1 .

9.3

Wavelet approximations on the sphere The central theme of approximation theory is the representation of a function by a truncated series expansion into a family of basis functions, for instance, the elements of a frame. Thus, in the flat case, one- or two-dimensional wavelets are widely used for approximation in various function spaces [Mal99]. The crucial advantage is their multiresolution character, which is optimally adapted to local perturbations. A natural framework is given by the Lebesgue spaces L p (Rn ), 1 p < ∞. One of the reasons is that approximation is often formulated in terms of convolution with an approximate identity, and many useful convolution identities are available in L p . Therefore, in order to apply these considerations to the sphere S 2 , it is necessary to have a good notion of convolution on S 2 . For that purpose, it is useful to represent the sphere as the quotient SO(3)/SO(2), since the convolution machinery extends almost verbatim to locally compact groups, and then partly to homogeneous spaces. For the convenience of the reader, we have collected in the Appendix the main definitions and essential properties of convolution on a locally compact group. In what follows, we will need two different cases. For simplicity, we write L 2 (SO(3)) ≡ L 2 (SO(3), d,), where d, is the Haar measure on SO(3), and L p (S 2 ) ≡ L p (S 2 , dµ).

333

9.3 Wavelet approximations on the sphere r

If f ∈ L 2 (SO(3)) and g ∈ L 1 (S 2 ), then their spherical convolution is the function on S 2 defined as f (,) g(,−1 ζ ) d,. (9.60) ( f % g)(ζ ) = SO(3)

Then f % g ∈ L 2 (S 2 ) and one has f % g2 f 2 g1 , r

(9.61)

where the norms refer to the corresponding spaces. If f ∈ L 2 (S 2 ) and g ∈ L 1 (S 2 ), their spherical convolution is the function on SO(3) defined as (f % g)(,) = f (,−1 ζ ) g(ζ ) dµ(ζ ). (9.62) S2

Then f % g ∈ L 2 (SO(3)) and f % g2 f 2 g1 ,

(9.63)

Here, however, we are only interested in functions on the sphere S 2 , that is, functions on SO(3) that are SO(2)-invariant. In particular, we will deal mostly with axisymmetric, or zonal, functions on S 2 , that is, functions of θ alone. Thus, we will focus on elements of L 2 ([−1, +1], dt), where t = cos θ, for which the Fourier series reduces to a Legendre expansion: ψ(t) =

∞ 2l + 1 l=0

= 2π ψ(l)

4π

+1

−1

Pl (t), ψ(l)

dt Pl (t) ψ(t) =

*

4π ψ(l, 0), 2l + 1

m) ≡ Y m |ψ denotes the Fourier coefficient of ψ. If f is an axisymmetric where ψ(l, l function, the spherical convolution (9.62) takes a simpler form [Fre97]: Proposition 9.3.1 Let f and g be two measurable functions on S 2 . If f is axisymmetric, the spherical convolution of f and g is a function on S 2 , which can be written:

dµ(ζ ) f ( ζ · ζ ) g(ζ ), (9.64) ( f % g)(ζ ) = S2

ζ is the R3 scalar product of unit vectors of directions ζ and ζ . where ζ · The proof amounts to a straightforward application of harmonic analysis (Fourier series) on S 2 and of the addition theorem for spherical harmonics, to the effect that f (,−1 ζ ) = f ( ζ · ζ ), where ζ ≡ ,˙ ∈ S 2 denotes the left coset of , ∈ SO(3) (see the geometrical discussion in Section 9.2.4.1).

334

Higher-dimensional wavelets

Now we may turn to the approximation problem proper. As in the Euclidean case [Lie97,Ste71], a convenient technique is to perform a convolution with a smoothing kernel, that acts as an approximate identity, that is, a kernel which tends to the identity (δ function) as the parameter goes to 0. For the sake of simplicity, we will only deal with zonal kernels, following mainly [Fre97]. Definition 9.3.2 Let Kτ , τ ∈ (0, τo ], τo ∈ R+ ∗ , be a family of elements of 1 ( L ([−1, +1], dt) satisfying Kτ (0) = 1. The functional Sτ [ f ] defined by Sτ [ f ] = Kτ % f,

f ∈ L p (S 2 ),

1 p < ∞,

is called a singular integral. It is called an approximate identity of L p (S 2 ) if lim f − Sτ [ f ] p = 0,

τ →0 τ >0

∀ f ∈ L p (S 2 ) .

(9.65)

The following theorem characterizes those spherical kernels which are associated with an approximate identity. Theorem 9.3.3 Let {Kτ } be a uniformly bounded spherical kernel, that is, there exists a constant M, independent of τ , such that +1 dt |Kτ (t)| M, ∀τ ∈ (0, τo ] . −1

Then the associated singular integral is an approximate identity of L p (S 2 ) if and only if (τ (n) = 1, lim K

τ →0 τ >0

∀n 0.

(9.66)

A proof may be found in [Fre97]. A particularly interesting case is given by positive definite kernels. In this case, since |Pl (t)| 1, {Kτ } is uniformly bounded, with (, (0). The following theorem gives a nice characterization of bound M = supτ ∈(0,τ0 ] K approximate identities associated with positive kernels. Theorem 9.3.4 Let {Kτ }, τ ∈ (0, τo ], be a positive kernel associated with a singular integral of L p (S 2 ). Then each of the following conditions is equivalent to (9.65) and (9.66), which means that {Kτ } is the kernel of an approximate identity: (τ (0) = 1, (i) lim K τ →0 τ >0

δ

(ii) lim

τ →0 τ >0

−1

dt Kτ (t) = 0, δ ∈ (−1, +1).

335

9.3 Wavelet approximations on the sphere

Condition (ii) is in fact a constraint on the localization of the kernel, as we shall see in an explicit example below (see Figure 9.15). Approximate identities are a very useful tool for harmonic analysis on the sphere and many applications can be found in [Fre97]. Thus it is gratifying that the spherical wavelet transform naturally yields a systematic way of deriving approximate identities on the sphere. This is actually an interesting way of handling functions on the sphere, because it allows us to represent information by means of localized, and hierarchically organized, coefficients. With such a representation, a local modification of the function would only result in a slight local perturbation of the original coefficients, a definite advantage over Fourier series (exactly as in the case of flat space). As a matter of fact, many examples of approximate identities are given in the textbook of Freeden et al. [Fre97], and they are applied extensively by these authors to geophysical data [Fre99]. Most of these examples are based on families of kernels indexed by a parameter which behaves like a dilation. However, since the latter is introduced directly as a parameter in those kernels, there is no unique way of generating approximate identities, as in Rn [Ste71]. But this problem disappears naturally if one uses the spherical dilation, since, as we shall see, the dilation operator generates an approximate identity in L 2 (S 2 ). However, we have to modify it first and adapt it to the L 1 environment. Using the notation of Section 9.2.2, we define, instead of Da , a new dilation operator: (D a f )(ζ ) ≡ f a (ζ ) = λ(a, θ) f (ζ1/a ),

(9.67)

and this operator clearly conserves the L 1 norm. Notice that the situation is more complicated here than in the flat case. There, indeed, changing the dilation operator from L 2 to L 1 simply amounts to changing the power of a in front of the transform [29]. Here, one replaces the factor λ(a, θ)1/2 by its square λ(a, θ), but this modifies the CWT itself in a nontrivial way. In particular, the admissibility condition (9.25) becomes ∞ da (a 8π 2 |ψ (l, m)|2 < c . (9.68) 2l + 1 |m| l 0 a First, we notice that our new dilation operator does not change the mean of a function, thus simplifying the statement of Proposition 9.2.4. Proposition 9.3.5 If ψ ∈ L 1 (S 2 ), then dµ(ζ ) ψ a (ζ ) = dµ(ζ ) ψ(ζ ). S2

(9.69)

S2

The proof reduces to a simple change of variables, taking into account the cocycle relation (9.23).

336

Higher-dimensional wavelets

Acting with the new dilation D a on a suitable function, one can now easily construct an approximate identity, as shown in the next proposition. Proposition 9.3.6 Let f ∈ C([−1, +1]), with f (0) = 1. Then the family { f a ≡ D a f, a > 0} is the kernel of an approximate identity. In view of Theorem 9.3.3, the proof consists in two steps. First, one shows that the fam +1 ily { f a } , a ∈ (0, 1], is uniformly bounded, which is obvious since −1 dt | f a (t)| = f 1 . Next, it remains to verify that f a (l) = 1, lim (

a→0 a>0

and this is done again by a change of variables and applying the cocycle relation (9.23) [35]. This technique is applied in Figure 9.15 to a zonal function of Gaussian shape, namely the mother wavelet of the spherical DOG wavelet, φG (θ, φ) = exp(− tan2 (θ/2)), θ ∈ [−π, π]. One clearly sees how dilation localizes the kernel better and better as a → 0. In the L 1 formalism, we recall from [29] that the necessary condition for admissibility becomes a genuine zero mean condition, exactly as in the flat case: 12

φ (θ, ϕ)

a = 0.7

G

a = 0.5

10

a = 0.3 8

6

4

2

0

π

−3

−2

−1

0

θ

1

2

3

−π

Fig. 9.15. Kernel of an approximate identity obtained by dilating a Gaussian mother function with scaling factor a = 0.7, 0.5 and 0.3 (from [35]).

337

9.3 Wavelet approximations on the sphere

S2

dµ(θ, ϕ) ψ(θ, ϕ) = 0,

(9.70)

Correspondingly, in view of Proposition 9.3.5, the difference wavelet ψφ(α) given in (9.32) is replaced by ψ˘ φ(α) (θ, ϕ) = φ(θ, ϕ) − D α φ(θ, ϕ)

(α > 1).

Now, combining the modified dilation operator D a with the usual rotation operator R, , we define a new set of spherical wavelets, starting from an admissible ψ, namely, ψ,a ≡ R, D a ψ = R, ψ a . Accordingly, we redefine as follows the spherical wavelet transform of a signal s ∈ L 2 (S 2 ): S˘ψ (,, a) =

S2

dµ(ζ ) ψ,a (ζ ) s(ζ ).

(9.71)

In particular, if the wavelet ψ is zonal, we get S˘ψ (ζ, a) =

S2

dµ(ζ ) ψ a ( ζ · ζ ) s(ζ ).

(9.72)

In addition, the correspondence between spherical wavelets and their stereographic projections on the tangent plane (Proposition 9.2.5) is modified, as follows. Proposition 9.3.7 If ψ ∈ L 2 (R2 ) is an admissible 2-D Euclidean wavelet, and I −1 denotes the inverse stereographic projection (9.44), then the function (1 + cos θ )−1 I −1 ψ is an admissible spherical wavelet for the transform defined with the L 1 normpreserving dilation operator D a . After this preparation, we proceed to show that the spherical CWT admits a reconstruction formula, valid in the strong L 2 topology, exactly as the usual CWT in Rn . As in the flat case, described in Section 2.6.1, we may distinguish between a bilinear and a linear formalism [Tor95]. But there is a crucial difference. In the flat case, it is advantageous, but not compulsory, to treat the large scales or low frequencies separately, in terms of a scaling function. Here, however, we are forced to do it. The reason is that, geometrically, only small scales are relevant and lead to the expected filtering behavior, as discussed in Section 9.2.5.2. We arbitrarily choose a = ao as reference scale and define the scales a > ao as large. Notice that we recover here, in precise mathematical terms, the argument behind the mixed spherical harmonics/spherical wavelets representation of object surfaces from [89], discussed in Section 9.2.4.4. Let us begin with the bilinear analysis. Given a wavelet ψ ∈ L 1 (S 2 ), we define the corresponding scaling function ≡ (ao ) by its Fourier coefficients:

338

Higher-dimensional wavelets

m)|2 = | (l,

∞

ao

da (a |ψ (l, m)|2 , a

l 1,

(9.73)

1 (9.74) 8π 2 (the integral in (9.73) converges in virtue of the admissibility condition (9.68) satisfied by ψ). Of course, (9.73) does not define the function uniquely. We can, for instance, m) 0, ∀ l, m, as in [Fre97]. Corresponding to (9.71), assume in addition that (l, we define the large-scale part of a signal s as o) ˘ (,, ao ) = / dµ(ζ ) (a (9.75) , (ζ ) s(ζ ), 0)|2 = | (0,

S2

(ao ) −1 o) where we have put (a (, ζ ). , (ζ ) ≡

Theorem 9.3.8 (Bilinear analysis) Let ψ ∈ L 1 (S 2 ) be a spherical wavelet and let ≡ (ao ) , ao > 0, denote the associated scaling function. Assume the following two conditions are satisfied: r for all l = 1, 2, . . . , 8π 2 ∞ da (a |ψ (l, m)|2 = 1, (9.76) 2l + 1 |m| l 0 a r

for all $ ∈ (0, ao ), there is a constant M > 0, independent of $, such that ao da ψ a 2 M. a $

Then, for all s ∈ L 2 (S 2 ), we have the equality ao da o) ˘ (,, ao ) (a s= d, S˘ψ (,, a) ψ,a + d, / , , a SO(3) 0 SO(3)

(9.77)

(9.78)

˘ is the large-scale where S˘ψ is the spherical CWT of s with respect to the wavelet ψ, / 2 2 part of s and the integral is understood in the strong sense in L (S ). Proof . We consider the first term in (9.78). Since ψ ∈ L 1 (S 2 ) and s ∈ L 2 (S 2 ), Young’s convolution inequality (9.62) shows that S˘ψ ∈ L 2 (SO(3)). As in the flat case, we define the infinitesimal detail at scale a: a d (ζ ) = d, S˘ψ (,, a) ψ,a (ζ ). SO(3)

This is a convolution on SO(3) and Young’s inequality (9.61) shows that d a ∈ L 2 (S 2 ). Explicitly, we have d a (ζ ) = dµ(ζ ) s(ζ ) d, ψ a (,−1 ζ ) ψ a (,−1 ζ ). (9.79) S2

SO(3)

339

9.3 Wavelet approximations on the sphere

Using the relation ψ a (,−1 ζ ) =

∞

l (a (l, n) Ylm (ζ ), Dmn (,) ψ

(9.80)

l=0 |m| l |n| l l where Dmn (,) denotes a Wigner function [Ros57,Tal68], we find

(a (l, n) ψ (a (l , n ) d a (ζ ) = dµ(ζ ) s(ζ ) Ylm (ζ ) Ylm (ζ ) ψ S2

lmn l m n

× SO(3)

l (,) D l (,). d, Dmn m n

Using the orthogonality of Wigner functions and the addition theorem for spherical harmonics (see Section A.4), this gives: ∞ a (a (l, m)|2 . d (ζ ) = 2π ζ · ζ |ψ dµ(ζ ) s(ζ ) Pl S2

l=0 |m| l

Now consider the following expression: ao da a d (ζ ) s$, (ζ ) = a $ ao ∞ da

(a (l, m)|2 . = 2π ζ · ζ |ψ dµ(ζ ) s(ζ ) Pl a l=0 |m| l S2 $ In virtue of condition (9.77), the double summation on the right-hand side of this equation is absolutely and uniformly convergent, since it is majorized by ao ao ∞ da da (a (l, m)|2 = |ψ ψ a 2 . a a $ $ l=0 |m| l Now let us introduce the quantity: ∞ ao da (a (ao ) 2 K$ (t) = 2π |ψ (l, m)| Pl (t), a $ l=0 |m| l so that s$(ao ) = K$(ao ) % s. By (9.77), we see that K$(ao ) ∈ L 1 ([−1, +1]), for all 0 < $ ao , and K$(ao ) 1 2π M. Next, we show in the same way that the second term in (9.78) equals H(ao ) % s, where H(ao ) (t) = 2π

∞

m)|2 Pl (t). | (l,

l=0 |m| l

Again, H(ao ) ∈ L 1 ([−1, +1]). Finally, we define the kernel K$ = K$(ao ) + H(ao ) , which also belongs to L 1 ([−1, +1]). Condition (9.77) shows that K$ is a uniformly bounded

340

Higher-dimensional wavelets

m), we deduce kernel. In addition, from (9.76) and the definition (9.73)–(9.74) of (l, the following constraint on its Legendre coefficients : ao 8π 2 da (a 2 2 ( lim K$ (l) = |ψ (l, m)| + | (l, m)| $→0 2l + 1 |m| l 0 a 8π 2 ∞ da (a |ψ (l, m)|2 = 1, l 1, 2l + 1 a |m| l 0 = 2 8π | (0, 0)|2 = 1, l = 0. Then Theorem 9.3.3 shows that K$ is the kernel of an approximate identity, which proves the strong convergence in L 2 (S 2 ) of the approximation: lim (K$ % s) = s.

$→0

As a check of the reconstruction formula (9.78), let us consider the unit function ι. Contrary to the case of the L 2 formalism, the L 1 -normalized CWT of ι vanishes identically, as a consequence of Proposition 9.3.5: a ˘Iψ (,, a) = dµ(ζ ) ψ (ζ ) = dµ(ζ ) ψ(ζ ) = 0. S2

S2

Hence only the second term, the large-scale part, subsists in (9.78). Using again the expansion (9.80), we find successively: 0), dµ(ζ ) (,−1 ζ ) = (0, I˘ (,, ao ) = S2

and, for (9.78), ι(ζ ) = (0, 0)

0)|2 = 1. d, (,−1 ζ ) = 8π 2 | (0, SO(3)

This result shows that the large-scale part of a signal must be treated separately, because constant functions on the sphere are square integrable, and hence must be reconstructible, although their CWT vanishes identically. In practice, of course, large scales should be irrelevant, since wavelet analysis is local, and we expect the second term in (9.78) to be numerically negligible (that is, one must choose ao large enough for this to be true). Theorem 9.3.8 applies, in particular, to a zonal wavelet. The only change is the parameter space of the spherical CWT which takes the form of the product S 2 × R+ ∗, −1 with the measure a da dµ(ζ ). A further simplification yet is to consider a singular reconstruction wavelet and build a framework similar to the Morlet linear analysis. As in the bilinear case, we begin by defining, through its Legendre coefficients, a scaling function φ ≡ φ (ao ) that takes care of the large scales :

341

9.3 Wavelet approximations on the sphere

= φ(l)

∞

ao

da (a ψ (l), a

l 1,

= 1. φ(0)

(9.81) (9.82)

The corresponding large part of a signal s is then σ˘ φ (ζ, ao ) = dµ(ζ ) φ( ζ · ζ ) s(ζ ).

(9.83)

S2

In these notations, the linear reconstruction formula is given by the following theorem. Theorem 9.3.9 (Linear analysis) Let ψ ∈ L 1 (S 2 ) be a zonal spherical wavelet satisfying the following two conditions: r for all l = 1, 2, . . . , ∞ da (a (9.84) ψ (l) = 1, a 0 r

for all $ ∈ (0, ao ), ∞ 2l + 1 ao da (a ψ (l) < ∞ . 4π a $ l=0

(9.85)

Then, for all s ∈ L 2 (S 2 ), we have the equality ao da ˘ Sψ (ζ, a) + σ˘ φ (ζ, ao ), s(ζ ) = a 0 the integral being again understood in the strong sense in L 2 . Proof . The same arguments as in the proof of Theorem 9.3.8 show that the partial sum ao da ˘ (ao ) s$ (ζ ) = Sψ (ζ, a) a $ belongs to L 2 (S 2 ). Expanding this expression and adding the large-scale term, we find ao da a s$ (ζ ) = dµ(ζ ) ψ (ζ · ζ ) s(ζ ) + σ˘ φ (ζ, ao ) a S2 $ ao da a

= dµ(ζ ) s(ζ ) ψ (ζ · ζ ) + φ(ζ · ζ ) a S2 $ ao ∞ 2l + 1 da (a

ψ (l) + φ(l) Pl ( = dµ(ζ ) s(ζ ) ζ · ζ ) 4π a S2 $ l=0 = (κ$ % s)(ζ ),

342

Higher-dimensional wavelets

where we have used (9.85) and set ao ∞ 2l + 1 da (a ψ (l) + φ(l) Pl (t). κ$ (t) = 4π a $ l=0 The Legendre coefficients of this kernel are ao da (a (l) = ψ (l) + φ(l). κ( $ a $ As in the proof of Theorem 9.3.8, we deduce from condition (9.84) that lim$→0 κ( $ (l) = 1, ∀ l = 0, 1, . . . . Thus we have again an approximate identity, which allows us to conclude that lim s − κ$ % s2 = 0 .

$→0

The conclusion of this analysis is that our spherical CWT, with the modified dilation operator D a , leads to the same approximation scheme as that developed by Freeden [Fre97,Fre99]. The present approach, however, has the additional advantage of giving a clear geometric meaning to the approximation parameter a. By the same token, it intuitively explains the validity of the Euclidean limit established in [29]. Indeed, taking a → 0 means going to the pointwise limit where curvature becomes unimportant, that is, going to the tangent plane and recovering the flat CWT.

10

Spatio-temporal wavelets and motion estimation

10.1

Introduction We live in a world where objects (cars, animals, men, birds, aeroplanes, the Sun, etc.) that surround us are constantly in relative motion. One would like to extract the motion information from the observation of the scene and use it for various purposes, such as detection, tracking and identification. In particular, tracking of multiple objects is of great importance in many real world scenarios. The examples include traffic monitoring, autonomous vehicle navigation, and tracking of ballistic missile warheads. Tracking is a complex problem, often requiring to estimate motion parameters – such as position, velocity – under very challenging situations. Algorithms of this type typically have difficulty in the presence of noise, when the object is obscured, in situations including crossing trajectories, and when highly maneuvering objects are present. Most motion estimation (ME) techniques such as the ones based on block matching, optical flow, and phase difference [Jah97,280,281] assume that the object is constant from frame to frame. That is, the signature of the object does not change with time. Consequently, these techniques tend to have difficulty handling complex motion, particularly when noise is present. The time-dependent continuous wavelet transform (CWT) is attractive as a tool for analysis, in that important motion parameters can be compactly and clearly represented. The CWT maps a given image sequence into a six-dimensional representation in which position, time, scale, and velocity (speed and orientation of the velocity) are explicit parameters. In effect, this transform provides a multiscale description of motion – the conventional continuous wavelet transform in two dimensions – with additional operators to provide control over the speed and orientation (i.e., velocity) in space– time [158]. It is a multidimensional filtering in all six variables, position, time, scale, speed, and orientation. Several authors have already shown how band-pass filtering can be used to evaluate the speed in an image. A large class of models of human motion sensing use this approach [3,218,371]. The early mechanisms involved in human perception of motion appeared to be sensitive to spatial and temporal frequencies. Some neurons of the

343

344

Spatio-temporal wavelets and motion estimation

visual cortex were found to respond best when they were stimulated with temporally modulated stimuli, the temporal frequency being included in a given range. One of these neurons may respond to the presentation of a slowly moving bar, and yet not respond to the same bar moving faster in its receptive field. The organization of these filters in spatial and temporal frequencies show some special characteristics related with our spatio-temporal sensitivity. In fact, our visual system seems to make a trade-off between its spatial and temporal resolution. For instance, if one stares at a train leaving the station, at low speed it is still possible to read the destination of the cars, but at a higher speed it becomes impossible to read those details, only global shapes are still available for our vision. Actually, the interpretation of this example is complicated by the fact that our eyes (and head) may be following the train. The organization of this chapter is parallel to that of Chapter 2. We start by describing the spatio-temporal signals and motions, emphasizing that motion is in fact orientation in space–time. Next we describe the elementary operations applied to spatio-temporal filters for motion extraction. Five of these operations are the same as the ones introduced in the 1-D and the 2-D CWT (but in a spatio-temporal setting). The sixth, called speed tuning operator, is different and acts on space and time in a way that allows us to cope with the presence of motion in a spatio-temporal signal. This is then cast into the group-theoretical language, in particular, we define the appropriate unitary irreducible representation. Since the latter is square integrable, wavelets in the usual sense may be constructed. Thus we define the (2+1)-D CWT and give its properties. Finally, we describe the CWT tracking algorithm, referred to here as the Mujica–Murenzi–Leduc–Smith (MMLS) tracking algorithm, and apply it to two test sequences that reflect difficulties associated with noise, accelerated motion, temporary occlusion, and time-varying signatures. The first scene corresponds to four moving objects with linear and nonlinear motion. The experiments are performed both for noiseless and for noisy cases. The results are compared with those of block matching algorithms (BMA). The second sequence corresponds to a circularly moving object under the conditions of increasing velocity and acceleration.

10.2

Spatio-temporal signals and their transformations We will consider (2+1)-dimensional signals s (image sequences) of finite energy, represented by square integrable complex valued functions on R2 × R, i.e., s ∈ L 2 (R2 × R, d 2 x dt): d 2 x dt |s( x , t)|2 < ∞ , (10.1) s2 = R2 ×R

where x is a vector of R2 and t is the time.

345

10.2 Spatio-temporal signals and their transformations

In practice, a black and white sequence of images will be represented by a bounded non-negative function: 0 s < M < ∞. As in the spatial case, we will also consider as admissible signals some generalized functions (distributions), such as a delta function δ( x − vo t) or a plane wave exp i(ko · x + ωo t). The Fourier transform of s is defined, as usual, by −3/2 s(k, ω) = (2π ) d 2 x dt e−i(k·x −ωt) s( x , t), (10.2) R2 ×R

where k is the spatial frequency, ω is the temporal frequency, and k · x is the Euclidean scalar product. If s moves with a constant speed vo , the resulting signal s may be written as s( x , t) = s( x − vo t, t),

(10.3)

and its Fourier transform becomes ω) = ω − k · vo ) . s(k, s(k,

(10.4)

This constant speed motion does not affect the spatial frequency of the signal, but the components of its Fourier transform corresponding to the spatial frequency k are translated by −k · vo in the direction of the ω axis. This operation transforms the plane defined by ω = 0 into a new plane defined by ω = k · vo . Consider now a static signal which, in Fourier space, is roughly concentrated around the plane defined by ω = 0. When the same signal moves with constant speed vo , its Fourier transform is concentrated around the plane defined by ω − k · vo = 0

voT

1

k ω

=0

(10.5)

where voT denotes the transposed vector. Thus the velocity plane defined by (10.5) is perpendicular to the velocity vector vo . Therefore a filter which is concentrated around this plane will be sensitive to the moving components of the image which have the same speed vo . This idea is the cornerstone of the band-pass filtering approach to visual motion perception. It is directly used in spatio-temporal energy models, such as that of Adelson and Bergen [3]. Their filters can see a pattern moving with a fixed speed if the ratio of their temporal and spatial mean frequency fits with the value of the speed and if the direction of the speed and the spatial frequency are close enough. As in Chapter 2, we begin by introducing the elementary operations that we want to apply to spatio-temporal signals for motion extraction. From our past experience, we know, however, this is equivalent to transforming the filters, in particular, the wavelets, and to analyzing a given object with help of the transformed filter (this is the so-called

346

Spatio-temporal wavelets and motion estimation

passive point of view described in Section 7.1). Thus, from now on, we consider transformations to be applied to a given spatio-temporal filter (in particular, a spatio-temporal wavelet), in order to have a basis on which to decompose a given spatio-temporal signal. As we will see in next section, appropriate filters, which are wavelets in our case, must vanish at zero temporal frequency and zero spatial frequency, (10.35)–(10.36). A fortiori, a good wavelet for motion analysis will have its support in the Fourier domain concentrated in a convex cone in the half space R2 × R+ ∗ , with apex at the origin. It is also supposed to be concentrated around the plane corresponding to a fixed speed vector vo . Let us consider a spatio-temporal filter ψ (of finite energy, as usual). We consider the following elementary operations. r Spatio-temporal translations The wavelet is shifted to a given point of space and time (R2 × R). This transformation in the spatio-temporal Fourier domain: is denoted by T in direct space and T t − τ) , x , t) = ψ( x − b, (Tb,τ ψ)(

x +ωτ ) ψ)( k, ω) = e−i(k· ω). ψ(k, (T b,τ

(10.6)

τ ), is used to detect the location of obThis transformation, with parameter q = (b, k, ω), remains jects. In the Fourier domain, the wavenumber–frequency spectrum, ψ( concentrated around the vo velocity plane, only a linear phase term is introduced. r Rotation This transformation, denoted by Rθ , rotates the wavelet in spatial coordinates around the temporal (or frequency) axis. In this way, filters can be tuned to a particular orientation associated with the velocity vo . It is defined by x , t) = ψ(r−θ ( x ), t) , (Rθ ψ)(

θ ψ)( k, −θ (k), ω) = ψ(r ω), (R

where rθ is the usual 2 × 2 rotation matrix cos θ − sin θ rθ = , 0 θ < 2π. sin θ cos θ

(10.7)

(10.8)

The effect of this transformation is to change the vo -plane into a new velocity plane associated with velocity rθ ( vo ), that is, · vo = k · rθ ( vo ). ω = r−θ (k)

(10.9)

Thus, the parameter θ allows for orientation changes and is used to estimate velocities in the context of motion estimation, together with the speed tuning parameter c, to be defined next. r Scaling in wave-number– This transformation, denoted by D in space–time domain and D frequency domain, is the analog of the one we used in Chapter 2 and is defined by

347

10.2 Spatio-temporal signals and their transformations

(Da ψ)( x , t) = a −3/2 ψ(a −1 x, a −1 t),

a ψ)( k, ω) = a 3/2 ψ(a k, aω), (D

a > 0. (10.10)

The dilation preserves the norm of L 2 (R2 × R, d 2 x dt). Since it operates in the same way on both time and space, the speed of the object is not affected by the dilation. In the Fourier domain, this means that the dilation keeps the wavelet on the same constant-speed plane. r Speed tuning transformation Unlike the dilation and the shift transformations, which do not affect the concentration of the energy of the wavelet around a plane of constant speed in the Fourier space, the speed tuning transformation allows the concentration of the energy of the wavelet to move from one velocity plane with speed | vo | to a velocity plane with a different speed. in wavenumber The transformation, denoted by # in the space–time domain and # domain, can be seen as an anisotropic scaling on space and time, that is, c ψ)(k, c−β ω), ω) = ψ(cα k, (#

c > 0.

(10.11)

We require the transformation to be unitary and to map the vo -plane into the c vo plane. This results in a system of two linear equations, that fixes α and β. The first constraint gives k, k, c ψ)( 2 ω)2 = (# ω)2 = c−2α+β ψ ψ(

(10.12)

that is, 2α = β. The second constraint gives c−β ω = cα k · vo = k · cα+β vo = k · c vo .

(10.13)

This requires that α + β = 1, which together with the unitarity constraint requires that α = 23 and β = 13 . The speed tuning transformation is thus explicitly defined by (#c ψ)( x , t) = ψ(c−1/3 x, c2/3 t), c ψ)( k, 1/3 k, c−2/3 ω). ω) = ψ(c (#

(10.14) (10.15)

The speed tuning transforms a wavelet which can see a speed vo into a wavelet sensitive to c vo . Therefore, the transformation allows a direct and natural tuning of the value of the speed that the wavelet is able to analyze. In Fourier space, the transformation generates a family of wavelets which are distorted and shifted along hyperbolas defined by the ω. The parameter c of the speed tuning allows to adapt the constancy of the product |k| speed analysis independently of the scale analysis. When analyzing a moving pattern, the dilation may be used to adapt the scale of the wavelet to the spatial extension of the pattern without affecting the speed that the wavelet analysis will detect. Therefore, the τ ) can be directly values of the wavelet transform for a fixed point in time and space (b, interpreted in terms of scale (with parameter a) and speed (c and θ).

348

Spatio-temporal wavelets and motion estimation

τ, θ; a, c) ≡ Combining all these operators, one obtains the operator U (b, Tb,τ Rθ Da #c , namely, a −1 c2/3 (t − τ )). τ, θ; a, c)ψ ( x − b), (10.16) U (b, x , t) = a −3/2 ψ(a −1 c−1/3r−θ ( τ, θ; a, c)ψ}, in terms of From the wavelet ψ, one obtains a family of wavelets {U (b, which one can decompose any spatio-temporal signal s.

10.3

The transformation group and its representations Clearly, the transformations presented in the previous section form a group. In this section we will develop the wavelet machinery associated with that group. As in the case of 2-D spatial wavelets studied in Chapter 2 and in Section 7.2, the key to the construction of spatio-temporal wavelets is to have at one’s disposal a unitary representation of this transformation group in the natural space of finite energy signals, namely, L 2 (R2 × R, d 2 x dt). We are looking for a group G acting on space–time R2 × R, whose restriction to space variables coincides with the usual continuous wavelet group in two dimensions, i.e., the similitude group SIM(2), while the restriction to the time variable coincides with the usual continuous wavelet group in one dimension, i.e., the affine group G + aff . Let us consider first the group G 1 = E(2) × R, the Euclidean group (rotations and translations) acting on the (2+1)-dimensional space–time. The elements of G 1 are τ, θ ), where b is the translation vector, τ is the time translation, and θ is the triplets (b, τ, θ) ∈ G 1 acts on the point ( rotation parameter. The element (b, x , t) in the space–time 2 R × R in a natural way: t + τ ), x ∈ R2 , t ∈ R. τ, θ) : ( x ) + b, (b, x , t) → (rθ (

(10.17)

+ Next, we consider the product G 2 = R+ ∗ × R∗ of two dilation groups with the following action on space–time: + (a, c) : ( x , t) → (ac3/2 x, act), a ∈ R+ ∗ , c ∈ R∗ .

(10.18)

This action allows one to define the semidirect product G mv ≡ G 1 G 2 = {g ≡ τ, θ; a, c)}, with multiplication: (b, τ, θ; a, c)(b , τ , θ ; a , c ) = (b + ac1/3rθ (b ), τ + ac−2/3 τ , θ + θ ; aa , cc ). (b, (10.19) The inverse is given by −a −1 c2/3 τ, −θ; a −1 , c−1 ), τ, θ; a, c)−1 = (−a −1 c−1/3r−θ (b), (b,

(10.20)

349

10.3 The transformation group and its representations

0, 0; 1, 1). In 4 × 4 matrix notation, we may write and the unit element is e = (0, (compare (7.37)): 1/3 ac rθ 0 b τ, θ; a, c) ≡ (b, (10.21) ac2/3 τ . 0 0 0 1 The group G mv is locally compact with right Haar measure dµR and left Haar measure dµL : da dc da dc , dµL = d 2 b dτ dθ 4 . (10.22) a c a c In the sequel, we will use systematically the left Haar measure dµL and, correspondingly, we will write 2π ∞ da ∞ dc 2 . (10.23) dµL (b, τ, θ; a, c) ≡ d b dτ dθ a4 0 c G mv R2 R 0 0 dµR = d 2 b dτ dθ

Having identified the appropriate transformation group G mv , we proceed to find a square integrable representation of it, in the Hilbert space of finite energy signals, ac τ, θ; a, c) ∈ cording to the general formalism sketched in Section 7.1. For every g ≡ (b, 2 2 2 G mv , consider the operator U (g) on L (R × R, d x dt) defined in (10.16), namely, a −1 c2/3 (t − τ )). x − b), x , t) = a −3/2 ψ(a −1 c−1/3r−θ ( [U (g)ψ] (

(10.24)

This may be written in the Fourier space as $ % b+ωτ ) (g)ψ (k, ac−2/3 ω). ω) = a 3/2 e−i(k· ψ(ac−1/3r−θ (k), U

(10.25)

Proposition 10.3.1 The operator U defines a unitary representation of G mv in L 2 (R2 × R, d 2 x dt). This is proven by a straightforward verification. defines two unitary irreducible representations of Proposition 10.3.2 The operator U 2 G mv in H+ ≡ H+ (R × R) and H− ≡ H− (R2 × R), respectively, where k, ω) = 0, ω < 0} H+ (R2 × R) = {ψ ∈ L 2 (R2 × R, d 2 x dt) : ψ( k, ω) = 0, ω > 0} H− (R2 × R) = {ψ ∈ L 2 (R2 × R, d 2 x dt) : ψ(

(10.26) (10.27)

with L 2 (R2 × R, d 2 x dt) = H+ (R2 × R) ⊕ H− (R2 × R).

(10.28)

Proof . The proof follows the same line as that of Proposition 2.1.2. Let ψ ∈ H+ be an arbitrary nonzero vector. We are going to show that ψ is cyclic for the representation

350

Spatio-temporal wavelets and motion estimation

U , i.e., the linear span of the orbit of ψ is dense in H+ . Let f be orthogonal to the span of {U (g)ψ}, that is, U (g)ψ| f = 0, ∀ g ∈ G mv . We show that f = 0. We have τ, θ; a, c) ∈ G mv , indeed, for all (b, (g)ψ| U (g)ψ| f = U f −1/3r−θ (k), ac−2/3 ω) d 2 k dω ei(k·b+ωτ ) ψ(ac = a 3/2 R 2 ×R

= 0. −1/3r−θ (k), ac−2/3 ω) ω) vanishes This means that the Fourier transform of ψ(ac f (k, (almost everywhere) for all θ, a, c. Then the transitivity of the action of (θ, a, c) ∈ ω) = 0 (almost everywhere). SO(2) × R∗+ × R∗+ on R2 × R∗+ implies that f (k, The same holds true for H− . Proposition 10.3.3 The representation U is square integrable, that is, there exists a function ψ ∈ L 2 (R × R, d 2 xdt), ψ = 0, such that the matrix ele τ, θ; a, c)ψ|ψ is square integrable with respect to the measure ment U (b, τ, θ; a, c), that is, in the notation of (10.23), dµL (b,

τ, θ; a, c)|U (b, τ, θ; a, c)ψ|ψ|2 < ∞. dµL (b,

I =

(10.29)

G mv

In addition, I = cψ ψ2

(10.30)

where cψ = (2π)

3 R 2 ×R

d 2 k dω |ψ(k, ω)|2 . 2 |ω| |k|

Proof . The result follows from a direct calculation: I =

τ, θ; a, c)ψ|ψ|2 dµL |U (b, G mv

(b, ψ| 2 τ, θ; a, c)ψ| dµL |U b+ωτ ) −1/3r−θ (k), ac−2/3 ω) e−i(k· ψ(k, ω) = dµL a 3 d 2 k dω ψ(ac =

G mv

G mv

×

R2 ×R

R2 ×R

τ ) −1/3r−θ (k ), ac−2/3 ω ) ei(k ·b+ω d 2 k dω ψ(ac ψ(k , ω )

(10.31)

351

10.3 The transformation group and its representations

=

2π

0

∞

0

∞

0

da dc dθ a c

×

2

d bdτ e

2

d k dω R2 ×R

k )·b+(ω−ω −i{(k− )τ }

d 2 k dω

R 2 ×R

R2 ×R

k, k , ω ) ω) ψ( ψ(

−1/3r−θ (k ), ac−2/3 ω ) −1/3r−θ (k), ac−2/3 ω) ψ(ac × ψ(ac = (2π)3

2π

∞

∞

dθ 0

0

0

× = (2π)3

R2 ×R

da dc a c

−1/3r−θ (k), k, ac−2/3 ω)|2 |ψ( ω)|2 d 2 k dω |ψ(ac

R2 ×R

k, ω)|2 d 2 k dω|ψ(

2π

× 0

0

∞

0

∞

! da dc −1/3 −2/3 2 dθ r−θ (k), ac ω)| . |ψ(ac a c

Perform now the following change of variables: k = ac−1/3r−θ (k),

−2/3 ω = ac ω,

(10.32)

which is equivalent to −1/3 k1 ac k1 k2 0 cos θ k2 = −k2 k1 0 ac−1/3 sin θ .

−2/3 0 0 ω ac ω Thus the measure dθ

(10.33)

d k dω da dc becomes and we have a c |k |2 |ω |

k , ω )|2 |ψ( k, ω)|2 d 2 k dω |ψ(

2

2 2 | k | |ω | R ×R R ×R 2 k, | ψ( ω)| = (2π )3 d k dω ψ2 . 2 2 | k| |ω| R ×R

I = (2π )3

d k dω

It is clear that the integral I converges if and only if ψ satisfies the admissibility condition (10.29).

352

Spatio-temporal wavelets and motion estimation

10.4

The spatio-temporal wavelet transform

10.4.1 Spatio-temporal wavelets: definition and examples By definition, a spatio-temporal wavelet [158,Duv91] is a complex-valued function ψ ∈ L 2 (R2 × R, d 2 x dt) satisfying the condition: d 2 k dω 3 (10.34) |ψ(k, ω)|2 < ∞ cψ = (2π) 2 |ω| R2 ×R |k| is the Fourier transform of ψ. where ψ If ψ is regular enough, the admissibility condition (10.34) simply means that the wavelet must be of zero mean with respect to space and time independently: ψ(0, ω) = 0 ⇐⇒ d 2 x ψ( x , t) = 0. (10.35) R2

and k, 0) = 0 ⇐⇒ ψ(

R

dt ψ( x , t) = 0.

(10.36)

Clearly the unitary operators T(b,τ ) , Rθ , Da , #c preserve the admissibility condition, τ, θ; a, c). Hence any function ψb,τ,θ and so does therefore U (b, ;a,c = U (b, τ, θ ; a, c)ψ obtained from a wavelet ψ by translation, dilation, rotation, or speed tuning is again a wavelet. Thus the given wavelet ψ generates the whole family {ψb,τ,θ ;a,c }, indexed by the 2 elements (b, τ ) ∈ R × R, θ ∈ [0, 2π ), a > 0, c > 0, that is, by the elements of G mv . We will consider only one example of spatio-temporal wavelet, the spatiotemporal Morlet wavelet, characterized by the wavenumber–frequency (k0 , ω0 ) and the anisotropy parameter $ and defined by −1 1 1 1 2 −1 2 −1 2 ψ$ ( x , t) = ei k0 ·A x e− 2 |A x| − e− 2 |A x| e− 2 |k0 | 1 2 1 2 1 2 (10.37) × eiω0 t e− 2 t − e− 2 t e− 2 ω0 in the spatio-temporal domain and by 1 1 1 2 2 2 k0 |2 2 2 $ (k, ω) = e− 12 |Ak− − e− 2 (|Ak| +|k0 | ) e− 2 (ω−ω0 ) − e− 2 (ω +ω0 ) ψ

(10.38)

in the wavenumber–frequency domain, where A = diag[$ −1/2 , 1], $ 1, is the usual anisotropy matrix. As in the pure spatial case, for large |k0 | and |ω0 |, typically |k0 | 6, |ω0 | 6, the counterterms (the second term in each factor of (10.37) and (10.38)) are small enough to be neglected. This wavelet is a good candidate for motion estimation applications. Its region of support can be appropriately located around a particular plane, vo , typically vo = (1, 0), for example in the case where k0 = (k0 , 0), k0 = ω0 . Figure 10.1 shows half-energy equisurfaces of the squared amplitude of Morlet wavelets for different values of the parameters, θ, a and c. This illustrates how the

353

10.4 The spatio-temporal wavelet transform

ω

ω

kx

kx ky

ky

(a)

(b)

ω

kx ky

(c)

(d)

Fig. 10.1. Wavenumber–frequency domain coverage of the spatio-temporal Morlet wavelet for different parameters of the CWT. (a) Rotation θ; (b) scale a; (c) speed tuning c; (d) rotation, scale, and speed tuning θ, a, c.

wavelet parameters distribute the energy of the resulting filter on the wavenumber– frequency domain. The energy is distributed around a circle by the rotation parameter, along a conic volume by the scale parameter, and along a hyperbolic-like path by the speed tuning parameter. In order to control the variance of the wavelet with respect to the reference velocity plane, an anisotropy parameter $ applied to the spatial variables is used. This indeed controls its temporal support. As it can be seen from Figure 10.1, the region of support of the filters is an ellipsoidal cone concentrated around a particular velocity plane. Thus, velocity detection and filtering are possible.

10.4.2 The spatio-temporal wavelet transform Let now s ∈ L 2 (R2 × R, d 2 xdt) be an image sequence. Its (2+1)-D continuous wavelet transform (with respect to the fixed wavelet ψ), S ≡ Wψ s is the scalar product of s τ, θ; a, c)ψ, considered as a function with the transformed wavelet ψb,τ,a,c,θ ≡ U (b, of (b, τ, a, c, θ ):

354

Spatio-temporal wavelets and motion estimation

τ, θ; a, c) = √1 < ψb,τ,θ;a,c S(b, |s > (10.39) cψ 1 a −1 c2/3 (t − τ )) s( =√ x − b), x , t) d 2 x dt a −3/2 ψ(a −1 c−1/3 rθ ( cψ (10.40) 1 1/3 rθ (k)ac −2/3 ω) ω). d 2 k dω a 3/2 e−i(k·b+ωτ ) ψ(ac =√ s(k, (10.41) cψ The main properties of the (2+1)-D CWT Wψ : s → S may be summarized as follows [Com89,Mey91]: (i) As the purely spatial version, Wψ is linear in the signal s; (ii) Wψ is covariant under all the operations considered, namely [Com89,Mur90,13]. Proposition 10.4.1 The map Wψ is covariant under translations, rotations and τ, θ; a, c) dilations, which means that the correspondence Wψ : s( x , t) → S(b, implies the following ones: s( x − bo , t) → S(b − bo , τ, θ; a, c) τ − to θ; a, c) s( x , t − to ) → S(b,

(10.42)

τ, θ − θo ; a, c). s(r−θo ( x ), t) → S(r−θo (b), −1 −1 −1 ao−1 τ, θ ; ao−1 a, c) ao s(ao x, ao t) → S(ao−1 b,

(10.44)

co−2/3 τ, , θ ; a, co−1 c) S(co1/3 b,

(10.46)

s(co−1/3 x, co2/3 )

→

(10.43) (10.45)

It is worth noting that, conversely, the wavelet transform is uniquely determined by the three conditions of linearity, covariance and energy conservation, plus some continuity [Mur90]. (iii) Energy conservation: 1 2 2 τ, θ; a, c)|U (b, τ, θ; a, c)ψ | s|2 . d x dt |s( x , t)| = dµL (b, cψ G mv R 2 ×R (10.47) Thus, Wψ is an isometry from the space of signals into the space of transforms. (iv) As a consequence, Wψ is invertible on its range and the inverse transformation is simply the adjoint of Wψ . Thus one has an exact reconstruction formula: 1 τ, θ; a, c)ψb,τ,θ τ, θ; a, c). dµL (b, x , t) S(b, (10.48) s( x , t) = ;a,c ( cψ G mv In other words, the (2+1)-D wavelet transform, like its 1-D and 2-D counterpart, provides a decomposition of the signal in terms of the analyzing wavelets ψb,τ,θ ;a,c , τ, θ; a, c). with coefficients S(b, (v) Redundancy: Exactly as in the 2-D case discussed in Chapter 2, one has: Proposition 10.4.2 The projection from L 2 (G mv , dµL ) onto the range Hψ of

355

10.4 The spatio-temporal wavelet transform

Wψ , the space of wavelet transforms, is an integral operator whose kernel τ, θ; a, c) is the autocorrelation function of ψ, also called K (b , τ , θ ; a , c | b, reproducing kernel: τ, θ; a, c) = cψ−1 ψb ,τ ,θ ;a ,c | ψb,τ,θ K (b , τ , θ ; a , c | b, ;a,c .

(10.49)

Therefore, a function f ∈ L 2 (G, dg) is the wavelet transform of a certain signal iff it satisfies the reproduction property: f (b , τ , θ ; a , c ) τ, θ; a, c)K (b , τ , θ ; a , c | b, τ, θ; a, c) f (b, τ, θ; a, c). dµL (b, = G mv

(10.50)

10.4.3 An alternative: relativistic wavelets The spatio-temporal wavelets just described, which could be called kinematical, may not always be sufficient, depending on the type of signal to be analyzed. One may wish to consider a specific form of movement, i.e., choose a particular relativity group. Three examples may be of interest (we begin again with one space dimension). (i) Galilean wavelets Here we add to the transformations discussed above the Galilei boosts, thus getting (x, t) → (a1 x + vt + b1 , a0 t + b0 ). The resulting group G aff 1 , called the affine Galilei group, is quite complicated. It has a natural unitary representation in the space of finite energy signals, which splits into the direct sum of four irreducible ones, and each of these is square integrable, so that wavelets may be constructed in the usual way [25]. In addition, more restricted wavelets may be constructed by taking as parameter space various quotient spaces G aff 1 /H , where H is not the stability subgroup of the basic wavelet. (ii) Schr¨odinger wavelets One obtains an interesting subclass of the previous one by imposing the relation a0 = a12 , so that the transformations leave invariant the Schr¨odinger (or the heat) equation. Then the unitary irreducible representation UG of G aff 1 splits into the direct sum of two square integrable ones of the (Schr¨odinger) subgroup. Thus again a CWT is at hand, which may prove useful for describing, for instance, the motion of quantum particles on the line. (iii) Poincar´e wavelets In order to get a CWT in the relativistic regime, it suffices to replace Galilei transformations by Poincar´e ones, while of course imposing the relation a0 = a1 to space and time dilations. The result is the affine Poincar´e group, that we have discussed at length in Section 7.4. The Poincar´e wavelets might be useful, for instance, in the presence of electromagnetic fields.

356

Spatio-temporal wavelets and motion estimation

Of course, this analysis extends in a straightforward way to higher dimensions, just by adding rotations. Details may be found in [Ali00; Section 15.3]. These three types of relativistic wavelets offer additional examples of the general group-based wavelet formalism. They have a definite mathematical interest, but they have not been tested on practical situations, indeed no motion estimation (ME) algorithm based on them has been designed. On the contrary, the kinematical spatio-temporal CWT does lead to an efficient ME algorithm, that we now describe in detail.

10.5

A motion estimation (ME) algorithm We shall now exploit the general CWT formalism developed in the previous section and describe an algorithm for motion estimation. A complete discussion can be found in [Muj99,281,282]. Velocity filtering approaches for motion estimation allow temporal information to be incorporated in the estimation process. These techniques have performance advantages over two-frame-at-a-time based approaches, like block matching and optical flow [Tek95,353] in nonideal environments. The spatio-temporal CWT facilitates adaptive velocity filtering and offers an elegant framework for motion analysis and estimation. It performs a mapping from the Hilbert space H to a parameter space meaningful for motion estimation purposes. It can also be seen as a tool for motionbased filtering, where the filter characteristics are appropriately determined by a set of parameters directly associated with motion features. The wavelet basis matches the motion characteristics of the object of interest rather than its spatial features. General spatial selectivity is taken into account through the scale parameter a. It is assumed that starting conditions, i.e., position and velocity, of the object of interest are known initially. Our ME algorithm deals with the problem of following timevarying motion parameters on a frame-by-frame basis, which allows us to determine object coordinates at any time [Muj99]. In this sense, our ME algorithm can be viewed as an object tracking algorithm after the initial detection has been performed. This section is organized in two parts. First, we formulate the rˆole of the spatio-temporal CWT as a motion parameter estimator and we describe the three partial energy densities used for this purpose. In the second part, we describe how the interaction of these energy densities can be used to track a particular object from incoming video. Our CWT-based tracking algorithm is then introduced.

10.5.1 Partial energy densities The multidimensional nature of the spatio-temporal CWT allows for the definition of a multitude of energy densities either by fixing a subset of the parameter space or, better, by partial integration of the CWT energy,

357

10.5 A motion estimation (ME) algorithm

+ +2 + +2 E(g) ≡ E[s](g) = +ψg |s+ = + Sψ (g)+ ,

τ, θ; a, c}, g = {b,

(10.51)

on subsets of the parameter space. As discussed at length in the spatial case, in Section 2.3.4, this approach has the nice property that it can result in invariant representations with respect to some parameters. Thus, partial integration on subsets of the parameter space results in different energy representations that can be used to extract relevant features. Three energy densities particularly interesting for motion estimation purposes are studied here. (1) Speed-orientation energy density: Here integration is performed over spatial translation, b = (bx , b y ), on a region B : (bxmin < bx < bxmax ) ∩ (b ymin < b y < b ymax ),

(10.52)

while the scale a and the temporal variable τ are fixed. As a result, we obtain the first energy density, +9 :+2 + EaI o ,τo (c, θ) = d 2 b + ψb,τ (10.53) o ,θ;ao ,c | s , b∈B

which can be interpreted as an estimator of local velocity. The boundaries of the spatial region B are updated on a frame-by-frame basis to reflect changes in object location. This allows the algorithm to focus on the object of interest, and reduce interference with other nearby objects. (2) Spatial energy density: In this representation, the speed tuning parameter c, the orientation θ, the scale a, and the temporal translation τ , are fixed, while the spatial translation, b is the variable of interest. The resulting energy density is given by, = EIIτo ,θo ;ao ,co (b)

:+2 1 ++9 + ψb,τ o ,θo ;ao ,co | s . 4 ao

(10.54)

We can think of this energy density as the output energy of a velocity selective filter, where the location of objects moving at a pre-specified velocity, v = co eiθo , can be easily determined. In addition, size selectivity (invariability) can be obtained by appropriately choosing (integrating over) the scale parameter a. (3) Scale energy density: The scale parameter is associated with spatial size. Consequently, there exists an optimum scale aopt that best matches the size of the object of interest. A measure of “scale optimality” can be defined by integrating the global energy density, E, over while fixing the temporal translation τ , and the velocity the spatial variables, b, parameters, c and θ, leading to +9 :+2 1 III + Eco ,θo ,τo (a) = 4 d 2 b + ψb,τ (10.55) o ,θo ;a,co | s . a b∈B

358

Spatio-temporal wavelets and motion estimation

The spatial integration is constrained to the region B to avoid interference with nearby objects. These energy densities can be used to derive local estimates of the motion parameters. It is noteworthy that these energy densities are different from the global energy densities where additional integration is performed over the temporal parameter τ [279]. The global approach can only handle linear motion, while the local approach can deal with accelerated motion and time varying signatures as well. The energy densities presented here are the computational core of the CWT-based tracking algorithm, which is presented in the next section.

10.5.2 Description of the algorithm Computer implementation of the CWT requires discretization of the spatial and temporal variables ( x , t). The energy densities EI , EII , and EIII of (10.53), (10.54), and (10.55) must then be discretized by replacing integrations with appropriate summations. It is assumed the input data is presented as an incoming video signal with one or more objects moving with a constant or accelerated speed. For instance, the l-th object in a given image sequence is denoted sl ( xn,m , ti ), where n and m are spatial indices (horizontal and vertical respectively), and i is the temporal index. The resulting model for the input image sequence is then s( xn,m , ti ) = sl ( xn,m , ti ) + w( xn,m , ti ), (10.56) l

where w is assumed to be zero-mean white noise. The sum in (10.56) must be interpreted carefully when occlusions occurs. In this case, if the point ( xn,m , ti ) is subject to occlusion from two or more objects, the correct expression is s( xn,m , ti ) = s L ( xn,m , ti ) + w( xn,m , ti ),

(10.57)

where L is the index denoting the object closest to the sensor. The CWT-based ME algorithm relies on the three energy densities defined above and consists of a frame-by-frame optimization of the motion parameters associated with a given object in the image sequence. These motion parameters are gathered in a state vector L(ti ) defined as T (10.58) L(ti ) = vt i , xt i , ati , for time t = ti . Note the convention we use for the position variable; xn,m represents the spatial location indexed by the pair (n, m), while xti represents the spatial state at time t = ti . We assume that the starting position, x0 , and velocity, v0 , for the object of interest are known initially. The CWT energy densities are used here as optimality criteria or cost functions for updating the state vector L. This update can be done either by searching for the local

359

10.5 A motion estimation (ME) algorithm

maximum or by a gradient-based approach like the LMS algorithm, on each of the energy densities. In either case, this process is denoted symbolically by EI EII EIII

L(ti ) −→ L(ti+1 ) .

(10.59)

The optimization of equation (10.59) is performed sequentially. That is, when optimizing one component of the state vector (i.e., v, x, or a), the others are kept constant. Indeed, we are searching for an optimal set of motion parameters in the 6-D CWT space. The sequential optimization of the parameters considerably reduces the search space and consequently the computational requirements (with respect to a simultaneous 6-D search). This is possible due to the implicit redundancy of the CWT representation and its ability to isolate motion features. Thus, each of the energy densities corresponds to a 2-D (or 1-D for the scale energy density) “slice” of the CWT parameter space, where one parameter is optimized independently of the others. The CWT can be seen as a spatio-temporal filtering operation where the filter characteristics are controlled by a set of parameters associated with motion features (i.e., velocity and size). More explicitly, manipulating the inner product of equation (10.40) the CWT can be defined as a convolution sum, i.e., t − τ) Sψ (g) = d 2 x dt s( x , t) ψθ,a,c ( x − b, R 2 ×R # d 2 x dt s( x , t) ψθ,a,c (b − x, τ − t) = R 2 ×R

τ ), = s ⊗ ψ θ,a,c (b, where the symbol ⊗ represents the (2+1)-D convolution operator, t − τ ) = ψb,τ,θ;a,c ψθ,a,c ( x − b, ( x , t), and # ( x , t) = ψθ,a,c (− x , −t). ψθ,a,c

The indices of the filtered signal correspond to the spatio-temporal translation parameters of the CWT (i.e., b and τ ). This allows us to take advantage of the Fast Fourier Transform (FFT) to compute the CWT (and its associated energy densities) efficiently in the wavenumber–frequency domain. The separability of the wavelets can be exploited to reduce the required computations to construct the motion and scale selective filters. A block diagram of the state updating process performed by the CWT-based ME algorithm is depicted in Figure 10.2. The N × M × K block of data represents a portion of the incoming video signal with K frames of N columns by M rows. This image sequence is first transformed to the Fourier domain by means of a 3-D DFT. As suggested by Figure 10.2 and equation (10.59) the parameter update is done in a specified order: velocity first, then position, and finally scale. Two arguments support this implementation choice. First, velocity is a motion parameter of higher order than position. For highly maneuvering objects, the relative change in velocity from frame to

360

Spatio-temporal wavelets and motion estimation

K N

3-D FFT

...

M

Spatio-Temporal CWT

Input image sequence

L(n)

Speed-Orientation Energy Density

L(n)’

Space-Time Energy Density

L(n)’’

Scale (size) Energy Density

L(n+1)

Fig. 10.2. CWT-based tracking algorithm.

frame is smaller than the relative change in position. Second, some of the calculations necessary for determining the position energy density are already done at the end of the speed-orientation energy density stage. The three update stages embodied in the tracking algorithm are now described in detail. It is important to point out that the diagram shown in Figure 10.2 represents the operations performed in order to update the state vector L(ti ) for each frame of a given data block.

10.5.2.1

Velocity update stage In this stage the speed-orientation energy density EI is used as the optimality criterion to update the velocity state at time ti+1 . The position and scale states are fixed to their corresponding values at time ti , which are, xti and ati respectively. The discrete version of the speed-orientation energy density EI is

Two-Dimensional Wavelets and their Relatives

Two-dimensional wavelets offer a number of advantages over discrete wavelet transforms when processing rapidly varying functions and signals. In particular, they offer benefits for real-time applications such as medical imaging, fluid dynamics, shape recognition, image enhancement and target tracking. This book introduces the reader to 2-D wavelets via 1-D continuous wavelet transforms, and includes a long list of useful applications. The authors then describe in detail the underlying mathematics before moving on to more advanced topics such as matrix geometry of wavelet analysis, three-dimensional wavelets and wavelets on a sphere. Throughout the book, practical applications and illustrative examples are used extensively, ensuring the book’s value to engineers, physicists and mathematicians alike. Jean-Pierre Antoine is a Professor of Mathematical Physics at the Institut de Physique

Th´eorique, Universit´e Catholique de Louvain. Romain Murenzi is currently Minister of Education, Science, Technology, and Scientific

Research of the Republic of Rwanda, on leave of absence from the Department of Physics, Clark Atlanta University, Atlanta, Georgia. Pierre Vandergheynst is a Professor at the Signal Processing Institute, Swiss Federal

Institute of Technology, Lausanne. Syed Twareque Ali is a Professor at the Department of Mathematics and Statistics, Concordia University, Montr´eal.

Two-Dimensional Wavelets and their Relatives Jean-Pierre Antoine Institut de Physique Th´eorique, Universit´e Catholique de Louvain

Romain Murenzi CTSPS, Clark Atlanta University, Ministry of Education, Science, Technology and Scientific Research, Rwanda

Pierre Vandergheynst Signal Processing Laboratory, Swiss Federal Institute of Technology

Syed Twareque Ali Department of Mathematics and Statistics, Concordia University

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge , UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521624060 © J.-P. Antoine, R. Murenzi, P. Vandergheynst and S. Twareque Ali 2004 This publication is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format 2004 - -

---- eBook (EBL) --- eBook (EBL)

- -

---- hardback --- hardback

Cambridge University Press has no responsibility for the persistence or accuracy of s for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Prologue

page ix

1

Warm-up: the 1-D continuous wavelet transform

1.1 1.2 1.3 1.4 1.5 1.6 1.7

What is wavelet analysis? The continuous wavelet transform Discretization of the CWT, frames Ridges and skeleton The discrete WT: orthonormal bases of wavelets Generalizations Applications of the 1-D CWT

1 5 10 14 19 24 29

2

The 2-D continuous wavelet transform

32

2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8

Derivation Basic properties of the 2-D CWT Implementation and interpretation of the 2-D CWT Discretization, frames Comparison with the 2-D discrete wavelet transform Bridging the gap: continuous wavelet packets and fast algorithms Steerable filters Redundancy: plus and minus

32 36 41 54 68 78 93 96

3

Some 2-D wavelets and their performance

97

3.1 3.2

Which wavelets? Isotropic wavelets

97 99

v

1

vi

Contents

3.3 3.4

Directional wavelets Wavelet calibration: evaluating the performances of the CWT

103 118

4

Applications of the 2-D CWT. I: image processing

125

4.1 4.2 4.3 4.4 4.5 4.6 4.7

Contour detection, character recognition Object detection and recognition in noisy images Image retrieval Medical imaging Detection of symmetries in patterns Image denoising Nonlinear extensions of the CWT

125 134 145 150 150 162 163

5

Applications of the 2-D CWT. II: physical applications

175

5.1 5.2 5.3 5.4 5.5 5.6

Astronomy and astrophysics Geophysics Applications in fluid dynamics Fractals and the thermodynamical formalism Texture analysis Applications of the DWT

175 192 197 205 210 212

6

Matrix geometry of wavelet analysis. I

214

6.1 6.2 6.3

Group theory and matrix geometry of wavelets Phase space analysis The case of Gabor wavelets

214 230 238

7

Matrix geometry of wavelet analysis. II

247

7.1 7.2 7.3 7.4

A group-adapted wavelet analysis The 2-D continuous wavelet transform 2-D wavelets on phase space The affine Poincar´e group

247 259 268 276

vii

Contents

8

Minimal uncertainty and Wigner transforms

281

8.1 8.2 8.3 8.4

Phase space distributions and minimal uncertainty gaborettes Minimal uncertainty wavelets Wigner functions Wigner functions for the wavelet groups

281 284 287 291

9

Higher-dimensional wavelets

300

9.1 9.2 9.3

Three-dimensional wavelets Wavelets on the 2-sphere and other manifolds Wavelet approximations on the sphere

300 308 332

10

Spatio-temporal wavelets and motion estimation

343

10.1 10.2 10.3 10.4 10.5

Introduction Spatio-temporal signals and their transformations The transformation group and its representations The spatio-temporal wavelet transform A motion estimation (ME) algorithm

343 344 348 352 356

11

Beyond wavelets

373

11.1 11.2 11.3 11.4

New transforms: ridgelets, curvelets, etc. Rate-distortion analysis of anisotropic approximations Sparse approximations in redundant dictionaries Algebraic wavelets

374 383 389 398

Epilogue Appendix Some elements of group theory References Index

413 415 431 455

Prologue

Wavelets are everywhere nowadays. Be it in signal or image processing, in astronomy, in fluid dynamics (turbulence), in condensed matter physics, wavelets have found applications in almost every corner of physics. In addition, wavelet methods have become standard in applied mathematics, numerical analysis, approximation theory, etc. It is hardly possible to attend a conference on any of these fields without encountering several contributions dealing with them. Correspondingly, hundreds of papers appear every year and new books on the topic get published at a sustained pace, with publishers strongly competing with each other. So, why bother to publish an additional one? The answer lies in the finer distinction between various types of wavelet transforms. There is, indeed, a crucial difference between two approaches, namely, the continuous wavelet transform (CWT) and the discrete wavelet transform (DWT). Furthermore, one has to distinguish between problems in one dimension (signal analysis) and problems in two dimensions (image processing), since the status of the literature is very different in the two cases. Take first the one-dimensional case. Beginning with the classic textbook of Ingrid Daubechies [Dau92], several books, such as those of M. Holschneider [Hol95], B. Torr´esani [Tor95] or A. Arn´eodo et al. [Arn95], cover the continuous wavelet transform, in a more or less mathematically oriented approach. On the other hand, the discrete wavelet transform is treated in many textbooks, more in the signal processing style, such as M. V. Wickerhauser [Wic94], M. Vetterli and J. Kovaˇcevi´c [Vet95], P. Wojtaszczyk [Woj97], or S. G. Mallat [Mal99], whereas others emphasize the algorithmic aspects, sometimes in a rather abstract way, for example, C. K. Chui [Chu92] or Y. Meyer [Mey94] (of course, there are many more on the market). Altogether these books tell a fascinating story, that is ideally depicted in the highly popular volume of B. Burke Hubbard [Bur98], which is based on interviews by the author with all the founding “fathers” of the theory (J. Morlet, A. Grossmann, I. Daubechies, Y. Meyer, etc.). It is a fact that DWT-inspired methods (multiresolution, lifting scheme, etc., that we shall describe in due time) constitute the overwhelming majority among the wavelet community, under the joint influence of electrical engineering (signal processing with ix

x

Prologue

filters and subband coding) and applied mathematics (numerical and algorithmic methods). Yet the CWT and, more generally, redundant representations of signals, offer distinct advantages in certain cases, as we shall see later. In two dimensions, that is, application to image processing, the situation is clearer. Discrete methods are somewhat trivial, since the basic structure is that of a tensor product, 2-D = 1-D ⊗ 1-D, enforcing a Cartesian geometry (x and y coordinates). Thus most textbooks on the DWT will cover, although briefly in general, the 2-D case as a straightforward extension of the 1-D setup. As for the 2-D CWT, it receives at best a cursory treatment in most cases. The raison d’ˆetre of the present volume is precisely to fill this gap in the literature and give a thorough treatment of the 2-D CWT and some of its applications in image processing and in various branches of physics. As a byproduct, we will also discuss in detail several extensions, such as 3-D wavelets, wavelets on the sphere or wavelets in space-time.

A historical note Before entering the subject proper, it may not be uninteresting to give some details on its origin, without pretension to completeness, of course; we are not historians. The first extension of the wavelet transform to imaging is due to Mallat [259,260], who developed systematically a 2-D discrete (but redundant) WT, combining the traditional concept of filter bank and the analogy with human vision. In fact, most of the concepts are indeed already present in the pioneering work of Marr [Mar82] on vision modeling, in particular the idea of multiresolution. Indeed, when we look at an object, our visual system works by registering first a global, low-resolution, image and then focusing systematically to finer and finer details. Thus, contrary to the 1-D case, the 2-D discrete WT preceded the continuous version. The 2-D continuous WT was born in a quite different way. The story starts in the coffee room of the Institut de Physique Th´eorique in UCL, Louvain-la-Neuve (LLN), in Spring 1987. Alex Grossmann from Marseille, one of the founding fathers of wavelets, was visiting J.-P. A., indeed they had already started to collaborate on the application of 1-D wavelets in NMR spectroscopy. Thus the two were discussing a possible Ph.D. topic for a young African student, called Romain Murenzi (R.M.). The latter had just concluded a Master’s thesis on five-dimensional quantum field theory, a subject hardly practical for a developing country! So the idea came up, why not try to do in two dimensions what had been so successful in 1-D, namely, wavelet analysis? The topic seemed tractable, involving moderate amounts of mathematics and some simple computing technology, and if it worked out, there could be very interesting practical applications. The problem was that nobody knew how to do it! The next summer, R.M. went down to Marseille and started to work with Grossmann and Ingrid Daubechies who happened to be there too. And when he came back 3 months later, the solution was clear. The key

xi

Prologue

is to start from the operations that one wants to apply to an image, namely, translations in the image plane, rotations for choosing a direction of sight, and global magnification (zooming in and out). The problem is to combine these three elements in such a way that the wavelet machine could start rolling (there are mathematical conditions to satisfy here). The result of R.M. was that the so-called similitude group yields a solution (actually, the only one). There remained to put it all together, to turn the mathematical crank and to apply the resulting formalism to a real problem, namely, 2-D fractals (the outcome of a visit of R.M. to Arn´eodo in Bordeaux), and the Ph.D. thesis was within reach [Mur90]. Several papers followed [12,13], more M.Sc. or Ph.D. students got involved over the years. We may cite Pierre Carrette, St´ephane Maes, Canisius Cishahayo, Pierre Vandergheynst, Y´eb´eni B. Kouagou, Laurent Jacques, Laurent Demanet. Each of them has brought his contribution to the edifice, small or big, but always useful. This is probably a good place for asking, why wavelets? After all, there are plenty of methods available for processing images. What is new here? A key fact is probably that wavelets are somehow a byproduct of quantum thinking. More precisely, it is an application of the quantum idea of a probe for testing an object, the result being given by the scalar product of the two functions (indeed the framework is a Hilbert space, that of finite energy signals). To get the transform, the probe is translated and scaled (zoom), and turned around in the 2-D case, and the result is plotted as a function of the corresponding parameters. (Actually the same could be said of the so-called Gabor or Windowed Fourier transform.) One gets in this way a highly flexible and efficient tool for signal/image processing, that sheds a different light and offers an alternative approach to many standard problems, in particular those involving the detection of singularities or discontinuities in signals. As somebody once remarked, wavelets do not solve all the problems, but they often help asking the right questions. Another sign of the quantum influence is the crucial role played by a unitary group representation, a tool largely absent in classical physics – and thus from signal processing as well. And it is no accident, in our opinion, that the crucial steps in developing wavelets were made by Alex Grossmann and Ingrid Daubechies, both educated as theoretical (quantum) physicists. Otherwise, it might have taken much longer for electrical engineers and mathematicians to meet!

About the contents of the book Now it is time to give some indications on the contents of the book. One can divide it into several stages. In a first part (Chapters 1–3), we develop systematically the continuous wavelet transform, first in one dimension (briefly), then in two dimensions. The emphasis here is on the practical use of the tool, with a minimum of mathematics. Then we devote two long chapters, 4 and 5, to applications. Three short chapters, 6–8,

xii

Prologue

set the general mathematical scene. This allows us, in Chapters 9 and 10, to describe wavelets in more general settings (3-D, sphere, space–time). In Chapter 11, finally, we discuss some recent developments that actually go beyond wavelets. This gradual structure is one of the original aspects of the book, in comparison with those on the market. Let us go into more details. As a warming up exercise, we begin, in Chapter 1, with a rather concise overview of the 1-D WT. This allows the reader to develop a feeling about the wavelet transform and to understand its success in signal processing. All aspects will be touched upon: the continuous WT, multiresolution and the discrete WT, various generalizations of the latter, some applications. One of the leitmotives is the role of redundancy, especially with respect to stability of the representation. Chapter 2, which forms the hard core of the first part, presents in a systematical way the theory of the 2-D CWT. As said above already, the starting point is to decide which elementary operations one wants to apply to an image. Choosing translations in the image plane, rotations (direction of sight), and global magnification (zooming in and out), together with the probe idea, leads uniquely to the 2-D CWT. We study in detail its basic properties: energy conservation, reconstruction formula, reproducing property, covariance under the chosen operations. Then we describe the interpretation of the WT as a singularity scanner and as a phase space representation of signals. Since the WT of a 2-D image is a function of four variables, visualizing it inevitably becomes problematic. Hence the need to reduce the number of parameters, either by fixing some of them, or integrating over them. This introduces a tool that will prove very useful in the applications, namely, the various partial energy densities, that is, the function obtained by integrating the squared modulus of the CWT over a subset of the parameters. In other words, various types of wavelet spectra, the analogs of the familiar power spectrum of a signal. As is well known in 1-D, the CWT is highly redundant, as one can expect from a transform that doubles the number of variables: one to two in 1-D, two to four in 2-D. This fact may be exploited in two ways. Either one limits oneself to a small subset of the transform, where most of the energy is concentrated, and thus one is led to the notions of local maxima, ridges and skeleton; or one discretizes the CWT and obtains wavelet frames. Such a representation is still redundant, but much less than the full CWT, and in many instances is a good substitute for a genuine orthonormal basis. An alternative is the so-called dyadic WT, originally due to Mallat, in which only the scale variable is discretized. Together with the latter, we also describe briefly the standard DWT, based on the multiresolution idea, and several generalizations, mostly the so-called lifting scheme. We conclude the chapter with a thorough discussion of a different scheme, called directional dyadic wavelet frames. Here, as in 1-D, there are two conflicting requirements: redundancy of the transform, which brings stability, and computing economy, that seeks fast algorithms. The formalism described here offers a good compromise.

xiii

Prologue

When it comes to treating a precise problem, the first question to ask is, which wavelet should one use? Thus there is a need for a sizable collection of them, well documented and calibrated. The aim of Chapter 3 is to provide this. The crucial distinction here is whether directions in the image are relevant or not. If they are not, a pointwise analysis suffices, and one can use rotation invariant (isotropic, radial) wavelets, the best known being the Mexican hat or LOG wavelet (already introduced by Marr [Mar82]). On the contrary, if directions must be detected, one needs a wavelet with a good orientation selectivity. The most efficient result is obtained with the so-called directional wavelets. These are filters living in a convex cone, with apex at the origin, in Fourier space. Examples are the 2-D Morlet wavelet and the family of conical wavelets. All these wavelets, and some more, are discussed in detail in Chapter 3, and their performances determined quantitatively. At this stage, the tool is ready and we turn to applications. Many of them are not easy to find, because they have appeared only in conference proceedings or in (unpublished) Ph.D. theses. For that reason, we have decided to present them in a rather detailed fashion, always giving original references, including personal websites when available. In each case, we emphasize the rationale for using wavelets in the particular problem at hand, rather than go into the technicalities. It is convenient (although not always unambiguous) to distinguish between two different fields of applications, image processing and physics. To the first type, the subject matter of Chapter 4, belong contour detection and character recognition; automatic target detection and recognition (for instance, in infrared radar imagery); image retrieval from data banks; medical imaging; detection of symmetries in patterns, in particular quasicrystals and other quasiperiodic patterns; and image denoising. The chapter concludes with two nonlinear extensions of the CWT, which both have important applications. The first one is contrast enhancement in images through an adaptive normalization. This technique, based on analogy with our visual system, may be of interest in medical imaging. Indeed typical images, such as those obtained by radiography or by NMR imaging, have rather weak contrast, which makes their interpretation sometimes difficult. The other problem we deal with is watermarking of images, which consists in adding an invisible “signature” (the watermark) to an image, that only the owner can recognize and is robust to manipulations. Clearly the field of image copyright offers a good market for such techniques. The novel method we present is based on the contrast analysis described previously, exploiting directional wavelets, and it turns out to be particularly efficient. The second class of applications, described in Chapter 5, concerns various fields of physics. Characteristically, they all belong to classical physics, as opposed to quantum physics, because the former relies much more on images. Indeed, there are very few applications of wavelet analysis in quantum problems. The first domain on which 2-D wavelets have made a substantial impact is astronomy and astrophysics, for several reasons. The Universe has a marked hierarchical structure.

xiv

Prologue

Nearby stars, galaxies, quasars, galaxy clusters and superclusters have very different sizes and live at very different distances. Thus the scale variable is essential and a multiscale analysis is in order. This, of course, suggests wavelet analysis, and indeed many authors have used it in problems such as determination of the large-scale structure of the Universe, galaxy or void counting, or analysis of the cosmic background radiation. In addition, we describe more in depth two applications of our own, namely, the detection of various magnetic features of the Sun, from satellite images, and the detection of distant gamma-ray sources in the Universe. In the latter case, difficult statistics problems arise, because of the extreme weakness of the signal (such a source emits very few high energy photons). The next topic is Earth physics: fault detection in geophysics, seismology, climatology (notably, thunderstorm prevision). A number of successful applications pertain to fluid dynamics, from the detection of coherent strucures in fully developed 2-D turbulence (a domain pioneered forcefully by Marie Farge [164]) to the measurement of the velocity field in a turbulent fluid, or the disentangling of a 2-D (or 3-D) wave train. Next comes the world of fractals. These are structures that are solely characterized by their behavior under a scaling transformation: ideal ground for wavelets! However, the self-reproducing properties of physical fractals are in general only approximate, so that methods from statistical mechanics are needed. Thus, a thermodynamical formalism has been designed by Arn´eodo and his group in Bordeaux for treating such problems, and we give a brief account of it. Finally we touch upon the problem of shape recognition, where wavelet descriptors have proven useful too. At this point, the book undergoes a sort of phase transition. Up to here, everything was done by hand, so to speak. The properties of the CWT have been derived by explicit calculations and very few mathematical prerequisites have been asked for. But now it is time to look over the hill and notice that the whole theory is firmly grounded in group theory. Indeed the wavelet transform and all its properties may be entirely derived from an appropriate representation of the affine group, both in one and in two dimensions. A mathematical condition, called square integrability of the representation, ensures the validity of the derivation, in particular the possibility of inverting the wavelet transform, that is, of obtaining reconstruction formulas. We devote two rather short chapters, 6 and 7 to these developments, with a double benefit. First, on the pedagogical level, we want to convince the reader that the group-theoretical approach is not only mathematically correct and pleasant, it is also natural and easy. It allows us indeed to understand in a simple and unified language the deeper mathematical structures involved. It is also quite efficient, in that it yields a general formalism (in fact, a special case of the coherent state formalism, well known in quantum physics, in particular, in quantum optics) that permits us to extend the CWT to more general manifolds, such as R3 , the two-sphere, or space–time, all generalizations that will be discussed in later chapters. Of course, we do not expect our reader to be fully conversant with group theory, and we will define all the needed ingredients along the way. Actually we will essentially restrict our treatment

xv

Prologue

to 2 × 2 or 3 × 3 matrices, without resort to abstract notions. Nevertheless, we found it convenient to gather all the group-theoretical information in a separate appendix. We begin, in Chapter 6, by revisiting the 1-D CWT in the light of the so-called ax + b or restricted affine group of the line, that is, the set of all translations and positive dilations. It turns out that the CWT may also be interpreted as a phase space representation of signals, in the sense of Hamiltonian mechanics, and the group-theoretical language makes this evident. The same treatment is then applied to the Gabor transform, also called Short Time or Windowed Fourier transform, simply replacing the affine group by the Weyl–Heisenberg group, that is, the group of phase space translations (this point of view has also been emphasized by Daubechies [Dau92]). Next, in Chapter 7, we repeat the procedure in two dimensions. Here the relevant group is the similitude group SIM(2), which consists of translations, rotations and dilations of the plane, that is, precisely all the transformations we have chosen to apply to images. Here, as in the 1-D case, the basic tool is a representation of the group by unitary operators acting in the space of finite energy signals, a natural representation that possesses the property of square integrability, meaning roughly that its matrix elements are square integrable functions of the group parameters. Here too, the CWT is a phase space realization of signals, and we spend some time exploring the consequences of this fact. In a third chapter with a mathematical flavor, Chapter 8, we discuss two less known properties of wavelets. First, some of them have minimal uncertainty, in the sense that they saturate some uncertainty relations linked to the Lie algebra of the wavelet group, exactly as Gaussians saturate those associated to the canonical commutation relations. Then we explore the relationship between wavelet transforms and the Wigner transform, well–known in physics and in radar theory (under the name of the closely related ambiguity function). The next two chapters are devoted to various extensions of the standard CWT, that can be derived with help of the general formalism just developed. First we treat, in Chapter 9, the higher dimensional cases. We begin with the 3-D CWT, which is a straightforward extension of the 2-D case. Then we examine in depth the CWT over the 2-sphere. Here, of course, there is a strong motivation from several domains, from geophysics to astrophysics. The former is clear. As for the latter, when one considers the whole Universe, as in the problem of gamma source detection mentioned above, it is necessary to take the curvature into account. However, there is an equally appealing aspect in the mathematics of the subject. Indeed, the group to consider here is the conformal group of the sphere S 2 , which is nothing but the proper Lorentz group SOo (3, 1). The same group is also the conformal group of the plane R2 , for instance, the tangent plane at the North Pole. The sphere and its tangent plane are mapped onto each other by the stereographic projection from the South Pole and its inverse. This operation is in fact the key to the construction of a spherical CWT. Indeed, the operations to be performed on spherical signals are motions on the sphere, given by rotations, and local dilations around a given point. In order to define

xvi

Prologue

these, one first defines dilations around the North Pole by lifting the corresponding ones in the tangent plane by inverse stereographic projection. Then, dilations around any other point of the sphere are obtained by combining the previous ones with an appropriate rotation. As a consequence, the parameter space of the spherical CWT is not the Lorentz group itself, but a homogeneous space of it, containing only rotations and the dilations just defined, that is, the quotient of SOo (3, 1) by a certain subgroup. Therefore, one needs the general formalism described in Chapter 7 in order to get a genuine spherical CWT. As an additional benefit, one recovers the natural link between the sphere and its tangent plane: the spherical CWT tends to the usual plane CWT when the radius of the sphere increases to infinity (the so-called Euclidean limit). It is gratifying that this aspect too is entirely described by the group-theoretical machinery, in terms of an operation called group contraction. Another byproduct of our spherical CWT is the possibility of designing good wavelet approximations of integrable functions on the sphere, another result previously known in the plane case. Here again practical applications are at hand, in the context of the so-called Geomathematics advertised by Freeden and his school [Fre97]. Then we turn, in Chapter 10, to the extension of the CWT to space–time. The problem of interest here is, of course, motion estimation, more precisely, detection, tracking, and identification of objects in (relative) motion. Examples include traffic monitoring, autonomous vehicle navigation, and tracking of ballistic missile warheads. This is a difficult problem, since the data is huge and often very noisy. As a consequence, most algorithms tend to lose track of the targets after a while, particularly if the latter changes its appearance (e.g., a maneuvering aeroplane) or in the case of an occlusion (one moving object hides another one). From the wavelet point of view, one designs a spatio-temporal CWT, whose parameters are space and time translations, rotations, global space–time dilations, that catch the size of the target, and a speed tuning parameter that measures its speed. The usual formalism goes through almost verbatim and allows one to design an efficient algorithm for motion estimation. One key ingredient again is the successive use of several partial energy densities. In the final Chapter 11, we turn to another kind of generalizations, namely, transforms specially adapted to the detection and modeling of lines and curves, called the ridgelet and the curvelet transforms. The motivation for these new transforms, and their superiority over standard wavelets, is that they take much better into account the geometry of the object to be analyzed. A curve in the plane is more 1-D than 2-D, and the conventional 2-D CWT simply ignores this fact – hence it is unnecessarily costly. Here, of course, one experiences the much bigger richness of the 2-D world, in particular, concerning singularities of functions. These transforms naturally lead to new approaches to image compression and various nonlinear approximations, that we also describe. We conclude the chapter and the book with a topic called ‘algebraic wavelets’. These are wavelets adapted to self-similar tilings on the line or the plane obtained by

xvii

Prologue

replacing the usual natural numbers by a different system of numeration, for instance, √ the golden mean τ = 12 (1 + 5). This is actually a generalization of the discrete WT, but it provides another example of wavelets adapted to a specific geometry, hence it is not out of place in this volume, and we found it interesting to give a short account of it, both in 1-D and in 2-D. In the latter case, typical examples are the famous Penrose tilings of the plane, with pentagonal symmetry, and this brings us back to the study of aperiodic patterns and to quasicrystals! The conclusion of the whole story is definitely optimistic. Wavelets, and in particular the continuous WT, have proven to be a versatile and extremely efficient tool for image processing, provided one uses the right wavelet on the right problem. Their future is undoubtly bright, in many fields of science and technology. Before concluding this introduction, several technical remarks are in order. First, most examples that are not reproduced from original papers have been computed using our own wavelet toolbox, called the YAW (Yet Another Wavelet) Toolbox, and freely accessible on the Louvain-la-Neuve website . Next, we have found it useful to split the references into two sections, devoted to books and Ph.D. theses, and regular journal articles (with a different presentation, viz. [Ald96] and [2], respectively). As we have already said, theses are an extremely rich source of information, although they are often only accessible on the web. In general, we have tried to trace most of the results to the original papers. Of course, there are omissions and misrepresentations, due to our ignorance and prejudices. We take responsibility for this and apologize in advance to those authors whose work we might have mistreated.

Acknowledgements The present volume results from some fifteen years of continuing research interest in wavelets in Louvain-la-Neuve, starting with a collaboration between Alex Grossmann (Marseille) and J.-P. A. Throughout these years, all four authors have lectured on wavelets in places as diverse as Louvain-la-Neuve, Paris, Atlanta, Zakopane, Havana, Cotonou, Amsterdam, Lyon, Brussels, many papers have been written, and the book reflects all the experience thus acquired. Many students have been involved and they all deserve thanks for their contribution to the edifice. But most of all, we have to express our gratitude to three people, Bernard Piette, Alain Coron, and Laurent Jacques, for the project could never have been completed without their computer skills and tireless help. As expected, the elaboration process involved many reciprocal visits, and we have to thank our respective institutions (UCL, CTSPS at Clark Atlanta U., EPFL, Concordia U.) for their hospitality and support, as well as the various funding agencies that made these travels possible. Finally, we thank many colleagues for stimulating

xviii

Prologue

discussions, such as Roberto Cesar, Alain Arn´eodo, Matthias Holschneider, Fran¸coise Bastin, Fabio Bagarello, Emmanuel Van Vyve, to name a few. Special thanks are due to Bruno Torr´esani, whose friendly comments and criticisms have helped us considerably, and who in addition proofread a large part of the manuscript. Jean-Pierre Antoine (Louvain-la-Neuve) Romain Murenzi (Atlanta and Kigali) Pierre Vandergheynst (Lausanne) Syed Twareque Ali (Montr´eal)

1

Warm-up: the 1-D continuous wavelet transform

1.1

What is wavelet analysis? Wavelet analysis is a particular time- or space-scale representation of signals that has found a wide range of applications in physics, signal processing and applied mathematics in the last few years. In order to get a feeling for it and to understand its success, we consider first the case of one-dimensional signals. Actually the discussion in this introductory chapter is mostly qualitative. All the mathematically relevant properties will be described precisely and proved systematically in the next chapter for the two-dimensional case, which is the proper subject of this book. It is a fact that most real life signals are nonstationary (that is, their statistical properties change with time) and they usually cover a wide range of frequencies. Many signals contain transient components, whose appearance and disappearance are physically very significant. Also, characteristic frequencies may drift in time (e.g., in geophysical time series – one calls them pseudo-frequencies). In addition, there is often a direct correlation between the characteristic frequency of a given segment of the signal and the time duration of that segment. Low frequency pieces tend to last for a long interval, whereas high frequencies occur in general for a short moment only. Human speech signals are typical in this respect: vowels have a relatively low mean frequency and last quite a long time, whereas consonants contain a wide spectrum, up to very high frequencies, especially in the attack, but they are very short. Clearly standard Fourier analysis is inadequate for treating such signals. Strictly speaking, it applies only to stationary signals, and it loses all information about the time localization of a given frequency component. In addition, it is very uneconomical. When the signal is almost flat, and thus uninteresting, one still has to sum an infinite alternating series to reproduce it. Worse yet, Fourier analysis is highly unstable with respect to perturbation, because of its global character. For instance, if one adds an extra term, with a very small amplitude, to a linear superposition of sine waves, the signal will barely be modified, but the Fourier spectrum will be completely perturbed. This does not happen if the signal is represented in terms of localized components. Indeed, as we shall see shortly, the basic idea of the wavelet transform is to decompose a signal locally

1

2

Warm-up: the 1-D continuous wavelet transform

into contributions living at different scales. This is a marked contrast with the Fourier components, which are sinusoidal waves repeating themselves indefinitely. As such, it is difficult to give them any physical reality. If a piece of audio signal is identically zero, it is because no sound is emitted, not because the Fourier components necessary to represent the zero signal interfere destructively. These components are a mathematical construction, rather than a genuine physical phenomenon. To quote J. Ville [364]: Si nous consid´erons en effet un morceau de musique . . . et qu’une note, la par exemple, figure une fois dans le morceau, l’analyse harmonique [de Fourier] nous pr´esentera la fr´equence correspondante avec une certaine amplitude et une certaine phase, sans localiser le la dans le temps. Or, il est e´ vident qu’au cours du morceau il est des instants o`u l’on n’entend pas le la. La repr´esentation est n´eanmoins math´ematiquement correcte, parce que les phases des notes voisines du la sont agenc´ees de mani`ere a` d´etruire cette note par interf´erence lorsqu’on ne l’entend pas et a` la renforcer, e´ galement par interf´erence, lorsqu’on l’entend; mais s’il y a dans cette conception une habilet´e qui honore l’analyse math´ematique, il ne faut pas se dissimuler qu’il y a e´ galement une d´efiguration de la r´ealit´e: en effet, quand on n’entend pas le la, la raison v´eritable est que le la n’est pas e´ mis.

That is, If we consider a piece of music . . . and if a note, an A for instance, appears once in that piece, Fourier analysis will yield the corresponding frequency with a certain amplitude and a certain phase, without localizing the A in time. Clearly the A will not be heard at certain instants. Yet the representation is mathematically correct, because the phases of the neighboring notes conspire to suppress the A by interference when it is not heard and to enhance it, again by interference, when it is heard. However, although this conception shows a skillfulness that honors mathematical analysis, one should not hide the fact that it also distorts reality: indeed, when the A is not heard, the true reason is that the A is not emitted.

Another eloquent comment along the same line by L. de Broglie may be found, together with the one above, in [Fla93; p.9]. Facing these problems, signal analysts turn to time–frequency representations. The idea is that one needs two parameters: one, called a, characterizes the frequency, the other one, b, indicates the position in the signal. This concept of a time–frequency representation is in fact quite old and familiar. The most obvious example is simply a musical score (see Figure 1.1). Clearly, it is not sufficient to give the pitch of a given note, that is, the frequency to which it corresponds, it is also important to know when to play it (time information)! Let s(x) be a finite energy signal, that is, a square integrable function s ∈ L 2 (R, d x). In most cases, x will be a time variable and the (Fourier) conjugate quantity a frequency, 2 4 Fig. 1.1. A traditional time–frequency representation of a signal (from Mozart’s Don Giovanni,

Act 1).

3

1.1 What is wavelet analysis?

but in general x simply represents position in the signal. Thus, following [Dau92], we prefer to keep a neutral notation (x, ξ ) for the couple of conjugate variables, instead of the more familiar (t, ω). Accordingly, the Fourier transform of the signal s is defined by ∞ 1 s(ξ ) = √ d x e−iξ x s(x). (1.1) 2π −∞ If one requires the transform to be linear, a general time–frequency transform of the signal s will take the form: ∞ d x ψb,a (x) s(x) , (1.2) s(x) → S(b, a) = −∞

where ψb,a is the analyzing function. Within this class, two time–frequency transforms stand out as particularly simple and efficient: the windowed (or short time) Fourier transform (WFT) and the wavelet transform (WT). For both of them, the analyzing function ψb,a is obtained by acting on a basic (or mother) function ψ, in particular, b is simply a time translation. The essential difference between the two is in the way the frequency parameter a is introduced: (1) Windowed Fourier transform: ψb,a (x) = ei(x−b)/a ψ(x − b).

(1.3)

Here ψ is a window function and the a-dependence is a modulation (1/a ∼ frequency); the window has constant width, but the smaller a, the larger the number of oscillations in the window (see Figure 1.2 (left)). (2) Wavelet transform: x −b 1 ψb,a (x) = √ ψ . (1.4) a a The action of a on the function ψ (which must be oscillating, see below) is a dilation (a > 1) or a contraction (a < 1): the shape of the function is unchanged, it is simply spread out or squeezed (see Figure 1.2 (right)). In particular, the effective support of ψb,a varies as a function of a. The windowed Fourier transform was originally introduced by Gabor (actually in a discretized version), with the window function ψ taken as a Gaussian; for this reason, it is sometimes called the Gabor transform. With this choice, the function ψb,a is simply a canonical (harmonic oscillator) coherent state [Kla85], as one sees immediately by writing 1/a = p. Since the new variables are the time (position) b and the frequency 1/a, the Gabor transform yields a genuine time–frequency representation of the signal. As for the wavelet transform, the variables are b and the scale a (or pitch in the case of music), hence we shall speak rather of a time-scale representation. We may remark here that the resemblance between the windowed Fourier transform and the wavelet transform is not accidental. They are both particular instances of a large

4

Warm-up: the 1-D continuous wavelet transform

1/a ~ ~ frequency ψb,a(x)

high

medium

low

a1

x Fig. 1.2. The function ψb,a (x) for different values of the scale parameter a, in the case of the windowed Fourier transform (left) and the wavelet transform (right). The quantity 1/a, which corresponds to a frequency, increases from bottom to top.

class of integral transforms constructed by the formalism of coherent states [Ali00]. This general analysis, however, has a more mathematical flavor and is not needed in a first approach, although it clarifies and unifies the picture considerably. Therefore, we postpone it to Chapter 6, since we want to emphasize first the practical aspects of the wavelet transform. One should note that the assumption of linearity is nontrivial, for there exists a whole class of quadratic or, more properly, sesquilinear time–frequency representations. The prototype is the so-called Wigner–Ville transform, introduced originally by E.P. Wigner [373] in quantum mechanics (in 1932!) and extended by J. Ville [364] to signal analysis: +∞ x x Ws (b, ξ ) = d x e−iξ x s(b − ) s(b + ), ξ = 1/a. (1.5) 2 2 −∞ Note that the signal s(x) is usually a real function, but, in quantum mechanics, s(x) represents a wave function, and is thus in general complex. This transform is entirely intrinsic to the signal, since it does not contain any extra function (wavelet, window)

5

1.2 The continuous wavelet transform

that inevitably influences the result. On the other hand, it is quadratic, which implies the appearance of interference terms whenever the signal is a superposition of two components. In order to minimize these as much as possible, one usually smoothes the Wigner–Ville transform with some function , thus obtaining a whole class of quadratic transforms, called Cohen’s class [109,Fla93], of the general form: Cs (b, ξ ) = db dξ (b − b , ξ − ξ ) Ws (b , ξ ). (1.6) R2

An example is the so-called smoothened pseudo-Wigner–Ville distribution, +∞ +∞

SPWs (b, ξ ) = db g(b − b ) d x h(x) e−iξ x s(b − x/2) s(b + x/2), (1.7) −∞

−∞

corresponding to a factorizable kernel (b, ξ ) = (2π )−1/2 g(b) h(ξ ), where h denotes the Fourier transform of h. Further information about quadratic transforms may be found in [Fla93], and as a general survey for time–frequency methods, we refer to [Gro01].

1.2

The continuous wavelet transform Actually one should distinguish two different versions of the wavelet transform, the continuous WT (CWT) and the discrete (or more properly, discrete time) WT (DWT) [Dau92,Hol95]. The CWT plays the same rˆole as the Fourier transform and is mostly used for analysis and feature detection in signals, whereas the DWT is the analog of the Discrete Fourier Transform (see for instance [Bur98] or [326]) and is more appropriate for data compression and signal reconstruction. The situation may be caricatured by saying that the CWT is more natural to the physicist, while the DWT is more congenial to the signal analyst and the numericist. The continuous wavelet transform is the main topic of this book. Nevertheless, for the sake of comparison, we will give short overviews of the discrete WT, both in one and two dimensions. The two versions of the WT are based on the same transformation formula, which reads, from (1.2) and (1.4): ∞ x −b −1/2 S(b, a) = |a| s(x), (1.8) dx ψ a −∞ where a = 0 is a scale parameter and b ∈ R a translation parameter (one often imposes only a > 0, which is more natural, but makes formulas slightly more complicated; see Chapter 6). Equivalently, in terms of Fourier transforms: ∞ 1/2 ) S(b, a) = |a| s(ξ ) eiξ b . dξ ψ(aξ (1.9) −∞

6

Warm-up: the 1-D continuous wavelet transform

In these relations, s is a square integrable function, representing a finite energy signal, and the function ψ, the analyzing wavelet, is assumed to be well localized both in the space (or time) domain and in the frequency domain. In addition ψ must satisfy the following admissibility condition, which guarantees the invertibility of the WT: ∞ )|2 |ψ(ξ cψ ≡ 2π dξ < ∞. (1.10) |ξ | −∞ In most cases, this condition may be reduced to the (only slightly weaker) requirement that ψ has zero mean: ∞ = 0 ⇐⇒ d x ψ(x) = 0. (1.11) ψ(0) −∞

Intuitively, it expresses the fact that a wavelet must be an oscillating function, real or complex (“little wave”). This is often thought to be the origin of the term “wavelet”, but it is not the case historically. Indeed the word was widely in use in the geophysics community, with quite a different meaning, when it was introduced by Grossmann and Morlet [205,206] in the present sense, under the name “wavelets of constant shape” – but, of course, this lengthy nomenclature did not survive the very first founding paper! ) is real and The wavelet ψ is said to be progressive if its Fourier transform ψ(ξ vanishes identically for ξ 0. (In the signal processing community, a signal with this property is called analytic, following the terminology introduced by J. Ville [364].) In addition, ψ is often required to have a certain number of vanishing moments: ∞ d x x n ψ(x) = 0, n = 0, 1, . . . N . (1.12) −∞

This property improves the efficiency of ψ at detecting singularities in the signal, since it is then blind to polynomials up to order N , which constitute the smoothest part of the signal. Notice that, instead of (1.8), which defines the WT as the scalar product of the signal s with the transformed wavelet ψb,a , S(b, a) may also be seen as the convolution of s with the scaled, flipped and conjugated wavelet ψa# (x) = |a|−1/2 ψ(−x/a) : ∞ S(b, a) = (ψa# ∗ s)(b) = d x ψa# (b − x) s(x). (1.13) −∞

In other words, the CWT acts as a filter with a function of zero mean. This property is crucial, for the main virtues of the CWT follow from it, combined are as well with the support properties of ψ. Indeed, we must assume that ψ and ψ localized as possible, but respecting, of course, the Fourier uncertainty principle. This means that, up to minute corrections, the product of the lengths of the supports of ψ is bounded from below by a fixed constant, usually taken as 1/2. Equivalently, and ψ 2 is bounded from below. the product of the variances of the distributions |ψ|2 and |ψ| More precisely, one defines the centers of gravity (which may in fact be normalized to

7

1.2 The continuous wavelet transform

zero by a suitable redefinition of the coordinates): ∞ ∞ )|2 , x0 = d x x |ψ(x)|2 , ξ0 = dξ ξ |ψ(ξ −∞

(1.14)

−∞

and the corresponding variances ∞ ( x)2 = ψ−2 d x (x − x0 )2 |ψ(x)|2 ; −∞ ∞ 2 −2 )|2 . dξ (ξ − ξ0 )2 |ψ(ξ ( ξ ) = ψ

(1.15) (1.16)

−∞

Then the Fourier uncertainty theorem [Fla93] says that

x ξ

1 . 2

(1.17)

Under these assumptions, the transformed wavelets ψb,a and ψ b,a are also well localized. Therefore, the WT s → S performs a local filtering, both in time (b) and in scale (a). The transform S(b, a) is nonnegligible only when the wavelet ψb,a matches the signal, that is, the WT selects the part of the signal, if any, that lives around the time b and the scale a. has a numerical support (bandwidth) of width ξ , then ψ In addition, if ψ b,a has a numerical support of width ξ/|a|. Thus, remembering that 1/a behaves like a frequency, we conclude that the WT works at constant relative bandwidth, that is,

ξ/ξ = constant. This implies that it is very efficient at high frequency, i.e., small scales, in particular for the detection of singularities in the signal. By comparison, in the case of the Gabor transform, the support of ψ b,a keeps the same width ξ for all a, that is, the WFT works at constant bandwidth, ξ = constant. This difference in behavior is often the key factor in deciding whether one should choose the WFT or the WT in a given physical problem. Another crucial fact is that the transformation s(x) → S(b, a) may be inverted exactly, which yields a reconstruction formula (this is only the simplest one, others are possible, for instance using different wavelets for the decomposition and the reconstruction): ∞ ∞ da −1 s(x) = cψ db ψb,a (x) S(b, a), (1.18) 2 −∞ −∞ a where the normalization constant cψ is given in (1.10) (incidentally, this relation shows why the admissibility condition cψ < ∞ is required for the transformation to be invertible). This means that the WT provides a decomposition of the signal as a linear superposition of the wavelets ψb,a with coefficients S(b, a). Notice that the natural measure on the parameter space (a, b) is da db/a 2 , and it is invariant not only under time translation, but also under dilation. This fact is important, for it suggests that these geometric transformations play an essential rˆole in the CWT.

8

Warm-up: the 1-D continuous wavelet transform

One should emphasize here that the choice of the normalization factor |a|−1/2 in (1.4) or (1.8) is not essential. This choice makes the transform unitary: ψb,a 2 = ψ2 and also S2 = s2 , where · 2 denotes the L 2 norm in the appropriate variables (the squared norm is interpreted as the total energy of the signal). In practice, one often uses instead a factor a −1 , which has the advantage of giving more weight to the small scales, i.e., the high frequency part (which contains the singularities of the signal, if any). Thus, defining 1 x −b ψ(b,a) = ψ , (1.19) |a| a we obtain the so-called L 1 -normalized transform: ∞ x −b ˘ a) = ψ(b,a) |s ≡ |a|−1 S(b, s(x), dx ψ a −∞

(1.20)

which preserves the L 1 -norm of the signal, as follows immediately from the corresponding convolution formula ˘ a) = (ψa# ∗ s)(b), S(b,

(1.21)

˘ 1 = s1 , where where ψa# (x) = |a|−1 ψ(−x/a). Thus indeed ψa# 1 = ψ1 and S 1 · 1 denotes the L -norm in the corresponding variables.

1.2.1

Examples In order to fix ideas, we exhibit here two simple examples of wavelets, both in the time domain and in the frequency domain. (1) The Mexican hat wavelet This wavelet is simply the second derivative of a Gaussian: H (ξ ) = ξ 2 exp(− 1 ξ 2 ). ψH (x) = (1 − x 2 ) exp(− 12 x 2 ), ψ 2

(1.22)

(2) The Morlet wavelet This wavelet is essentially a plane wave within a Gaussian window: M (ξ ) = exp(− 1 (ξ − ξo )2 ) + c(ξ ). ψM (x) = exp(iko x) exp(− 12 x 2 ) + c(x), ψ 2 (1.23) Here the correction term c must be added in order to satisfy the admissibility condition (1.11), but in practice one will arrange that this term be numerically negligible ( 10−4 ) and thus can be omitted (it suffices to choose the basic frequency |ξo | large enough, typically |ξo | > 5.5). These two wavelets have very different properties and, naturally, they will be used in quite different situations. Typically, the Mexican hat is sensitive to singularities in the signal, and it yields a genuine time-scale analysis. On the other hand, since it is complex, the Morlet wavelet will catch the phase of the signal, hence will be sensitive to frequencies, and will lead to a time-frequency analysis, somewhat closer to a Gabor

1.2 The continuous wavelet transform

1

1.5

Log Scale

9

2

2.5

3

100

200

300

400

500

600

700

800

900

1000

Position Fig. 1.3. Wavelet analysis with a Mexican hat wavelet of the discontinuous signal bumps (shown in

the bottom panel).

analysis. In both cases, additional flexibility is obtained by adding a width parameter to the Gaussian (see (3.8) in the equivalent 2-D situation). As an illustration of the performance of the CWT as a singularity scanner, we first show in Figure 1.3 the analysis with a Mexican hat wavelet of a discontinuous signal, called bumps and consisting of three pieces, a δ function, a boxcar function and a tent function. Clearly the wavelet locates all discontinuities in the signal and in its successive derivatives well. However, if one wants to discriminate between the various types of singularities, one has to invoke the concept of vanishing moment, defined in (1.12). Let us consider the successive derivatives of a Gaussian: ψH(n) (x) = −

dn exp(− 12 x 2 ). dxn

(1.24)

For increasing n, these wavelets have more and more vanishing moments, and are thus sensitive to increasingly sharper details. As an example, we consider a continuous signal obtained by glueing together an arc of parabola (the so-called function x+2 ) and a linear piece and we analyze it successively with the first three derivatives of a Gaussian, ψH(n) (x), n = 1, 2, 3. The result is shown in Figure 1.4. In (a), the first-order wavelet ψH(1) has only one vanishing moment, hence it sees the full content of the two pieces of the signal. In (b), the second-order wavelet ψH(2) (x) ≡ ψH does not see the linear part anymore, only the singularities at the two ends, but still sees the quadratic piece on the left (in technical terms, one would say that this wavelet is blind to a linear trend in the signal). In (c), finally, the third-order wavelet correctly erases both pieces of the

10

Warm-up: the 1-D continuous wavelet transform

3.6

Log Scale

3.8 4 4.2 4.4 4.6 4.8

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

3.6

3.6

3.8

3.8

4

4

Log Scale

Log Scale

(a)

4.2

4.2

4.4

4.4

4.6

4.6

4.8

4.8

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

(b)

200

400

600

800

1000

1200

1400

1600

1800

2000

Position

(c)

Fig. 1.4. Analysis of a composite signal (bottom panel) with successive derivatives of a Gaussian.

(a) First order; (b) second order; (c) third order.

signal, keeping only the three singularities. This example shows the advantage of the local filtering effect of the CWT. Notice that a Gabor analysis would be utterly unable to achieve such a discrimination between singularities, let alone to detect them! As a direct application of this behavior, an interesting technique has been designed by A.Arn´eodo et al. [49], which consists in analyzing the same signal with several wavelets ψH(n) , for different n. The features common to all the transforms surely belong to the signal, they are not artifacts of the analysis.

1.3

Discretization of the CWT, frames All this concerns the continuous WT (CWT). But, in practice, for numerical purposes, the transform must be discretized, by restricting the parameters a and b in (1.8) to the points of a discrete lattice = {a j , bk , j, k ∈ Z} in the (a, b)-(half)-plane. Then we

11

1.3 Discretization of the CWT, frames

say that yields a good discretization if an arbitrary signal s(x) may be represented as a discrete superposition jk (x), s(x) = ψ jk |s ψ (1.25) j,k ∈ Z

jk should be instead of the reconstruction formula (1.18). In (1.25), ψ jk ≡ ψbk ,a j and ψ explicitly constructible from ψ jk . We emphasize that (1.25) must be an exact representation, i.e., there is no loss of information as compared to a direct discretization of the continuous reconstruction (1.18). Notice that here also, as in the latter, the reconstruction formula is in general not unique, which offers an additional degree of freedom in a given situation. One may wonder whether a discrete representation of the type (1.25) is really possible. The answer lies in the reproducing property, da db −1

S(b , a ) = cψ ψb ,a |ψb,a S(b, a) , (1.26) a2 that every wavelet transform must satisfy. Indeed (1.26) implies that the information content of the wavelet transform S(b, a) is highly redundant. In fact the signal has been unfolded from one to two dimensions, and this explains the practical efficiency of the CWT for disentangling parts of the signal that live at the same time, but on different scales. This redundancy (which is the source of the nonuniqueness of the reconstruction formula) may be eliminated – this is the rationale behind the discrete wavelet transform. It may also be exploited, for instance, by observing that it must be possible to obtain the full information about the signal from a small subset of the values of the transform S(b, a). In particular, the validity of a representation (1.25) means that a discrete subset will do the job, and this is precisely what is needed for the reconstruction of a signal from its wavelet transform. The problem is to find the minimal sampling grid ensuring no loss of information. In order to formulate it in mathematical terms, one relies on the theory of discrete frames or nonorthogonal expansions, that we now sketch. See [121,Dau92] for a complete treatment. In fact, the discrete representation (1.25) means that the signal s(x) may be replaced by the set {ψ jk |s} of its wavelet coefficients. Since s ∈ L 2 , it is natural to require that the sequence of coefficients be also square integrable and that the map F : s → {ψ jk |s} be continuous from L 2 (R) to 2 , i.e., |ψ jk |s|2 Bs2 , 0 < B < ∞. (1.27) j,k ∈ Z

In addition, one wants the reconstruction of s(x) from its coefficients to be numerically stable, that is, a small error in the coefficients implies a small error in the reconstructed signal. In particular, if the left-hand side of (1.27) is small, s2 should be small also. Therefore, there must exist a constant A > 0 such that

12

Warm-up: the 1-D continuous wavelet transform

A s2

|ψ jk |s|2 B s2

(1.28)

j,k ∈ Z

(the lower bound indeed guarantees the numerical stability [Dau92]). By definition, this relation means precisely that the set {ψ jk } constitutes a (discrete) frame, with frame bounds A and B. This frame is said to be tight if A = B. Note that (1.28) is in fact a weakened form of the Parseval relation. The latter is recovered in the case of a tight frame, in particular for an orthonormal basis. For a general frame, however, (1.28) is sufficient for inverting the wavelet representation (we will discuss this in detail in the 2-D case, in Section 2.4). We will present a detailed analysis of these concepts in Chapter 2. Here we simply observe that, for all practical purposes, a good frame is almost as good as an orthonormal basis. By “good frame,” we mean that the expansion (1.25) converges sufficiently fast. The detailed analysis of [121,122] shows this to be the case if |B/A − 1| 1, thus in particular if the frame is tight. Many functions ψ satisfying the admissibility condition (1.10) will yield a good frame (of course, this must be proved for every given ψ). However, we will not get an orthonormal basis, since the functions {ψ jk , j, k ∈ Z} are in general not orthogonal to each other! Yet orthonormal bases of wavelets can be constructed, but by a totally different approach, based on the concept of multiresolution analysis. We emphasize that the discretized version of the CWT just described is totally different in spirit and method from the genuine discrete wavelet transform, that we will sketch in Section 1.5 below. The full story may be found in [Dau92], for instance. Of course the practical question is: how does one build a good frame? Clearly, the question of the existence of a discrete frame must take into account the geometry of the parameter space. In the present case, this means that the lattice must be invariant under discrete dilations and translations: r for scale, one chooses naturally a = a λ− j , j ∈ Z, for some λ > 1; j o −j r for time, one takes b ≡ b = k b a λ , j, k ∈ Z. k k, j o o Thus we get ψ jk (x) = λ j/2 ψ(ao−1 λ j x − kbo ),

j, k ∈ Z.

(1.29)

The most common choice is λ = 2 (octaves!) and ao = bo = 1, which results in ψ jk (x) = 2 j/2 ψ(2 j x − k),

j, k ∈ Z.

(1.30)

It is worth noticing that this so-called dyadic lattice {(k2− j , 2− j ), j, k ∈ Z} is exactly the same that indexes the DWT (see Section 1.5), which may create some confusion (and sometimes did so!). For a given choice of ψ, λ, one finds a range of values of bo such that {ψ jk }, as given in (1.29), is a frame. Detailed results may be found in [122,Dau92]. Here we will restrict ourselves to the following simplified version.

13

1.3 Discretization of the CWT, frames

Theorem 1.3.1 . Let ψ and ao be such that: ∞ − j ξ )|2 > 0; (i) inf |ψ(λ 1ξ λ

(ii)

j=−∞

)| C |ξ |α (1 + |ξ |)−γ , α > 0, γ > α + 1. |ψ(ξ

Then there exists boo such that {ψ jk } constitutes a frame for all choices bo < boo . Both the Mexican hat and the Morlet wavelet satisfy the conditions of the theorem for the dyadic case, λ = 2, ao = bo = 1, thus they both generate discrete frames on the dyadic lattice. Explicit values for the corresponding frame bounds A, B may be found in [Dau92]. However, the numerical implementation of such a dyadic frame is unwieldy, in particular, the reconstruction works well only if the frame is very redundant. Two variants are used in practice, and both of them amount to increase the redundancy. The first one consists in subdividing further the octaves, introducing what are called additional voices. This means that one further subdivides each octave, replacing in (2.77) the exponent j by Nν j, ν = 0, 1, . . . , N − 1 (ν is called a voice and N the number of voices). The effect is to “densify” the dyadic lattice, which improves the ratio B/A and thus speeds up convergence of the discrete approximation. By taking sufficiently many voices by octaves, one may even get frames which are numerically tight. For further details, we refer the reader to [Dau92]. However, if the speed is the determining criterion, one can do better by using the continuous wavelet packets described in Section 1.6.1. The other solution consists in replacing the dyadic lattice with a rectangular one, that is, taking the sampling rate independent of the scale. This has the advantage of making the whole analysis invariant under global discrete translations. The resulting version of the discretized CWT has been advocated by Mallat [Mal99,265] (who, curiously, calls it the dyadic WT!) and Torr´esani [Tor95], and we will meet it again repeatedly in the sequel (see Section 1.6.1). Note that similar ideas were already used in the classical Littlewood–Paley analysis [Fra91] and the famous Laplacian pyramid from vision analysis [91]. The scheme was also rediscovered by statisticians, who called it “stationary wavelet transform” [112,295]. To illustrate this technique, we show in Figure 1.5 a five-level decomposition of the signal bumps with a translation invariant frame of quadratic spline wavelets, as used in [265]. The figure shows, from bottom up, the low resolution approximation and the five levels of details with increasing resolution. As we shall see in detail in the next chapter, Section 2.5, the whole machinery of frames extends almost verbatim to the two-dimensional case. In particular, both the 2-D Mexican hat and the 2-D Morlet wavelet yield reasonably good 2-D frames. Besides their simplicity and their efficiency, this explains the widespread use of these two wavelets in image processing, as was already the case in one dimension.

14

Warm-up: the 1-D continuous wavelet transform

Signal 2 1 0 2 0 −2 1 0 −1 1 0 −1 2 0 −2 2 0 −2 10 5 0

100

200

300

400

500

600

700

800

900

1000

Fig. 1.5. Five-level decomposition of the signal bumps with a translation invariant frame of quadratic spline wavelets. The low resolution approximation is shown on the bottom panel and the five levels of details with increasing resolution a = 2− j , j = −5, −4, −3, −2, −1, in the next five panels.

1.4

Ridges and skeleton Real life signals are frequently very entangled and noisy, and their WT is difficult to interpret. However, a clever exploitation of the intrinsic redundancy of the CWT is often able to bypass the difficulty and thus to improve the efficiency and the range of applicability of wavelet analysis. The technique consists in using the skeleton of the CWT instead of its modulus. Roughly speaking, the skeleton is a collection of lines, called ridges, which are approximately lines of local maxima. The concept is easy to visualize in two extreme situations [Mal99,262,358]. Assume first that the signal s(t) consists of a singularity γα (x − xo ), of order α, at time xo , superimposed on a smooth background and some stochastic noise [262]. Here the singularity function γα is defined as follows: γα (x − xo ) =

0, x xo , α (x − xo ) , x > xo .

(1.31)

15

1.4 Ridges and skeleton

[The index α is in fact an index of homogeneity or a Lipschitz regularity exponent. For instance, a δ function has α = −1.] Thus we have d α+1 γα (x − xo ) = (α + 1) δ(x − xo ). d x α+1 Let the wavelet be the nth derivative of a smooth positive function φ, that is, ψ(x) = dn φ(x), with n α + 1 (typically, a derivative of a Gaussian, like the Mexican hat dxn and its higher order analogs (1.24)). Then the CWT of γα with respect to ψ to reads: n−α−1 φ xo − b α d . (1.32) Sγα (b, a) = (α + 1) a d x n−α−1 a Assume now that the modulus of the (n − α − 1)th derivative of φ has N maxima {φl , l = 1, . . . , N } at positions {xl , l = 1, . . . , N }. Then, for each a, the modulus |Sγα (b, a)| has N maxima localized at positions {bl = axl + xo , l = 1, . . . , N }, which converge toward xo as a → 0. Furthermore, the maxima of |Sγα (b, a)| lie on N lines, called (vertical) ridges {b = axl + xo , l = 1, . . . , N }, which converge toward the singularity xo of the signal, and the modulus of |Sγα (b, a)| along the lth ridge behaves as aα : |Sγα (b = axl + xo , a)| = (α + 1) a α φl .

(1.33)

Hence the strength α of the singularity may be read off a log–log plot: ln |Sγα (axl + xo , a)| ∼ α ln a + ln φl .

(1.34)

This technique, introduced by Mallat and Hwang [262], has been developed to a considerable extent for the analysis of fractals by Arn´eodo and his collaborators under the name of Wavelet Transform Modulus Maxima (WTMM) (see [Arn95] for a survey). The important point is that the restriction of the WT to its skeleton (the set of its ridges) characterizes the signal completely [264,265]. Thus, in practice, it is enough to compute the skeleton. We will discuss the 2-D extension of the WTMM method in Section 2.3.5 and its application to fractals in Section 5.4. To give a simple example, we take again our signal bumps, and compute the skeleton of its wavelet transform shown in Figure 1.3. The result, presented in Figure 1.6, clearly confirms the analysis above. As a second example, we consider the analysis of the behavior of a material under impact made in [358]. The physical context is that of a so-called “instrumented falling weight impact” testing. During such a test, a striker falls from a certain height on a clamped disk, so that either the striker rebounds or the disk breaks. In both cases, one records the time and the force on the striker. This type of event occurs on a very short time scale and is thus essentially transient, so that a time–frequency method is required for the analysis. Among the various methods discussed in [358], we focus here on the case of a rebound and a wavelet analysis of the force signal. The Mexican hat detects precisely three discontinuity points, namely, first contact, maximal penetration and last

Warm-up: the 1-D continuous wavelet transform

1

1.5

Log Scale

16

2

2.5

3

100

200

300

400

500

600

700

800

900

1000

Position Fig. 1.6. Skeleton of the CWT of the signal bumps, as presented in Figure 1.3.

contact. The results are shown in Figure 1.7. The signal is given in panel (a), whereas the CWT and its skeleton are presented in (b) and (c). The latter, in particular, shows the three ridges that point towards the three instants mentioned. A further analysis exploits the behavior of the modulus of the CWT along each ridge. This yields precious insight into the physics of the phenomenon, particularly in the case of the rupture of the sample, not shown here. We refer to [358] for more details. More generally, vertical ridges allow us to discriminate between genuine signal features and noise. First, noise ridges are usually much shorter, being visible mostly at small scales [21]. Then the modulus of the CWT tends to increase for increasing scale a on a noise ridge, whereas it decreases along genuine signal ridges. This fact has been exploited, for instance (in a discrete set-up), for the correction of aberrated images of the Hubble Space Telescope [85]: noise and signal have opposite behavior with increasing scale. As for the second typical situation, the idea is that many signals are well approximated by a superposition of simple spectral lines: s(x) =

N

sl (x),

sl (x) = Al (x) eiξl x ,

(1.35)

l=1

where the amplitude Al (x) varies slowly. By linearity, the WT of this signal is a sum

of terms, S(b, a) = l Sl (b, a), where, from (1.9),

1.4 Ridges and skeleton

250 (a)

Force [Volts]

200

150

100

50

500

600

700

800

900 1000 1100 1200 1300 1400 1500

Time [10−4 s]

(b) 1.86

Scale

3.59

6.90

13.29

25.58

49.25

600

700

800

900

1000 1100 1200 1300 1400 1500 −4

Time [10

s]

(c) 1.86

3.59

Scale

17

6.90

13.29

25.58

49.25

600

700

800

900 1000 1100 1200 1300 1400 1500 −4

Time [10

s]

Fig. 1.7. Analysis of the rebound signal, with a Mexican hat wavelet: (a) the signal and the points

detected by the respective ridges; (b) the modulus of the CWT; and (c) the corresponding skeleton (from [358]).

18

Warm-up: the 1-D continuous wavelet transform

Sl (b, a) =

√ a

+∞

−∞

√ = a eiξl b

l (ξ − ξl ) eiξ b )A dξ ψ(aξ

+∞

−∞

l (ξ ) eiξ b . dξ ψ(a(ξ + ξl )) A

(1.36) (1.37)

around aξl , one obtains the following expansion Inserting the Taylor expansion of ψ for Sl : l )sl (b) − iaeiξl b Sl (b, a) = ψ(aξ

d Al dψ (aξl ) (b) + higher order terms dξ db

(1.38)

Assuming the amplitude to be a smooth function (C 1 ), all terms beyond the first are easily bound using the intermediate value theorem and one can limit oneself to the lowest order, namely, l )sl (b), Sl (b, a) ψ(aξ

(1.39)

and thus S(b, a)

N

l )sl (b). ψ(aξ

(1.40)

l=1

) has a unique maximum in frequency space at ξ = ξo , like Assume that the wavelet ψ(ξ the Morlet wavelet (1.23). Then, if the values of the frequencies ξl are sufficiently far l ) allows to treat each spectral line independently. away from each other, the factor ψ(aξ In this case, the contribution of the lth spectral line to S(b, a) is localized on the scale al = ξo /ξl and, along the line of maxima a = al , called the lth (horizontal) ridge, the CWT is approximately proportional to the lth spectral line: S(b, al ) sl (b). o) ψ(ξ

(1.41)

The set of all the ridges is again called the skeleton of the CWT. Thus the restriction of the WT S(b, a) to its skeleton contains the whole information. The analysis extends to the more general case where the spectral lines have the form sl (x) = Al (x) eiφl (x) ,

(1.42)

with the amplitude Al (x) varying slowly with respect to the phase φl (x) (such signals are called asymptotic). Typical examples are spectra in NMR spectroscopy [210]. In this case, each term Sl in the CWT (1.8) in the time domain is a rapidly oscillating integral, the essential contribution to which is given by the stationary points of the phase of the integrand. These points are the solutions xs (b, a) of the equation dφl ξo (xs ) = . (1.43) dx a Then the corresponding ridge of the WT is defined as the set of points (b, a) for which xs (b, a) = b. These constitute a curve in the (b, a)-half-plane, which essentially

19

1.5 The discrete WT: orthonormal bases of wavelets

reduces to a line of local maxima. A detailed analysis [131] shows again that, on this curve, the WT S(b, a) coincides, up to a small correction, with the component sl (b) of the signal. Taking all ridges together, one obtains the skeleton of the CWT, and the analysis shows that the restriction of S(b, a) to it essentially coincides with the analytic signal Z (b) associated to s(x). It follows again that the restriction of the WT S(b, a) to its skeleton contains the whole information. In particular, the so-called frequency modulation law x −1 arg{s(x)} of s(x) is easily recovered from it. Thus, it is not necessary to compute the whole CWT, but only its skeleton [99]. This is, of course, much less costly computationally, because there are fast algorithms available. Spectacular applications of this method may be found, for instance, in spectroscopy [131], geomagnetism [4], chirp detection/estimation in gravitational waves [Mor02,277] or shape determination [21]. The last quoted paper, in particular, contains a thorough analysis of the two types of ridges, vertical and horizontal, interpreted in both cases as lines of local maxima (see Section 5.4.2).

1.5

The discrete WT: orthonormal bases of wavelets One of the successes of the WT was the discovery that it is possible to construct functions ψ for which {ψ jk , j, k ∈ Z} is indeed an orthonormal basis of L 2 (R). In addition, such a basis still has the good properties of wavelets, including space and frequency localization. Moreover, it yields fast algorithms, and this is the key to the usefulness of wavelets in many applications. The construction is based on two facts. First, almost all examples of orthonormal bases of wavelets can be derived from a multiresolution analysis, and then the whole construction may be transcripted into the language of digital filters, familiar in the signal processing literature. Notice that it is precisely at this point that arises the basic difference between the discretized continuous wavelet transform, discussed in the previous section, and the discrete wavelet transform (DWT). In the former case, the wavelet ψ is chosen a priori (with very few constraints, as we have seen above), and the question is whether one can find a lattice such that {ψ jk } is a frame with decent frame bounds A, B. In the other approach, one imposes from the beginning that the set {ψ jk } be an orthonormal basis and tries to construct a function ψ to that effect. The construction is rather indirect and the resulting function is usually very complicated (sometimes it has a fractal behavior). Definition 1.5.1 . A multiresolution analysis of L 2 (R) is an increasing sequence of closed subspaces . . . ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ . . . , with j ∈ Z V j = {0} and j ∈ Z V j dense in L 2 (R), and such that

(1.44)

20

Warm-up: the 1-D continuous wavelet transform

(1) f (x) ∈ V j ⇔ f (2x) ∈ V j+1 (2) There exists a function φ ∈ V0 , called a scaling function, such that the family {φ(x − k), k ∈ Z} is an orthonormal basis of V0 . Combining conditions (1) and (2), one gets an orthonormal basis of V j , namely {φ jk (x) ≡ 2 j/2 φ(2 j x − k), k ∈ Z}. Note that the scaling function φ is often required only to generate a Riesz basis of V0 , that is, a frame {φ0k , k ∈ Z} of linearly independent vectors spanning V0 . However, since the Riesz basis can be orthonormalized, for instance by a Gram–Schmidt procedure, the condition (2) is equivalent, and simpler. Remark: Some authors (for instance, [Dau92] or [Mal99]) use the opposite convention for the index j, namely V j ⊂ V j−1 (both have advantages and inconveniences). With the present convention, large j means small scale (of order) 2− j or high frequency 2 j , thus high resolution 2 j . Each subspace V j can be interpreted as an approximation space. The approximation of f ∈ L 2 (R) at the resolution 2 j is defined by its projection onto V j , and the larger j, the finer the resolution obtained. Then condition (1) means that no scale is privileged. The additional details needed for increasing the resolution from 2 j to 2 j+1 are given by the projection of f onto the orthogonal complement W j of V j in V j+1 : V j ⊕ W j = V j+1 ,

(1.45)

and we have:

Wj. L 2 (R) =

(1.46)

j∈ Z

Equivalently, fixing some lowest resolution level jo , one may write ∞

L 2 (R) = V jo ⊕ Wj .

(1.47)

j= jo

The crucial theorem then asserts the existence of a function ψ, sometimes called the mother wavelet, explicitly computable from φ, such that {ψ jk (x) ≡ 2 j/2 ψ(2 j x − k), j, k ∈ Z} constitutes an orthonormal basis of L 2 (R): these are the orthonormal wavelets. The construction of ψ proceeds roughly as follows. First, the inclusion V0 ⊂ V1 yields the relation (called the scaling, or two-scale, or refinement equation): φ(x) =

∞ √ 2 h k φ(2x − k),

h k = φ1,k |φ.

(1.48)

k=−∞

Taking Fourier transforms, this gives ∞ ) = h(ξ ) φ(ξ ), with h(ξ ) = √1 h k e−ikξ . φ(2ξ 2 k=−∞

(1.49)

21

1.5 The discrete WT: orthonormal bases of wavelets

Thus h is a 2π-periodic function and it satisfies the relation |h(ξ )|2 + |h(ξ + π)|2 = 1,

a.e.

(1.50)

Iterating (1.49), one gets the scaling function as the infinite product ) = (2π)−1/2 φ(ξ

∞

h(2− j ξ ),

(1.51)

j=1

which may be proven to be convergent [Dau92]. Then one defines the function ψ ∈ W0 ⊂ V1 by the relation ) = g(ξ ) φ(ξ ), ψ(2ξ

(1.52)

where g is another 2π -periodic function. By the relation (1.45) and the orthonormality of the functions {φ jk }, the functions h, g must satisfy the identity g(ξ ) h(ξ ) + g(ξ + π) h(ξ + π) = 0,

a.e.

(1.53)

The simplest solution is to put g(ξ ) = eiξ h(ξ + π), which implies, in particular, |h(ξ )|2 + |g(ξ )|2 = 1, a.e. Then one obtains ψ(x) =

∞ √ 2 (−1)k−1 h −k−1 φ(2x − k),

(1.54)

k=−∞

and one proves that this function indeed generates an orthonormal basis with all the required properties. Another, equivalent, solution is ψ(x) =

√

2

∞

(−1)k h −k+1 φ(2x − k).

(1.55)

k=−∞

Various additional conditions may then be imposed on the basic wavelet ψ, such as arbitrary regularity, several vanishing moments (in any case, ψ has always mean zero), symmetry, fast decrease at infinity, and even compact support [Dau92]. Remark: Some authors (for instance, [Dau92] or [Tor95]) denote the functions h, g by m 0 , m 1 , respectively. The simplest example of this construction is the Haar basis, which comes from the scaling function φ(x) = 1 for 0 x < 1, and 0 otherwise √ (boxcar function). The coefficients of the corresponding filter h are h 0 = h 1 = 1/ 2, h k = 0, for k = 0, 1. Applying the recipe (1.55) then yields the Haar wavelet 1, if 0 x < 1/2, (1.56) ψHaar (x) = −1, if 1/2 x < 1, 0, otherwise.

22

Warm-up: the 1-D continuous wavelet transform

Similarly, various B-spline bases may be obtained along the same line. Other explicit examples may be found in [Chu92] or [Dau92]. In order to set up the discrete WT, the technique consists in translating the multiresolution structure into the language of digital filters, which is precisely what we have just done. Indeed, a filter is simply a multiplication operator in frequency space or a linear convolution in the time variable, and the discussion above amounts to nothing more than expanding (filter) functions in a Fourier series. For instance, h(ξ ) is a filter, with Fourier coefficients h n , g(ξ ) is another one, and {h, g} are called Quadrature Mirror Filters or QMFs whenever they satisfy the identities (1.50) and (1.53). Then the various restrictions imposed on ψ translate into suitable constraints on the filter coefficients h n . For instance, ψ has compact support if only finitely many h n differ from zero (one then speaks of a finite impulse response or FIR filter). Of course, the goal is to obtain a fast algorithm, and this relies on two aspects, namely, short filters and a pyramidal structure, already familiar in signal processing. Indeed, the rapidity of the algorithms depends crucially on the length of the filters involved, because the pyramidal structure rests on a concatenation of several filters. One major stumbling block is the dilation. This is easy to understand from the very definition of the WT. Indeed, as we have seen above, the WT (1.8) or (1.13) basically a convolution. Once discretized, these formulas become discrete convolutions of digital sequences. Then the point is that, if the sequence ψ # (n) has length N , then the dilated sequence ψ2# (n) has length 2N , and this leads to an algorithm of exponential increase, clearly not admissible. One trick for avoiding this difficulty is to replace the (natural) dilation by a so-called pseudodilation, which consists in inserting a zero between any two successive entries of ψ # (n), and then correcting for the distortion so introduced. In this way, one obtains a fast algorithm. Since the sequences resulting from successive dilations have all the same length, but are full of zeros or holes, this algorithm is known as the “algorithme a` trous” [157,222]. The interesting fact is that this procedure may be extended to very general situations, involving wavelet transforms on abelian groups, which form a kind of intermediate step (“missing link”) between the CWT and the DWT [Kou00,32]. Several other fast algorithms have been designed, mostly along the line proposed by Mallat [Mal99,339]. Altogether, there does exist a Fast Wavelet Transform, exactly as a Fast Fourier Transform. In practical applications, the (sampled) signal is taken in some V J , and then the decomposition (1.47) is replaced by the finite representation V J = V jo ⊕

J −1

Wj .

(1.57)

j= jo

Figure 1.8 shows an example of a decomposition of order 5, namely our familiar signal bumps decomposed over an orthonormal basis of Daubechies d6 wavelets [Dau92]. Thus we take J = 0 and jo = −5 in formula (1.57):

23

1.5 The discrete WT: orthonormal bases of wavelets

Signal 2 1 0 1 0 −1 1 0 −1 1 0 −1 2 0 −2 2 0 −2 2 0 −2 10 0 −10

100

200

300

400

500

600

700

800

900

1000

Fig. 1.8. Five-level decomposition of the bumps signal on an orthonormal basis of Daubechies d6

wavelets. The low resolution approximation c−5 ∈ V−5 is shown on the bottom panel and the five levels of details with increasing resolution, d j ∈ W j , j = −5, . . . , −1, in the next five panels.

V0 = V−5 ⊕ W−5 ⊕ W−4 ⊕ W−3 ⊕ W−2 ⊕ W−1 .

(1.58)

Correspondingly, the signal s ∈ V0 is decomposed as s=

c−5,k φ−5,k +

k∈Z

−1

d jk ψ jk

j=−5 k∈Z

≡ c−5 +

−1

dj.

j=−5

As in Figure 1.5, the figure shows, from bottom up, the low resolution approximation c−5 and the five levels of details with increasing resolution, successively d−5 , d−4 , d−3 , d−2 , d−1 . At this point, we should add a word of caution concerning the numerical implementation of the reconstruction formula associated to (1.57), s=

k∈Z

c jo ,k φ jo ,k +

J −1 j= jo k∈Z

d jk ψ jk .

(1.59)

24

Warm-up: the 1-D continuous wavelet transform

If s is an analog signal, i.e. a function, the approximation coefficients c jo ,k and the wavelet coefficients d j,k are calculated in the standard way, c jo ,k = φ jo ,k |s,

d j,k = ψ jk |s.

However, if the signal is only accessible through sampled values, then the coefficients must be estimated from the latter. This may be done by making some assumption about the signal (for instance, in the “algorithme a` trous” [157,222], the signal is supposed to be a spline function), which amounts to some pre-filtering. Alternatively, one takes for c jo ,k the sampled values themselves. This is a good approximation [Mal99; Section 7.2.3], but without any real theoretical justification. In any case, these procedures generate errors that have to be controlled. See [132] for a comprehensive discussion.

1.6

Generalizations As we just saw, appropriate filters generate orthonormal wavelet bases. However, this result turns out to be too rigid and various generalizations have been proposed (see also the comments in Section 1.6.2). To name a few: biorthogonal wavelet bases, wavelet packets and the Best Basis Algorithm, the lifting scheme and second generation wavelets. We shall refrain from describing these here. Instead, we will give a rather detailed treatment in Chapter 2, for the two-dimensional case. Further information may also be found in [Mey94].

1.6.1

Continuous wavelet packets Besides the full discretization described in Section 1.3, and the discrete WT just discussed, there is an intermediate procedure, which consists in discretizing the scale variable alone, on an arbitrary sequence of values (not necessarily powers of a fixed ratio), but leaving translations fully continuous. The resulting transform has the advantage of being completely covariant with respect to translations, a very desirable feature, for instance, in pattern recognition. If we use dyadic scales, the result is called a dyadic wavelet transform and was introduced in [264,265], precisely for that reason. A detailed account will be given in the 2-D case, in Section 2.4.4. An elegant way of deriving this dyadic WT from the CWT was described in [159], under the name of infinitesimal multiresolution analysis, or continuous wavelet packets. This approach leads to fast algorithms that could put the CWT on the same footing as the discrete WT in terms of speed and efficiency, by extending the advantages of the latter to cases where no exact QMF is available [Tor95,Vdg98,291,360]. While already interesting in 1-D, this method displays its full potential in 2-D, offering a very fast implementation of the so-called directional 2-D wavelets. Accordingly we shall describe it in detail in the next chapter, Section 2.6. Here we shall only sketch briefly the idea, in the version called linear formalism in [Tor95,291].

25

1.6 Generalizations

Given a wavelet ψ, one lumps together all low-frequency components in a scaling function (here we take a > 0) ∞ ∞ x da 1 x ) da , (x) = (1.60) ψ(aξ ψ = ψ(s) ds, (ξ ) = 2 a a x 0 a 1 1 and introduces the integrated wavelet 1 x da 1 2x ψ = ψ(s) ds, (x) = a a2 x x 1/2

)= (ξ

1

) da . ψ(aξ a 1/2

(1.61)

These functions satisfy two-scale relations: (x) = 2 (2x) − (x),

) = (ξ/2) ). (ξ − (ξ

(1.62)

Next, one chooses a regular grid, as opposed to the dyadic one used in the discrete case, namely, xj ≡ 2 j (2 j (· − x)),

xj ≡ 2 j (2 j (· − x))

(1.63)

[note that one uses here the L 1 -normalization (1.19)]. Although the resulting transform will be redundant, it has the great advantage over the conventional DWT of maintaining (integer) translation covariance. Then, exactly as for the DWT, one gets a discrete reconstruction formula: s(x) = xjo |s +

∞

xj |s.

(1.64)

j= jo

Truncating the summation, as usual, one gets thus a finite sum, to be compared with the decomposition (1.57). However, there still remains a major problem. Indeed, we may try and mimick the ˇ gˇ satisfying the following continuous formalism and assume there exist two functions h, relations, analogous to (1.49), (1.52): ) = h(ξ ), ˇ ) (ξ (2ξ

) = g(ξ ), ˇ ) (ξ (2ξ

a.e.

(1.65)

The difficulty now is that these functions are in general not 2π -periodic, which precludes designing any fast (pyramidal) algorithm. There is a way out, however. Since using the regular grid means sampling (x) at unit rate, we have to assume that the function ˇ gˇ always appear is essentially supported in [−π, π ]. Therefore, since the functions h, in a product with , according to the relations (1.65), it is reasonable to approximate them in a neighborhood of zero by 2π-periodic functions h a , g a . In fact [Tor95], there exists a unique pair h, g that minimizes the distance between hˇ and h a , respectively gˇ a and g . These approximate filters, called pseudo-QMFs, will be described at length in Section 2.6.4. The end result is a very fast implementation of the continuous wavelet transform, truly competitive with the discrete wavelet transform.

26

Warm-up: the 1-D continuous wavelet transform

1.6.2

Orthogonal or redundant wavelet expansions? Now we have seen the full spectrum of possible wavelet decompositions, from the minimalist, that is, nonredundant, orthonormal bases to increasingly redundant systems, frames and pseudo-QMFs. Which one to use in practice? Of course, there is no unique answer, it depends on the problem at hand. If it comes to signal compression, for instance, the most economical representation is certainly preferable, thus orthonormal bases will be the first choice. In addition, they often yield the least correlation between wavelet coefficients, and, of course, mathematicians have a long tradition in orthogonal expansions. This explains the popularity of orthonormal bases in many communities. However, this is certainly not the last word. In statistical analysis, for instance, one prefers overcomplete systems to orthonormal bases, because they have higher adaptivity properties and they allow to control the degree of redundancy. In addition, the lack of (even discrete) translation invariance is a serious drawback for all applications involving some pattern recognition. In fact, increasing redundancy has many advantages, in particular, it improves both the quality of reconstruction and the stability with respect to perturbation, e.g., by noise. We will now comment on these two aspects. As an illustration, let us compare the three analyses of the bumps signal made above, the full CWT analysis of Figure 1.3, the translation invariant frame decomposition of Figure 1.5 and the orthonormal basis decomposition of Figure 1.8. The CWT sees correctly all discontinuities, with intensities depending of their strength, measured by the singularity index α, defined in (1.31). The three parts of the signal have, respectively, α = −1 for the δ function, α = 0 for the boxcar function, and α = 1 for the tent or triangle function. Correspondingly, see (1.32), the wavelet coefficients behave as a α , which fits with Figure 1.3 for a small enough. Compare now the two finite decompositions. When a → 0, the wavelet coefficients of the δ function increase, whereas those of the triangle function decrease. As a consequence, both the orthonormal basis and the translation invariant frame barely see the discontinuities of the triangle. The high resolution (small scale) wavelet coefficients are so small that one cannot see them on the figure! Indeed, if we redo the same analysis on the triangular signal alone, using appropriate units, the high resolution coefficients are perfectly seen (Figure 1.9). The reason is that the δ function singularity is so strong that it swamps the whole picture. This is a perfect illustration of the fact that wavelet orthonormal bases are too rigid: the same basis cannot analyze correctly the three pieces of the signal, in a sense it is not local enough! For that reason, substitutes were invented for improving the local character of the analysis. One of the most popular examples is that of a local cosine basis. The idea is to divide the time axis into arbitrary intervals (which overlap slightly for continuity), depending on the signal itself, and performing into each of these an independent discrete Fourier-type decomposition (DCT) [see Figure 1.11 (d)]. We refer to [Mal99] for further information (we quote this technique here for the sake of comparison only).

27

1.6 Generalizations

Signal 1 0.5 0 x 10−3 2 0 −2 x 10−3 5 0 −5 0.01 0 − 0.01 0.02 0 − 0.02 0 0.5 0 − 0.5 10 0 −10

100

200

300

400

500

600

700

800

900

1000

Fig. 1.9. Five-level decomposition of the triangular part of the bumps signal, with the same

presentation as in Figure 1.8.

On the other hand, the frame of Figure 1.5 performs better. The reason here is spatial (time) resolution. In the orthogonal basis case, the resolution becomes so loose at larger scales that even the wavelet coefficients at j = −3 are barely visible on Figure 1.8. But, since the frame is translation invariant, the spatial resolution remains the same at all scales, and the wavelet coefficients are now visible up to j = −1. Another bonus of redundancy is an increased robustness of the representation with respect to small perturbations. Suppose indeed we are working in finite dimension N . Let s ∈ R N be a signal and denote by si , i = 1 . . . N the coefficients of s in a given orthonormal basis. Suppose we perturb each si by a random variable n i (modeling noise, for example). For simplicity, we choose the n i s independent, with zero mean and variance σ 2 . It is easy to verify that the mean square error (MSE) of the perturbed signal s˜ is N σ 2 . Now, if we decompose s in a frame composed of M > N elements, the intuition is that this result should be improved because we have diluted noise in a much larger space. Actually, if the frame is tight and denoting by r = M/N its redundancy, the MSE reduces to [200]: MSE =

Nσ2 . r

Warm-up: the 1-D continuous wavelet transform

5

0

−5 Log MSE

28

−10

−15

−20

0

2

4

6

8 Number of Bits

10

12

14

16

Orthogonal Wavelet Basis Redundant Wavelet Frame

Fig. 1.10. Mean square error of the two reconstructions of the bumps signal, using an orthonormal wavelet basis and a redundant translation invariant wavelet frame, respectively.

Redundancy thus implies stability, or robustness to perturbations of the decomposition coefficients. This is illustrated on Figure 1.10. We have decomposed the bumps signal using an orthonormal wavelet basis and a redundant translation invariant wavelet frame. We have then quantized the respective coefficients and measured in each case the MSE of the reconstructed signal. The plot shows that the redundant representation always yields a better MSE. Furthermore, the distance between the two curves is constant and equal to the redundancy of the frame used. Finally, as a visual aid to the various discrete wavelet schemes, it is instructive to characterize each of them by the tiling of the time–frequency plane it induces. We show a schematic presentation of these in Figure 1.11. On the top row, we have, from left to right, the cases of a Gabor transform and of the DWT (dyadic partition). Note, however, that the Gabor tiling is idealized, since this sharp partition can never be realized, because the time–frequency localization is intrinsically limited (the so-called Balian–Low theorem [Dau92]). On the bottom row, we show the cases of wavelet packets and that of a local cosine basis, respectively.

1.7 Applications of the 1-D CWT

Frequency

Frequency

29

Time

Time

(b)

Frequency

Frequency

(a)

Time

(c)

Time

(d)

Fig. 1.11. Tilings of the time–frequency plane corresponding to four different analysis schemes. (a) The (idealized) Gabor analysis; (b) the discrete WT; (c) wavelet packets; (d) a local cosine basis.

1.7

Applications of the 1-D CWT The CWT has found a wide variety of applications in various branches of physics and/or signal processing. Of course, our main concern in this book will be the practical applications of the 2-D wavelet transform, so we will devote two full chapters to them (Chapters 4 and 5). Nevertheless, we will list here a representative selection of onedimensional applications, in order to convey to the reader a feeling about the scope and richness of the field. Most of the early applications, and the original references, may be found in the proceedings volumes [Com89,Mey91,Mey93]. Another interesting source for applications is the recent volume of Addison [Add02]. In all cases, the CWT is primarily used for analyzing transient phenomena, detecting abrupt changes in a signal or comparing it with a given pattern. r Sound and acoustics The first applications of the CWT were in the field of acoustics. A few examples are musical synthesis, speech analysis [123] and modeling of the sonar system

30

Warm-up: the 1-D continuous wavelet transform

r

r

r

r

r

of bats and dolphins. Other examples include various problems in underwater acoustics, such as the disentangling of the different components of an underwater refracted wave (see Section 5.3 for the 2-D case) and the identification of an obstacle. Geophysics This is the origin of the method, which was designed in an empirical fashion by J. Morlet for analyzing the recordings of microseisms used in oil prospection. More recently, the CWT has been applied to the analysis of various types of geophysical data, e.g., in gravimetry (fluctuations of the local gravitational field), in seismology (arrival time of the various waves), in geomagnetism (fluctuations of the Earth magnetic field [4]) or in astronomy (fluctuations of the length of the day, variations of solar activity, measured by the sunspots, etc). Fractals, turbulence The CWT is an ideal tool for studying fractals, or more generally phenomena with particular properties under scale changes [221]. Thus it is quite natural that the CWT has found many applications in the analysis of (1-D and 2-D) fractals, artificial (diffusion limited aggregates) or natural (arborescent growth phenomena) [Arn95,43,44]. Related to these is the use of the CWT in the analysis of developed turbulence (identification of coherent structures, uncovering of hierarchical structure) [Abr97,163–165]. An interesting example of fractal or self-similar behavior is that of telecommunications network traffic, and here too the WT (although rather the DWT) has given interesting results [1]. We will come back to 2-D fractals and similar objects in Section 5.4. Atomic physics When an atom is hit by a short intense laser pulse, it emits radiation that cover a whole spectrum of harmonics of the laser frequency (experimentally, harmonics of order larger than 400 have been observed). This is a fast and complex physical process, which cannot be understood without a time-frequency analysis. This has been done, both with a Gabor analysis and with wavelets (CWT), yielding for instance the time profile of each individual harmonic [30,31] and the effect of the polarization of the laser field on harmonic generation [39,40]. Spectroscopy This was one of the earliest and most successful applications, in particular for NMR spectroscopy, where the method proves extremely efficient in subtracting unwanted spectral lines or filtering out background noise [60,131,210]. We may note that here, as in the previous application, a Gabor analysis may be fully competitive with the wavelet analysis [33,34]. Medical and biological applications The CWT has been used for analyzing or monitoring various electrical or mechanical phenomena in the brain (EEG, VEP) or the heart (ECG) [Ald96,Tho98,354]. It also yields good models for the auditory mechanism [123]. Another striking result is the

31

1.7 Applications of the 1-D CWT

characterization of long-range correlations in DNA sequences (and the solution of a long-standing puzzle) by Arn´eodo and his group [49]. r Analysis of local singularities The strong point of the CWT is to detect singularities in a signal, but it yields also a fine characterization of their strengths (Lipschitz regularity, expressed via the local H¨older exponents), using the homogeneity relations (1.32) and (1.34), in particular in the case of oscillating singularities [48,50,51]. r Shape characterization A particular case of analysis of local singularities is the determination of the shape of an object, a standard problem in image processing, for instance in robotic vision. A novel approach [Ces97,Cos01,21] consists in treating the contour of the object as a complex curve in the plane and analyzing it with the 1-D CWT. The method benefits from all the good properties of the wavelet transform, for instance its robustness to noise, and looks promising for applications. Since this is in fact a 2-D problem, we will analyze it in detail, in Section 5.4.2. r Industrial applications Here again the important aspect is monitoring, for instance in detecting anomalies in the functioning of nuclear, electrical or mechanical installations. A typical application is the analysis of the behavior of materials under impact made in [358] and discussed above for illustrating the concept of ridge and skeleton. In that paper, the same signal is analyzed with a Gabor transform, a CWT with a Morlet wavelet, a CWT with a Mexican hat, a Wigner–Ville transform, and the respective merits of each method are compared. Another, closely related, application is the determination, with a Morlet wavelet, of the vibration normal modes of a high tower excited by wind [259]. The results of this paper fully confirm the previous ones and, in particular, emphasize the role of the width parameter of the Gaussian.

2

The 2-D continuous wavelet transform

Chapter 1 has given us a brief overview of the basic facets of the CWT in the simpler one-dimensional (1-D) case, including its relationship with the various discrete approaches and a glimpse of some applications. Now it is time to enter the proper subject of the book, namely, the two-dimensional (2-D) wavelet analysis.

2.1

Derivation In 1-D, the CWT (1.8) amounts to projecting the signal onto the wavelet ψb,a , obtained by translation and dilation of the mother wavelet ψ. Thus the transform is fully determined by these elementary operations of the line. Accordingly, in order to derive the CWT in 2-D, a good starting point is to consider first the elementary operations we want to apply to our signals. Actually, as we will see later (Chapter 6), this point of view allows one to extend the CWT to much more general situations, such as wavelets in higher dimensions, wavelets on the sphere, time-dependent wavelets, etc.

2.1.1

Images and elementary operations on them By an image, we mean a two-dimensional signal of finite energy, represented by a complex-valued function defined on the real plane R2 and square integrable, i.e., a function s ∈ L 2 (R2 , d 2 x): 2 s = d 2 x |s( x )|2 < ∞ (2.1) R2

(sometimes it is useful to take s integrable as well). In practice, a black and white image will be represented by a bounded non-negative function: 0 s( x ) M, ∀ x ∈ R2 (M > 0),

(2.2)

the discrete values of s( x ) corresponding to the level of gray of each pixel. However it is useful to keep general functions s as above. In fact, one often considers also as admissible signals generalized functions (distributions), such as a delta function 32

33

2.1 Derivation

δ( x − xo ), a plane wave exp(i k · x), a fractal measure, etc. This will be justified below (see the comments after Definition 2.1.3). The Fourier transform of the signal s is defined, as usual, by 1 s(k) ≡ (Fs)(k) = d 2 x e−i k·x s( x ), (2.3) 2π R2 where k ∈ R2 is the spatial frequency and k · x = k1 x1 + k2 x2 is the Euclidean scalar 2 = k · k. Of course, the Fourier transform is unitary (Parseval product. We also write |k| relation): s ∈ L 2 (R2 , d 2 k)

and

s2 = s2 .

(2.4)

Given an image s, all the geometric operations we want to apply to it are obtained by combining three elementary transformations of the plane, namely, rigid translations in the plane of the image, dilations or scaling (global zooming in and out) and rotations. (More complicated operations are sometimes applied, such as deformations (shearing), but we will not consider them here. We will come back to this point in Chapter 7.) Explicitly, the transformations act on x ∈ R2 in the familiar way: (i) translation by b ∈ R2 : x → x = x + b; (ii) dilation by a factor a > 0 : x → x = a x; (iii) rotation by an angle θ : x → x = rθ ( x ), where rθ is the usual 2 × 2 rotation matrix: cos θ − sin θ , 0 θ < 2π. rθ ≡ sin θ cos θ It will prove convenient to combine a rotation by an angle θ and a dilation by a > 0 into a single 2 × 2 matrix, namely, a cos θ −a sin θ h = h(a, θ) = . (2.5) a sin θ a cos θ Using this form, we verify that x · x = a cos θ | x |2 , demonstrating that the angle be

tween x and x is indeed θ and that x is scaled by the amount a relative to x . Finally, combining all three operations, we get as general transformation in the plane x + b. x → x = h

(2.6)

In the present context, these transformations are represented by the following unitary operators in the space L 2 (R2 , d 2 x) of finite energy signals: b ∈ R2 ; (i) translation : (Tb s)( x ) = s( x − b), (2.7) −1 −1 (ii) dilation : (Da s)( x ) = a s(a x), a > 0; (2.8) (iii) rotation : (Rθ s)( x ) = s(r−θ ( x )), θ ∈ [0, 2π). (2.9)

34

The 2-D continuous wavelet transform

In addition, we introduce the modulation operator: (E b s)( x ) = ei b·x s( x ), b ∈ R2 .

(2.10)

Then a straightforward calculation yields the commutation rules among the operators (2.7)–(2.9), and with the Fourier operator (2.3): Tb Da = Da Tb/a F Da = D1/a F, , Tb Rθ = Rθ Tr−θ (b) , F Tb = E −b F, F Rθ = Rθ F. Rθ Da = Da Rθ ,

(2.11)

Combining now the three operators (2.7)–(2.9), we define the unitary operator a, θ ) = Tb Da Rθ , U (b,

(2.12)

which acts on a given function s as a, θ )s ( U (b, x ) ≡ sb,a,θ ( x ) = a −1 s(a −1 r−θ ( x − b)),

(2.13)

or, equivalently, in the space of Fourier transforms,

= a e−i b·k s(ar−θ (k)). s (k) b,a,θ

(2.14)

If the function s is rotation invariant, we simply omit the index θ: sb,a x ) = a −1 s(a −1 ( x − b)). (

(2.15)

We may remark that, here, contrary to the 1-D case, it is sufficient to take positive dilations, since the effect of a negative dilation a < 0 may be obtained by combining a positive one, a > 0, with a rotation by θ = π . The geometrical effect of these transformations is easily visualized assuming s and s to be well localized, for instance in an ellipse. An example is shown in Figure 2.1 (in Section 2.3.1).

2.1.2

Wavelets and continuous wavelet transform As in 1-D, a wavelet is a particular type of finite energy signal, whose properties make it a good analyzing tool. Thus we define, as in (1.10): Definition 2.1.1 . A two-dimensional wavelet is a complex-valued function ψ ∈ L 2 (R2 , d 2 x) satisfying the admissibility condition: k)| 2 |ψ( cψ ≡ (2π)2 d 2 k < ∞, (2.16) 2 |k| R2 is the Fourier transform of ψ and |k| 2 = k · k = (k1 )2 + (k2 )2 . where ψ The origin of this condition will be clarified below.

35

2.1 Derivation

If ψ is regular enough, the admissibility condition (2.16) implies the following easier one, which simply means that the wavelet has zero mean: 0) = 0 ⇐⇒ ψ( d 2 x ψ( x ) = 0. (2.17) R2

Strictly speaking, the condition (2.17) is only necessary, but in fact it is almost sufficient (see [Dau92] for a precise mathematical statement), and for all practical purposes (2.17) may be taken as admissibility condition. Intuitively, as in one dimension, it expresses the fact that a wavelet must be an oscillating function. Clearly the three unitary operators Tb , Da , Rθ preserve the admissibility condition, a, θ ). Hence any function ψb,a,θ a, θ)ψ obtained and so does therefore U (b, = U (b, from a wavelet ψ by translation, rotation or dilation is again a wavelet. Thus the given wavelet ψ generates the whole family Dψ = {ψb,a,θ }, indexed by the elements b ∈ R2 , a > 0, θ ∈ [0, 2π). In the sequel we will denote by G this four-dimensional parameter space. a, θ) ∈ G} is a Proposition 2.1.2 . The linear span of the family Dψ = {ψb,a,θ , (b, 2 2 dense subspace of L (R ). Proof . Let f ∈ L 2 (R2 , d 2 x) be orthogonal to every vector in the family Dψ , that is a, θ) ∈ G. This means ψb,a,θ | f = 0, ∀ (b, a, θ ) ∈ G, | f = a d 2 k ei b·k ψ(ar ∀ (b, ψ −θ (k)) f (k) = 0, b,a,θ R2

−θ (k)) f (k) = 0 a.e., for all a > 0, θ ∈ [0, 2π ). Now the joint which implies that ψ(ar is a “patch” action of rotations and dilations on R2 is transitive. Thus, if the support of ψ −θ (k)) will cover (for instance, a disk or an ellipse) in the k-plane, the supports of ψ(ar the whole plane when a and θ vary over their range. Therefore, this implies f (k) = 0 a.e., that is, f = 0. Note that rotations are needed to get that result. A naive generalization of the 1-D formalism would consist in combining translations with separate dilations along the x- and the y-axis. But then, if only positive dilations are used, each quadrant in the k-plane is invariant and additional conditions on the wavelet ψ would be necessary for the argument of the proof above to work – thus spoiling the result. Actually, this is precisely the technique used in the 2-D DWT, where the 2-D multiresolution is obtained by taking the tensor product of two 1-D copies, one in x, one in y (see Section 2.5.1). As we shall see later (see Section 2.5.2), this approach, while commonly used in practice, has severe shortcomings. As a consequence of Proposition 2.1.2, any vector in L 2 (R2 , d 2 x) is uniquely determined by its projections on the vectors of the family Dψ . This justifies the basic definition of the CWT.

36

The 2-D continuous wavelet transform

Definition 2.1.3 . Given an image s ∈ L 2 (R2 , d 2 x), its continuous wavelet transform (with respect to the fixed wavelet ψ), S ≡ Tψ s is the scalar product of s with the a, θ): transformed wavelet ψb,a,θ , considered as a function of (b, a, θ ) = ψb,a,θ S(b, |s s( d 2 x ψ(a −1 r−θ ( x − b)) x) = a −1 R2 s(k). d 2 k ei b·k ψ(ar =a −θ (k))

(2.18) (2.19) (2.20)

R2

The relations (2.18)–(2.20) permit us to extend the formalism beyond the Hilbert space framework. As explained above, the signal s may be taken as a singular function (a distribution), provided the wavelet ψ is sufficiently regular (most wavelets used in practice are smooth functions, see below). Before exploring in detail the mathematical properties of the CWT, it is instructive to exhibit two typical 2-D wavelets (actually the simplest ones). (1) The isotropic Mexican hat wavelet This wavelet is simply the Laplacian of a Gaussian: ψH ( x ) = (2 − | x |2 ) exp(− 12 | x |2 ), H (k) = |k| 2 exp(− 1 |k| 2 ). ψ

(2.21)

2

(2) The Morlet wavelet This wavelet is essentially a plane wave within a Gaussian window: ψM ( x ) = exp(i ko · x) exp(− 12 | x |2 ) + corr.; M (k) = exp(− 1 |k − ko |2 ) + corr. ψ

(2.22)

2

As in 1-D, a a correction term must be added in order to satisfy the admissibility condition (2.17), but in practice one will arrange that this term be numerically negligible and thus can be omitted (it suffices to choose the norm |ko | of the wave vector large enough). The first wavelet is real, the Morlet wavelet is complex. They have very different properties and, naturally, they will be used in quite different situations. Both wavelets, and many more, will be studied in detail in Chapter 3, and many examples of applications in Chapters 4 and 5.

2.2

Basic properties of the 2-D CWT The main properties of the continuous wavelet transform are conveniently expressed in terms of a linear map Wψ from the space of finite energy signals L 2 (R2 , d 2 x) into the space of transforms. We summarize them in three propositions [Mur90,13,15,283].

37

2.2 Basic properties of the 2-D CWT −1/2

Proposition 2.2.1 . Let the map Wψ : s → cψ

S be defined by

a, θ) = cψ−1/2 ψb,a,θ (Wψ s)(b, |s, s ∈ L 2 (R2 , d 2 x), where cψ is the constant given in (2.16). Then: (1) Wψ conserves the norm of the signal, thus its total energy: 2 da 2 d b 3 dθ |S(b, a, θ )| = cψ d 2 x |s(x)|2 , a G R2

(2.23)

(2.24)

i.e., it is an isometry from the space of signals into the space of transforms. The latter is a closed subspace Hψ of L 2 (G, dg), where dg ≡ a −3 d 2 b da dθ is the natural measure on G. Equivalently, the family of wavelets {ψb,a,θ }, with b ∈ R2 , a > 0, and 0 θ < 2π, generates a resolution of the identity: da −1 cψ d 2 b 3 dθ |ψb,a,θ ψb,a,θ | = I. (2.25) a G (2) Since it is an isometry, the map Wψ is invertible on its range Hψ , and the inverse transformation is the adjoint of Wψ . This means that the image s( x ) may be a, θ) by the formula: reconstructed from its wavelet transform S(b, da a, θ). s( x ) = cψ−1 d 2 b 3 dθ ψb,a,θ ( x ) S(b, (2.26) a G Proof . The relation (2.24) follows from a straightforward calculation: da a, θ)|2 = d 2 b 3 dθ |S(b, a G da 2 2 dθ = d k d k d 2 b a R2 R2 G

−θ (k)) −θ (k )) ψ(ar s(k ) s(k) × ei b·(k−k ) ψ(ar ∞ 2π da −θ (k))| 2 | 2 d 2 k dθ |ψ(ar s(k)| = (2π)2 a 0 0 R2 (the exchange of integrals is justified by Fubini’s theorem). Introducing polar coordi k) p (ρ, φ), with ρ ≡ |k|, ≡ψ we get nates: ψ( ∞ ∞ da 2π da 2π −θ (k))| p (aρ, φ − θ )|2 2= dθ |ψ(ar dθ |ψ a a 0 0 0 0 ∞ 2π dρ p (ρ , θ )|2 = dθ |ψ ρ 0 0 d 2 k 2 = |ψ(k )| . |2 R2 |k By comparison with the definition (2.16) of cψ , and using Plancherel’s theorem, this proves the statement.

38

The 2-D continuous wavelet transform

Then (2.25) is simply a reformulation of (2.24), in view of the definition (2.23) of −1/2 Wψ : s → cψ S. Finally, the reconstruction formula (2.26) follows immediately by applying both sides of (2.25) to a signal s( x ) and taking into account the definition (2.23) of Wψ .

Three remarks are in order here. First, the relation (2.25) must be taken as a weak integral, that is, both sides are equal when sandwiched between arbitrary vectors. This precisely means that the reconstruction formula (2.26) holds in the weak sense. But, in fact, much more is true, and actually necessary for obtaining good approximation schemes, namely, the relation (2.26) holds in strong L 2 convergence. This will be demonstrated in Section 2.6.1, in two different versions. Second, the measure dg ≡ a −3 d 2 b da dθ on G is precisely the unique measure (up to normalization) that is invariant under all the operations of translation, dilation, and rotation (this is why we have called it natural). Third, the reconstruction formula (2.26) may also be proven by an explicit calculation of the adjoint map of Wψ : −1/2

f |sL 2 (R2 ) = cψ

= cψ−1

−1/2

f |Wψ∗ SL 2 (R2 ) = cψ Wψ f |SL 2 (G) da a, θ). d 2 b 3 dθ f |ψb,a,θ S(b, a G

It is the possibility of having a reconstruction formula (2.26) that justifies the admissibility condition cψ < ∞ imposed on wavelets. However, (2.26) is not only a reconstruction formula, it also means that the wavelet transform, like its 1-D counterpart, provides a decomposition of the signal in terms of the analyzing wavelets ψb,a,θ , a, θ ). Under both interpretations, this formula leads in practice with coefficients S(b, to discretization problems (see Section 2.4). In the same spirit, it is interesting to see the inverse Fourier transform 1 s( x) = s(k), (2.27) d 2 k ei k·x 2π R2

as the decomposition of the signal into the improper basis {ei k·x , k ∈ R2 } of eigenvectors of the translation operators (2.7). In view of the crucial importance of dilations in the wavelet context, it is useful to write down also the polar coordinate version of the Fourier transform, which involves the (improper) eigenvectors {einϕ , n ∈ Z} of the rotation operator (2.9) and those of the dilation operator (2.8), {riν , ν ∈ R}: 2π ∞ 1 dr −iν f (ν, n) = dϕ e−inϕ (2.28) r r f (r, ϕ), 2π 0 r 0 ∞ ∞ 1 inϕ f (ν, n). (2.29) e dν r iν r f (r, ϕ) = 2π n=−∞ −∞

39

2.2 Basic properties of the 2-D CWT

Of course, we recover the well-known fact that the polar coordinate version of the Fourier transform is a combination of a Mellin transform in the radial variable r and a Fourier series in the angle ϕ. Actually, as in one dimension, the reconstruction formula (2.26) may be generalized in several ways. First, the wavelet used for the analysis, ψ, and the one used for the reconstruction, χ , need not coincide, they have only to satisfy a cross-admissibility condition [Hol95,223], namely, 0 < |cψχ | < ∞, where d 2 k 2 < ∞. ψ(k) χ (k) (2.30) cψχ = (2π) 2 R2 |k| Then one gets a more general reconstruction formula: da −1 a, θ). d 2 b 3 dθ χb,a,θ ( x ) (Wψ s)(b, s( x ) = cψχ a G

(2.31)

The proof consists in a straightforward verification, including some interchanges of integrals justified by Fubini’s theorem. As we shall see later, this is the analog, in the continuous case, of the bilinear scheme commonly used in the discrete approach, namely the construction of biorthogonal wavelet bases (see Section 2.5.2). In particular, if one takes for the reconstruction wavelet χ a delta function, one obtains the simplified reconstruction formula: ∞ da 2π −1 s( x ) = cψδ dθ S( x , a, θ), (2.32) a2 0 0 where cψδ =

R2

d 2 k ψ(k). 2 |k|

On the other hand, if ψ is rotation invariant, the wavelet transform S does not depend on θ and we obtain, instead of (2.26), a simpler reconstruction formula: ∞ da a). d 2 b ψ ( x ) S(b, (2.33) s( x ) = 2π cψ−1 a 3 b,a R2 0 Finally, combining the two preceding points, one obtains the simplified reconstruction formula originally used by Morlet in 1-D [206], in which one reconstructs the original image by summing over scales only: ∞ da −1 s( x ) = 2π cψδ S( x , a). (2.34) a2 0 Next, a characteristic feature of the CWT, in fact shared by a large class of transformations, as we shall see later (Chapter 6), is the existence of a so-called reproducing kernel, which actually is nothing but the wavelet transform of the wavelet itself, that is, the autocorrelation function of the wavelet. More precisely:

40

The 2-D continuous wavelet transform

Proposition 2.2.2 . The projection from L 2 (G, dg) onto the range Hψ of Wψ , the space a, θ ) is the of wavelet transforms, is an integral operator whose kernel K (b , a , θ |b, autocorrelation function of ψ, also called reproducing kernel: a, θ ) = cψ−1 ψb ,a ,θ |ψb,a,θ K (b , a , θ |b, .

(2.35)

Therefore, a function f ∈ L 2 (G, dg) is the wavelet transform of a certain signal iff it satisfies the reproduction property: da

a, θ) f (b, a, θ). d 2 b 3 dθ K (b , a , θ |b, (2.36) f (b , a , θ ) = a G Proof . Since Wψ is an isometry from L 2 (R2 , d 2 x) into L 2 (G, dg), i.e., Wψ∗ Wψ = I , its range Hψ is a closed subspace and the corresponding projection operator is Pψ = Wψ Wψ∗ . Thus a vector f ∈ L 2 (G, dg) belongs to Hψ iff f = Pψ f . Explicitly, this gives: f (b , a , θ ) = Wψ Wψ∗ f (b , a , θ ) da −1 a, θ ), = cψ d 2 b 3 dθ ψb ,a ,θ |ψb,a,θ f (b, a G which proves (2.35)–(2.36).

Because it may be interpreted as the autocorrelation function of the wavelet, the reproducing kernel leads to the notion of correlation length, that is, it determines the a, θ parameter space. As such, it plays region of influence of a given wavelet in the b, a role in the determination of the capabilities of a given wavelet (calibration), and in particular in the process of discretization. We will discuss these features in Chapter 3, Section 3.4. Finally, the continuous wavelet transform has the important property of covariance (improperly called invariance in the signal processing literature) under all the operations used in its definition.

Proposition 2.2.3 . The map Wψ is covariant under translations, dilations and rotations, a, θ) implies the following which means that the correspondence Wψ : s( x ) → S(b, ones: s( x − bo ) → S(b − bo , a, θ ) ao−1 a, θ) ao−1 s(ao−1 x) → S(ao−1 b,

(2.38)

a, θ − θo ). x )) → S(r−θo (b), s(rθo (

(2.39)

(2.37)

41

2.3 Implementation and interpretation of the 2-D CWT

It is worth noting that, conversely, the wavelet transform is uniquely determined by the three conditions of linearity, covariance and energy conservation, plus some continuity [Mur90]. These covariance relations, which are proved by a straightforward calculation, have a crucial importance for the applications. Translation covariance (2.37), often called improperly shift invariance, is lost in the standard formulation of the discrete WT, based on multiresolution (see Definition 1.5.1 and Section 2.5.1), and this generates many problems in practice, for instance in pattern recognition. Covariance under dilations, (2.38), is the basis for the application of the wavelet transform to the analysis of fractals (see Section 5.4). Finally, joint covariance under rotations and dilations justifies the use of the CWT for detecting rotation–dilation (inflation) properties of several classes of 2-D patterns, for instance, Penrose tilings of the plane or diffraction patterns of quasicrystals. We will discuss this recent application in Chapter 4, Section 4.5. As a final remark, we emphasize that, as in the 1-D case, the choice of the normalization factor a −1 in (2.8) or (2.12) is not essential and is made mainly for mathematical reasons. It is the only one that makes the dilation operator Da , and thus the wavelet transform, unitary: ψb,a,θ 2 = ψ2 and Wψ s2 = s2 , as stated in Proposition 2.2.1. In practice, one often uses instead a factor a −2 , so as to enhance the high-frequency part of the signal, and thus to make more conspicuous its singularities, if any. This amounts to introducing, instead of the unitary Da , a nonunitary dilation operator D a , which preserves the L 1 -norm of the signal: D a ψ1 = ψ1 and ψ(b,a,θ ) 1 = ψ1 , where a −2 −1 −1 = a ψb,a,θ = Tb D Rθ ψ = a ψ(a r−θ ( x − b)) . Correspondingly, one deψ(b,a,θ) 1 fines the L -normalized transform ˘ b, a, θ ) = ψ(b,a,θ) |s. S(

(2.40)

This transform is also useful for making contact with the so-called dyadic wavelet transform (see Section 2.4.4), in particular, for the design of fast algorithms, using the continuous wavelet packets developed in Section 2.6. We will meet it again in Chapter 9, while extending the CWT to the 2-sphere.

2.3

Implementation and interpretation of the 2-D CWT

2.3.1

Interpretation of the CWT as a singularity scanner In order to get a physical interpretation of the CWT, we notice that in signal analysis, as in classical electromagnetism, the L 2 norm is interpreted as the total energy of the a, θ)|2 as the energy signal. Therefore, the relation (2.24) suggests we interpret |S(b, density in the wavelet parameter space [284]. Assume now, as in 1-D, that the wavelet ψ is fairly well localized both in position Then so is the transformed wavelet ψb,a,θ space ( x ) and in spatial frequency space (k). ,

42

The 2-D continuous wavelet transform

rotated by θ and dilated by a. Because with effective support suitably translated by b, (2.19) is essentially a convolution with a function ψ of zero mean, the transform a, θ ) is appreciable only in those regions of parameter space (b, a, θ) where the S(b, signal is. Thus we get an appreciable value of S only where the wavelet ψb,a,θ “matches” the features of the signal s. In other words, the CWT acts on a signal as a local filter in a, θ: S(b, a, θ) sees only that portion of the signal that lives around all four variables b, a, θ and filters out the rest. Therefore, if the wavelet is well localized, the energy b, density of the transform will be concentrated on the significant parts of the signal. This is the key to all the approximation schemes that make wavelets such an efficient tool. In order to clarify the filtering effect in scale and angle variables, we rewrite the expression (2.20) of the CWT in polar coordinates k = (ρ, φ): s (k) S(b, a, θ ) = a d 2 k ei b·k ψ(ar (2.41) −θ (k)) 2 R 2π ∞ φ − θ ) s (ρ, φ). (2.42) ρ dρ dφ eibρ cos φ ψ(aρ, =a 0

0

On the last relation, we see that the CWT amounts to a convolution in the scaleangle variables ρ, φ. In order to better appreciate this, we switch throughout to polar coordinates, so the Fourier transform turns into a Mellin transform, as seen in (2.29). In position space, write (2.19) as 1 S(b, a, θ ) = d 2 x ψ(r−θ (a −1 x)) sb ( x ), a R2 In polar coordinates x = (r, ϕ), we get x ) ≡ s( x + b). where sb ( ∞ 2π dr dϕ a −1r ψ(a −1r, ϕ − θ) sb (r, ϕ), S(b, a, θ ) = 0

(2.43)

0

or, in conjugate variables [see (2.29)], ∞ ∞ 1 inθ n) S(b, a, θ ) = e dν a iν ψ(ν, sb (ν, n). 2π n=−∞ −∞

(2.44)

Performing the change of variables r = eu , a = ev , we obtain ev , θ ) = S(b,

(2.45)

∞

du −∞

0

2π

dϕ eu−v ψ(eu−v , ϕ − θ) eu sb (eu , ϕ).

(2.46)

Upon introducing the functions u, ϕ) = eu sb (eu , ϕ), G(u, ϕ) = ψ(eu , ϕ), F(b,

(2.47)

43

2.3 Implementation and interpretation of the 2-D CWT

(2.46) turns into

ev , θ ) = S(b,

∞

du −∞

2π

u, ϕ). dϕ G(u − v, ϕ − θ) F(b,

(2.48)

0

a, θ ) reduces into that of computing a convolution Thus, the problem of computing S(b, of the signal F with a function G of zero mean, hence the filtering effect – and the rapidity of the algorithm [286]. to be as Let us make more precise the support properties of ψ. Assume ψ and ψ well localized as possible (but in a way still compatible with the Fourier uncertainty property), namely, ψ has for numerical support (i.e., the region outside of which the while function is numerically negligible) a “disk” of diameter T , centered around 0, ψ has for numerical support a “disk” of diameter , centered around ko . Then, for the transformed wavelets ψb,a,θ and ψ we have, respectively: b,a,θ r supp ψ is a “disk” of diameter aT centered around b and rotated by an angle b,a,θ θ; r supp ψ is a “disk” of diameter /a, centered around rθ (ko )/a and rotated by b,a,θ θ. Notice that the product of the two diameters is constant (we know it has to be bounded below by a fixed constant, by Fourier’s theorem). These support properties are illustrated in Figure 2.1 for an elliptic shape. As a consequence, we may characterize the filter properties of the wavelet. r If a ! 1, ψ is a wide window, whereas ψ is very peaked around a small b,a,θ b,a,θ spatial frequency rθ (ko )/a: this transform will be most sensitive to low spatial frequencies. r If a 1, ψ is a narrow window and ψ is wide and centered around a high b,a,θ b,a,θ spatial frequency rθ (ko )/a: this wavelet has a good localization capability in the space domain and is mostly sensitive to high spatial frequencies. Thus wavelet analysis operates at constant relative bandwidth, k/k = const, where Therefore, the analysis is most efficient at high spatial frequencies or small k ≡ |k|. scales, and so it is particularly apt at detecting discontinuities in images, either point singularities (contours, corners) or directional features (edges, segments). In other words, the CWT is a singularity scanner. Actually, we will see later that it is also a singularity analyzer. For instance, the CWT allows one to measure fractal dimensions in images (see Chapter 5, Section 5.4). In addition to these localization properties, one often imposes on the analyzing wavelet ψ a number of additional properties, for instance, restrictions on the support Or ψ may be required, as in the 1-D case, to have a certain number of ψ and of ψ. of vanishing moments, up to order N 1 (by the admissibility condition (2.17), the moment of order 0 must always vanish):

The 2-D continuous wavelet transform

(b)

(a)

............ ..... ........ ... .... ... .. .. .. ... .. .. ... ... ... ... ... .. ... ... .. .. .. . . .. . . ... ... ... .... ... ........ ........... ....

✻

.. ..................... ............ ... .......... ........ .. ....... . ...... . . . . ... . .. .. ... ... .. . . . . . . .... ... .. ... .. ... .... ... ... .... . .. ..... .. ...... ..... ....... ...... ... . . . . . . . ..... .... .................................

✻

❅ I ❅

✻

❅

T

✟✟ b

✲

✟

✯ ✟✟

✟

✟

θ

❅ aT ❅ ❘ ❅ ✲

❄

(c)

✻

✒ ✒ /a θ✠

................................................. ........... ...... ...... .... .... .. .... .... ... o ....... .... . . . . . . ............. . ............................................

✲ k

✛

(d)

✻ .......................... .... ...... .... ... ... ... ... .... ... .... .. ...... . ........ ..................

44

✲

✲

Fig. 2.1. Support properties under the basic operations: (top) in the time domain: (a) the original

signal ψ( x ); (b) the modified signal ψb,a,θ ( x ), with b = (2.4, 1.2), a = 1.5, θ = 45◦ ; (bottom) in k); (d) the modified signal ψb,a,θ the frequency domain: (c) the original signal ψ( (k).

x ) = 0, d 2 x x α y β ψ(

x = (x, y),

0 α + β N.

(2.49)

This property improves its efficiency at detecting singularities in the signal. Indeed, the transform (2.19) is then blind to the smoothest part of the signal, that which is polynomial of degree up to N – and less interesting, in general. Only the sharper part remains, including all singularities (jumps in the signal or one of its derivatives, for instance). Equivalently, ψ detects singularities in the (N + 1)th derivatives of the signal [264]. For instance, if the first moments (N = 1) vanish, the transform will erase any linear trend in the signal, such as a linear gradient of luminosity. Conversely, if the signal is rough, a fortiori if it is a measure (as in the analysis of fractals [43,221]), it is sufficient to take a wavelet with no nontrivial vanishing moment, i.e., no condition has to be imposed beyond (2.17). Altogether, as in the 1-D case, the 2-D wavelet transform may be interpreted as a mathematical, direction selective, microscope, with optics ψ, magnification 1/a and orientation tuning parameter θ [Arn95,44]. Two features must be emphasized here. First, the magnification (zoom) 1/a is global, independently of the direction, because we have excluded distortions of the image. Then, there is the additional property of directivity, given by the rotation angle θ. This last feature opens the way to a whole new

45

2.3 Implementation and interpretation of the 2-D CWT

class of applications, in which directions play an essential role. We will detail some of them in Chapters 4 and 5.

2.3.2

The CWT as a phase space representation In order to get a better insight, it is worth recasting the basic formulas (2.18)–(2.20) into a different form. First, we notice that the CWT is in fact a phase space representation (in the usual sense of Hamiltonian mechanics). To see this in a simple way, we observe onto itself. that the correspondence k ⇔ (a −1 , θ) is a bijection from R2∗ ≡ R2 \ {0} Thus, writing κ = a −1 and p ≡ (a −1 , θ) = (κ, θ) ∈ R2∗ , we get da dθ = κ dκ dθ = d 2 p, a3

(2.50)

so that the measure on G becomes simply the volume element of R2 × R2∗ : da dθ = d 2 b d 2 p. (2.51) a3 Thus, the full four-dimensional parameter space of the 2-D WT, G, may be interpreted as phase space, with q ≡ b the position variable and the pair (a −1 , θ) ≡ (κ, θ) playing the rˆole of spatial frequency p, expressed in polar coordinates. The same result holds in the 1-D case [122,259]: a −1 defines the frequency scale, so that the full parameter space of the 1-D WT, the time-scale half plane, is in fact a time–frequency space, thus a phase space. Of course, this interpretation is borne out by mathematical analysis (see Chapter 7). The variable p follows also the common practice in image processing: a = 0 is the horizon in spatial frequency plots, corresponding to extremely high frequencies. It is amusing to note that the same interpretation is even supported by some physiological evidence, namely the so-called orientation hypercolumns of Hubel and Wiesel [DeV88,Duv91,226]. In certain species, cortical neurons are organized into columns, whose sensitivity to position, orientation, and frequency variables correspond exactly to the geometry of R2 × R2∗ just described. In order to manifest the fact that the CWT is really a phase space realization of the signal, we express it explicitly into phase space variables ( q , p). For any vector x = (x, y) = (r cos ϕ, r sin ϕ), with polar coordinates (r, ϕ), define the matrix x −y s(x ) = = r rϕ (rϕ is the 2 × 2 rotation matrix). (2.52) y x d 2 b

x )z = s(z ) x and rθ ( x ) = s( x )eθ , where eθ denotes a unit One shows immediately that s( vector in the direction θ. Then one has = s( p)−1 k = s(k) R p , ar−θ (k) (2.53) | p |2 where R denotes the reflection with respect to the x-axis. We come back now to the expression (2.20) of the CWT and rewrite it in terms of the phase space variables ( q , p):

46

The 2-D continuous wavelet transform

a, θ ) ≡ S(b, S( q , p) = | p |−1

R2

s (k). d 2 k ei q·k ψ( s( p)−1 k)

(2.54)

Alternatively, one may consider the “inverse” phase space variables v = (a, −θ ) ≡ R p/| p |2 . Clearly s( p)s( v ) = I. Although the variable v is less natural, its use some = times simplifies the computations (see Section 5.3, for instance). Since ar−θ (k) v , one gets for the CWT s(v )k = s(k) ˇ s (k). S(b, a, θ ) ≡ S(b, v) = | v| d 2 k ei b·k ψ( s(v )k) (2.55) R2

2.3.3

Visualization of the CWT: the various representations In practice, once the CWT of a given signal s( x ) has been computed, one immediately a, θ) is a function of four variables: faces a problem of visualization. Indeed, S(b, 2 two position variables b = (bx , b y ) ∈ R2 , and the pair (a, θ ) ∈ R+ ∗ × [0, 2π) R∗ −1 (equivalently, (a , θ)). Now, to compute and visualize the full CWT in all four variables is hardly possible. Therefore, in order to obtain a manageable tool, some of the variables, a, θ, bx , b y must be eliminated. There are two ways of achieving this. The first one consists in fixing the value of some of the variables. In other words, one must restrict oneself to a section of the parameter space. Of course, this makes sense only if the variables in question may take arbitrary values in a continuous range. Alternatively, one may integrate out the variables in question. Using the proper part of the natural measure a −3 d 2 b da dθ , one obtains in this way partial energy densities, a, θ )|2 over all the variables is interpreted as the total energy since the integral of |S(b, of the signal, as results from (2.24). This procedure turns out to be crucial whenever the relevant values of the variables to be eliminated (typically, the scale variable a) take only discrete values. We will see an illuminating example of the difference between the two approaches in the problem of symmetry detection in patterns, discussed in Section 4.5. Let us treat first the problem of sections, the partial energy densities will be discussed in detail in Section 2.3.4. In general, one considers two- and three-dimensional sections. While there are six possible choices of 2-D sections, the geometrical considerations made above indicate that two of them are more natural. Either (a, θ) or (bx , b y ) are fixed, and the WT is treated as a function of the two remaining variables. The corresponding representations have the following characteristics [13,19]. (1) The position representation: a and θ are fixed and the CWT is considered as a function of position b alone (this amounts to take a set of snapshots, one for each value of (a, θ ), which may then be collected together into a video sequence). The position representation is the standard one, and it is useful for the general purposes of image processing: detection of position, shape and contours of objects; pattern recognition; image filtering by resynthesis after elimination of unwanted features α), in (noise, for instance). Alternatively, one may use polar coordinates, b = (|b|,

47

2.3 Implementation and interpretation of the 2-D CWT

and aspect or perception angle which case the variables are interpreted as range |b| α, another familiar representation of images. the CWT is considered as a function of (2) The scale-angle representation: for fixed b, scale a and anisotropy angle θ, i.e., of spatial frequency. In other words, one looks and observes all scales and all at the full CWT as through a keyhole located at b, directions at once. The scale-angle representation will be particularly interesting whenever scaling behavior (as in fractals) or angular selection is important, in particular, when directional wavelets are used. Clearly, these two representations are complementary, together they provide the full information contained in the signal. Accordingly, both are needed for a full understanding of the properties of the CWT in all four variables, as demonstrated in [13]. In addition to these two familiar representations, there are four other two-dimensional α, a, θ), and analyzing the sections, obtained by fixing two of the four variables (|b|, CWT as a function of the remaining two. and anisotropy angle (3) The scale-perception angle representation: for fixed range |b| θ , one obtains an analysis at all scales a and all perception angles α. (4) The range-anisotropy angle representation: one fixes the scale a and the perception and all anisotropy angles θ. angle α. This gives an analysis at all ranges |b| (5) The scale-range representation: fixing the perception angle α and the anisotropy angle θ gives an analysis at all scales a and all ranges |b|. and the (6) The angle-angle representation: on the contrary, if one fixes the range |b| scale a, one gets an analysis at all perception angles α and all anisotropy angles θ . This case is particularly interesting, because the parameter space is now compact (it is a torus) and the discretization easy (linear) in both variables. This representation will be used in Section 3.4, for illustrating the difference in angular selectivity between two standard wavelets. For the numerical evaluation, in particular for exploiting the reconstruction formula (2.26), one has to discretize the CWT. In any of these representations, a systematic use of the FFT algorithm will lead to a numerical complexity of 3N1 N2 log2 (N1 N2 ), where N1 , N2 denote the number of sampling points in the two free variables. In the case of the position representation, where (bx , b y ) are free, the geometry is Cartesian and a square lattice will give an adequate sampling grid. In the scale-angle representation, the CWT is naturally expressed in polar coordinates, like (a, θ ) or (a −1 , θ), and the discretization must be logarithmic in the scale variable a and linear in the anisotropy angle θ. For each variable, the size of the sampling mesh may be estimated from the support properties of the reproducing kernel K , which plays the rˆole of a correlation length. We shall come back to this discussion in Section 3.4 of Chapter 3 (see also [13] and [18]). Similar considerations apply to the remaining four representations. In addition, one may also consider three-dimensional sections, for which a single variable is fixed. Two of them look promising for applications.

48

The 2-D continuous wavelet transform

(1) The position-scale representation: suppose the anisotropy angle θ is fixed, or that it is irrelevant, because the wavelet is rotation invariant. Then the transform is a function of position b and scale a. This representation is optimal for detecting the presence of coherent structures, that is, structures that survive through a whole range of scales. Examples may be found, for instance, in astrophysics (hierarchical structure of galaxy clusters and superclusters) [343] or in the analysis of turbulence in fluid dynamics [164,165]. Further information on these two topics will be found in Chapter 5. (2) The position-anisotropy representation: here the scale a is fixed, and the transform is viewed as a function of position b and anisotropy angle θ. If the latter is plotted on the vertical axis of a three-dimensional graph, this means that the plane θ = θo selects all features (targets) that live in the corresponding line of sight. Similarly, an angular sector of opening θ is represented in such a plot by a horizontal slice of thickness θ . This visualization may offer distinct advantages over the conventional ones.

2.3.4

Partial energy densities of the CWT As explained in the previous section, the visualization problem of the CWT is solved by eliminating a certain number of variables, either by fixing their values, or by integrating them out. The principal example is the scale variable a. If the signal has significant features for a discrete set of scales only, {a j , j ∈ J }, the corresponding properties will be visible only if one chooses one of these values a j . Otherwise, nothing will be seen, and the transform is useless. A typical example is the problem of dilation symmetry in patterns, discussed in Section 4.5. In such a situation, clearly one should not fix the scale variable, but integrate over all scales (exactly as in the construction of continuous wavelet packets discussed in Section 2.6). Of course, the measure to use is the dilation invariant one, that is, the scale part a −3 da of the natural measure of the parameter space G. Proceeding in this way with the squared modulus of the wavelet transform, one obtains a quantity which has the physical meaning of a partial energy density. The same reasoning applies to any combination of “ignorable” variables. Thus, one gets such a partial energy density for each of the representations described in the previous section. The most important ones, of course, are the following. (1) Position (or range and aspect) energy density ao , θo ) is In the position representation, a = ao and θo are fixed and the CWT S(b, considered as a function of position b alone, either in Cartesian coordinates bx , b y , α (range and aspect). Accordingly, if one integrates the or in polar coordinates |b|, a, θ)|2 over all scales and orientations, one obtains phase space energy density |S(b, the position energy density, ∞ da 2π a, θ )|2 , dθ |S(b, (2.56) P[s](b) = a3 0 0

49

2.3 Implementation and interpretation of the 2-D CWT

either in Cartesian coordinates (position) or in polar coordinates (range and aspect). This density has been used as the basis of a CWT-based algorithm for automatic detection and recognition of targets (ATR) in forward-looking infrared radar (FLIR) imagery [285]. This application will be discussed in Section 4.2.2. (2) Scale-angle energy density In the scale-angle representation, the CWT is looked at from a fixed position bo as a function of scale a and anisotropy angle θ , i.e., of spatial frequency. In the phase space language of Section 2.3.2, this means considering | S( qo , p)|2 as a function of p alone, for fixed qo . The corresponding partial energy density is obtained by integrating over all positions b or q: a, θ)|2 . d 2 b |S(b, (2.57) M[s](a, θ ) = R2

This energy density, called the scale-angle measure or, better, the scale-angle spectrum of the signal, may be used, for example, for discriminating objects of interest according to their size and orientation [16], in particular target classification in FLIR imagery [287]. It yields also an efficient technique for detecting symmetries, even local ones, in patterns such as quasicrystals or Penrose tilings, the rationale being that such objects have no exact translation invariance, so that any dependence on position variables must be eliminated [24]. Both of these applications will be discussed in Chapter 4. A related concept, introduced in [249], is the relative scale-angle spectrum of the signal, obtained by normalizing M[s] over all angles: Z[s](a, θ ) = 2π 0

M[s](a, θ) dθ M[s](a, θ)

.

(2.58)

Whereas M[s] gives the distribution of energy at different scales and directions, Z[s] gives the relative distribution of energy at different directions at a particular scale with respect to the total energy at that scale. It turns out that Z[s] reveals more efficiently the scale-space anisotropic behavior of the signal. Similar partial energy densities may be introduced for the other representations, corresponding to other choices of “ignorable” variables. For instance, one may write the anisotropy angle and aspect energy density as ∞ da ∞ A[s](α, θ ) = | p | d| p | |S(| p |, α, a, θ )|2 . (2.59) 3 a 0 0 Altogether, there are four one-dimensional partial energy densities, six two-dimensional and M[s](a, θ ), and four three-dimensional ones. Besides the two main ones P[s](b) only two have found an application so far. One is the angular spectrum (called angular measure in [24]): ∞ da α[s](θ ) = M[s](a, θ), (2.60) a3 0

50

The 2-D continuous wavelet transform

which is used for detecting discrete rotation invariances of patterns (see Section 4.5). The other one is the wavelet spectrum, or scale spectrum, a function of scale only, obtained by integrating the scale-angle spectrum M[s](a, θ ) over all angles, or, more often, taking the latter in a rotation-invariant situation: a)|2 . W[s](a) = d 2 b |S(b, (2.61) R2

This is the proper generalization to the wavelet setup of the familiar Fourier power spectrum. It has been used, under various names (wavelet spectrum, wavelet (auto-)power spectrum, wavelet variance, scalogram), by many authors both in 1-D [211,227,273] and in 2-D [175].

2.3.5

Ridges in the 2-D CWT As we saw in Chapter 1 for the 1-D case, the reproduction property (2.36) means that the a, θ) is highly redundant. As a consequence, we information contained in the WT S(b, might hope that no content will be lost if we restrict the WT to a subset of the parameter space. As in 1-D again, there are two ways of achieving this. The first possibility is to take as determining subset the regions where the energy of the signal is concentrated, that is, essentially the lines of local maxima or ridges, the set of all ridges being called again the skeleton of the WT. This we will do in the present section. The alternative is to choose a discrete subset of the parameter space (a lattice), and this leads to the theory of frames that we shall discuss in detail in Section 2.4. When trying to extend the notion of ridge to 2-D signals, one faces again the two extreme situations described for the 1-D case in Section 1.4. We begin with the vertical a, θ ). The square modulus ridges. Let s( x ) be a 2-D signal (an image), with CWT S(b, of the latter is to be interpreted, as we have seen already, as the energy density of the signal, that we shall denote a, θ) = |S(b, a, θ)|2 . E[s](b,

(2.62)

In the case of a rotation-invariant wavelet, the θ-dependence drops out, and we write a) = |S(b, a)|2 . This is the case we will meet in applications, in Chapter simply E[s](b, 5, for instance, in the analysis of astrophysical images (Section 5.1). a) [272]. In this situation, we define the ridges as the lines of local maxima of E[s](b, More precisely, we will define a (vertical) ridge R as a 3-D curve ( r (a), a) such that, + for each scale a ∈ R , E[s]( r (a), a) is locally maximum in space and r is a continuous function of scale. Figure 2.2 gives a concrete example. The signal (left panel) is a set of singularities in a smooth background, simulating a set of bright points on the surface of the Sun and modeled by a random distribution of Gaussians of small (but random) width. The corresponding vertical ridges of the CWT of that signal are shown on the right panel, clearly each ridge points towards a singularity.

51

2.3 Implementation and interpretation of the 2-D CWT

Fig. 2.2. An example of a 2-D ridge: (left) the signal: a field of singularities, simulating a set of bright points on the surface of the Sun; (right) the corresponding vertical ridges of the CWT of that signal.

Given such a vertical ridge, one may distinguish three characteristic features. The first one is the amplitude of the ridge, that is, the value of E[s] on the ridge when a tends to zero, AR = lim E[s]( r (a), a).

(2.63)

a→0

The second one is the slope order, or slope, of E[s] on the ridge when a is close to 0: d ln E[s]( r (a), a) . a→0 d ln a

SR = lim

(2.64)

The last feature is the ridge energy, that is, the integral of E[s] along the ridge, assuming the latter to have a finite length, corresponding to the scale interval [0, amax ]: amax da E[s]( r (a), a). (2.65) ER = a3 0 Here the measure da/a 3 follows from the L 2 normalization of the wavelet. If one uses the L 1 normalization (see Section 2.6.1), one gets instead da/a. As in 1-D, one can show that the restriction of the CWT to its skeleton characterizes the signal completely [265]. A more sophisticated definition of ridge has been introduced by Mallat and Hwang [262], extending to 2-D the WTMM representation described in Section 1.4. The idea is to consider as (directional) wavelets the partial derivatives of a smoothing function ζ ( x ), a Gaussian, for instance: ψ1 ( x ) = ∂x ζ ( x ),

ψ2 ( x ) = ∂ y ζ ( x ),

x ≡ (x, y).

52

The 2-D continuous wavelet transform

Then, given any function s ∈ L 2 (R2 ), its WT with respect to ψ1 and ψ2 can be expressed as a vector b, a) = ∇(S a)), ζ (b, S( where a) = a −2 Sζ (b,

R2

(2.66)

s( d 2 x ζ (a −1 ( x − b)) x)

(using the L 1 -normalization). At a given scale a, the WTMM are defined by the positions b, a)| is locally maximum in the direction of the gradient b where the WT modulus | S( b, a). It turns out that the WTMM lie on connected chains in the b plane. vector S( One then defines the WTMM maxima (WTMMM) as the local maxima of the modulus b, a)| along the WTMM chains. Globally, these WTMMM live along connected | S( lines across scales, that is, the (vertical) ridges. We will describe in Section 5.4.1 the application of this new concept of ridge to the analysis of fractal surfaces. Here again, one can reconstruct from the ridges a good approximation of the original image [262]. An alternate possibility is to introduce horizontal ridges, as in [21]. The algorithm applies to signals which are superpositions of terms of the type s( x ) = A( x ) eiφ(x ) ,

(2.67)

where the amplitude A( x ) varies slowly with respect to the phase φ( x ). A first approximation of the CWT of this signal with the wavelet ψ is obtained by a Taylor expansion, which yields

−θ (∇φ( + R(b, a, θ), a, θ ) = A(b) eiφ(b) ψ(ar b))) S(b,

(2.68)

where the last term is a remainder that can be estimated. Assuming that the wavelet ψ has a unique maximum at a given frequency k0 , one sees that the CWT is concentrated along a surface in the parameter space, the corresponding ridge (more precisely, the 2-D ridge is a vector field, as we shall see below). A mathematically more precise approximation may be obtained by a stationary phase argument, following [195,196]. Take again the signal (2.67) and write the wavelet as ψ( x ) = Aψ ( x ) eiφψ (x ) . Then the CWT (2.19) of the signal s reads −1 ( x) ei b,a,θ d 2 x A( x ) Aψ (a −1 r−θ ( x − b)) , S(b, a, θ ) = a

(2.69)

R2

where ( x ) = φ( x ) − φψ (a −1 r−θ ( x − b)). b,a,θ

(2.70)

53

2.3 Implementation and interpretation of the 2-D CWT

This is an oscillatory integral and, therefore, the main contribution comes from the a, θ) such that stationary points xs of the phase b,a,θ , that is, the points xs = xs (b, ( xs ) = 0, or ∇ b,a,θ xs ). ψ (a −1 r−θ ( xs ) = a −1rθ (∇φ x − b)))( ∇φ(

(2.71)

to second order then yields Expanding the phase b,a,θ C(b, a, θ)−1 + R(b, a, θ ), a, θ ) = s( xs − b)) S(b, xs ) ψ(a −1 r−θ (

(2.72)

a, θ) is a correction term depending on the phase and R(b, a, θ) is again where C(b, a remainder that can be estimated. Because of the support properties of the wavelet ψ, a, θ), if we see from (2.72) that the CWT is essentially localized around the points (b, a, θ) = b. The set of these points is, by definition, the (horizontal) any, such that xs (b, ridge. Thus, on the latter, we have = a −1rθ (∇φ ψ (0)). b) ∇φ(

(2.73)

Writing this relation as ψ (0)), x ) = ar ( x )−1rθr (x ) (∇φ ∇φ(

(2.74)

of polar coordinates we see that the ridge takes the form of a vector field kr = kr (b) −1 (ar (b) , θr (b)), to be interpreted as a local wave vector. We call again skeleton of the WT the restriction of the latter to the ridge. Moreover, a, θ ) to the ridge it may be shown [196] that the restriction of the correction term C(b, is completely determined by the ridge itself, so that the skeleton of the WT reads as ar (b), θr (b)) = Ss (b,

2π ψ(0) s(b). ar (b), θr (b)) C(b,

(2.75)

It follows that the knowledge of the ridge and the skeleton of the WT is sufficient to characterize the signal s( x ). The discussion extends to a multicomponent signal, i.e., a linear superposition of N terms of the form (2.67). Assuming that the wavelet is sufficiently well localized to prevent any overlap, it will see the ridge corresponding to each term separately, and the skeleton will simply consist of N separate ridges, from which the N components can be extracted and reconstructed individually. Explicit examples may be found in [196]. In general, however, ridges may interact, but this case is much more difficult to handle. Finally, additional information may be obtained, as in 1-D, from the length of the various ridges (short vertical ridges tend to come from noise) and the behavior of the modulus of the CWT along each ridge, as a function of scale. Techniques based on 2-D ridges have been exploited in the problem of texture determination and, in particular, in the “shape from texture” problem, that we will discuss briefly in Section 5.5.

54

The 2-D continuous wavelet transform

2.4

Discretization, frames

2.4.1

Generalities on frames As we saw in Chapter 1 for the 1-D case, the reproduction property (2.36) means that the a, θ) is highly redundant. As a consequence, we information contained in the WT S(b, might hope that no content will be lost if we restrict the WT to a subset of the parameter space, in particular, a discrete subset (for instance, a lattice). Then the integral is replaced by a sum over a discrete (but infinite) family of wavelets ψbm ,a j ,θl : s( x) =

ψbm ,a j ,θl ( x ) S(bm , a j , θl ).

(2.76)

m jl

Here too, and by the same reasoning, one is thus led to the introduction of frames. Whereas we have barely sketched this topic in Chapter 1, we will now give a fairly detailed treatment. Further information (albeit mostly in 1-D) may be found in [121, 122,Dau92]. Let us start with a precise definition. According to the terminology introduced by Duffin and Schaefer [156] in the context of nonharmonic Fourier series, one has: Definition 2.4.1 . A countable family of vectors {ψn } in a Hilbert space H is called a (discrete) frame if there are two positive constants A, B, with 0 < A B < ∞, such that A f 2

∞

|ψn | f |2 B f 2 , ∀ f ∈ H.

(2.77)

n=1

The two constants A, B are the frame bounds. If A = B > 1, the frame is said to be tight. If A = B = 1, and ψn = 1, ∀ n, the set {ψn } is simply an orthonormal basis. The properties of a frame are best discussed in terms of the frame operator F : H → , defined by 2

F : f → {ψn | f }. As discussed in Section 1.3, the upper bound in (2.77) simply means that F is a bounded operator, whereas the left inequality guarantees the numerical stability for the recovery of the signal f from its frame coefficients {ψn | f } – in other words, it gives an estimation of the inverse operator F −1 . As for the frame bounds A, B, they measure the redundancy of the representation of the signal in terms of its coefficients. For a tight frame, in particular, A = B > 1 means that the frame is redundant, and B is its index of redundancy.

55

2.4 Discretization, frames

In terms of the frame operator F, the inequalities (2.77) may be written as A I F ∗ F B I,

(2.78)

where I is the identity operator. This in turn implies that F ∗ F is invertible and B −1 I (F ∗ F)−1 A−1 I.

(2.79)

Define now, for each n ∈ N : n = (F ∗ F)−1 ψn , ψ

(2.80)

n . Then the following is true: so that ψn = F ∗ F ψ n } constitute a frame, with frame bounds B −1 , A−1 and Theorem 2.4.2 . The vectors {ψ = F(F ∗ F)−1 . In addition, the expansion frame operator F f (x) =

∞

n (x), ψn | f ψ

(2.81)

n=1

∗ F = I . converges strongly in H, that is, F ≡ {ψ n } is the frame described in the statement results from the equalities Proof. That F n | f |2 = |ψ |(F ∗ F)−1 ψn | f |2 n

n

=

|ψn |(F ∗ F)−1 f |2

n

= F(F ∗ F)−1 f 2 = f |(F ∗ F)−1 f and the inequalities (2.77). Furthermore ∗ F = (F ∗ F)−1 F ∗ F = I, F that is, (2.81) is an identity.

= ∗ F = F ∗ F In other words, the duality between the two frames may be written as F I or explicitly, n | = n ψn | = I. |ψn ψ |ψ (2.82) n

n

= {ψ n } is called the dual or reciprocal frame of F = {ψn }. Notice that the The frame F dual of a tight frame is again a tight frame. This notion is crucial for applications. In the case of wavelet expansions, it is the basis of the so-called biorthogonal scheme [Dau92], briefly discussed in Section 2.5.2. In practice, orthonormal bases are not always available for representing arbitrary functions, but one may often use instead a good frame. By this, we mean that the expansion (2.81) converges sufficiently fast. How could one estimate the speed of this

56

The 2-D continuous wavelet transform

n = (F ∗ F)−1 ψn . If B and A are close convergence? By (2.81), we need to compute ψ 1 2 to each other, F ∗ F is close to 2 (B + A)I , hence (F ∗ F)−1 is close to B+A I and thus n is close to 2 ψn . Hence we may write ψ B+A f =

2 ψn | f ψn + R f, B+A n

(2.83)

where R=I−

2 F ∗ F. B+A

(2.84)

Hence 2 (I − R)−1 B+A ∞ 2 = Rk . B + A k=0

(F ∗ F)−1 =

(2.85)

The series converges in norm, since, by (2.84), −

B−A B−A I R I, B+A B+A

(2.86)

which implies R

B/A − 1 B−A = < 1. B+A B/A + 1

Therefore the expansion (2.81) converges essentially as a power series in |B/A − 1|. Thus the frame is good if |B/A − 1| 1, in particular if it is tight. To the first order, the expansion (2.81) becomes 2 f = ψn | f ψn . (2.87) B+A n The quantity w(F ) =

B−A B+A

(2.88)

is called the width or the snugness of the frame F. It measures the lack of tightness, since w(F ) = 0 iff the frame F is tight. Notice that a frame and its dual have the same width. In practical applications, the infinite sum in (2.81) or (2.87) will be truncated and the approximate reconstruction so obtained is numerically stable. If the width of the frame is sufficiently small, a few terms will suffice. More details on frames may be found in [121,122,Dau92,220,Mal99]. Let us now come back to the notion of redundancy of a frame, which may be characterized in terms of the frame operator F. Let Ran F ⊂ 2 denote the range of F, that is, the set of sequences F f = (ψn | f ), f ∈ H. First, we remark that the inclusion is

57

2.4 Discretization, frames

strict if the frame vectors {ψn } are linearly dependent. In that case, indeed, there exists

a nonzero vector y = (yn ) ∈ 2 such that n yn ψn = 0. But then, for any f ∈ H, one

has n yn ψn | f = y|F f = 0, that is, y ∈ (Ran F)⊥ = {0}, where (Ran F)⊥ is the orthogonal complement of Ran F in 2 . Moreover, the more redundant the frame, the larger the orthogonal complement (Ran F)⊥ . Since F is injective by the left inequality in (2.77), it may be inverted on its image, but the inverse operator is not uniquely defined, because it remains arbitrary on the complement (Ran F)⊥ . Among all possible inverses, the pseudo-inverse F˘ −1 is defined as the inverse operator that vanishes on (Ran F)⊥ . It is easy to show [Mal99] that it is also the inverse with the smallest norm, and it is given by F˘ −1 = (F ∗ F)−1 F ∗ (thus ∗ . Thus, it is uniquely defined). From the discussion above, it is clear that F˘ −1 ≡ F the pseudo-inverse of the frame operator is always bounded. This explains the rˆole of the lower bound in (2.77) as guaranteeing numerical stability in the computation of the inverse operator F −1 . In addition, we see that the dual frame is built from the pseudo-inverse. One should also notice that there exist alternative inversion techniques, for instance, the conjugate gradient method, that sometimes converge faster than the pure wavelet reconstruction formulas. For more information, see for instance [167]. Another useful notion (already met in the Definition 1.5.1 of a multiresolution analysis, in Section 1.5) is that of a Riesz basis, namely, a frame F of linearly independent vectors. Of course, this definition implies that the corresponding frame operator F maps is also a Riesz basis, and in H onto 2 , Ran F = 2 . It follows that the dual frame F fact, it is biorthogonal to F. To see this, apply (2.82) to a basis vector ψk : n |ψk . ψk = ψn ψ n

n |ψk = δn,k , i.e., Since the vectors {ψn } are linearly independent, this implies that ψ the two bases are biorthogonal. Next, if the vectors {ψn } were not linearly independent,

n = 0. But then, taking the there would exist nonzero numbers {λn } such that n λn ψ inner product with any ψk gives n |ψk = λk , 0= λn ψ n

a contradiction. Further information about Riesz bases and their numerical implementation may be found in [Mal99].

2.4.2

Two-dimensional wavelet frames Thus there remains the question: given a specific wavelet, does it generate a frame? The first problem is how to choose the sampling grid # in an optimal fashion. As in 1-D, one should take into account the geometry of the parameter space, that is, the lattice # must be invariant under discrete sets of dilations, rotations and translations.

58

The 2-D continuous wavelet transform

Note, however, that in practice the sampling points are quite often fixed empirically. For the (a, θ) variables, in particular, they are mostly chosen on the basis of biological considerations or symmetry requirements [126,260,264]. Proceeding as in 1-D, one thus obtains the following natural discretization scheme. r For the dilations, a logarithmic scale a = a λ− j , j ∈ Z, for some λ > 1; here again j 0 we will put a0 = 1. r For the rotations, one subdivides the interval [0, 2π ) uniformly into L pieces, for 0 some natural number L 0 ∈ N, that is, θl = lθ0 , θ0 = 2π , l ∈ Z = {0, ..., L 0 − 1}. L 0 L0 r For the translations, one takes into account the two previous discretizations, putting u m 0 m 1 ), bm ≡ b jlm 0 m 1 = λ− j rlθ0 ( with um 0 m 1 ≡ (m 0 β0 , m 1 β1 ), m 0 , m 1 ∈ Z, β0 , β1 0. Thus the discretization grid reads:

−j

# = #(λ, L 0 , β0 , β1 ) = (λ

2π 2 , l , b jlm 0 m 1 ), ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z . L0 (2.89)

The resulting discretized wavelet transform, which is a map from L 2 (R2 , d 2 x) to l 2 (Z × Z L 0 × Z2 ), reads now: S jlm 0 m 1 ≡ S(bm , λ− j , lθ0 ) = ψ jlm 0 m 1 |s = λj d 2 x ψ(λ j r−lθ0 ( x ) − um 0 ,m 1 ) f ( x) R2 −j d 2 k ei bm0 m1 ·k ψ(λ r−lθ0 (k)) f (k), = λ− j

(2.90) (2.91)

R2

with wavelet coefficients

S jlm 0 m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z2 .

Our task now is to find conditions on the grid #(λ, L 0 , β0 , β1 ), that is, on the parameters λ, L 0 , β0 , β1 , such that the family of wavelets {ψ jlm 0 m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z L 0 × Z2 } is a frame. As in 1-D, the answer is that the 2-D wavelet transform obeys a sampling theorem, that gives a lower bound on the density of sampling points, like the standard Shannon theorem of signal analysis. The following result, first proven in [Mur90], is the exact counterpart of [122; Theorem 2.7], and the proof follows closely [Dau92; Section 3.3.2].

59

2.4 Discretization, frames

Theorem 2.4.3 . Assume the wavelet ψ satisfies the following conditions: L0 ∞

(i) s(λ, L 0 , ψ) = ess inf R2 k∈

=

− j r−lθ0 (k))| 2 |ψ(λ

(2.92)

j=−∞ l=0

ess inf

L0 ∞

)∈[0,λ)×[0,2π) (|k|,θ

p (λ− j |k|, ϕ − lθ 0 )|2 > 0, |ψ

j=−∞ l=0

(2.93) p is the Fourier transform of ψ in polar coordi where k = |k|(cos ϕ, sin ϕ) and ψ nates. (ii) S(λ, L 0 , ψ) = sup

R2 k∈

=

L0 ∞

− j r−lθ0 (k))| 2 |ψ(λ

(2.94)

j=−∞ l=0

sup

L0 ∞

)∈[0,λ)×[0,2π) j=−∞ l=0 (|k|,θ

p (λ− j |k|, ϕ − lθ0 )|2 < ∞. |ψ (2.95)

u |)1+$ α( u ) < ∞, (iii) sup (1 + |

(2.96)

u∈R2

where $ > 0 and α( u ) = sup

L0 ∞

R2 j=−∞ l=0 k∈

− j r−lθ0 (k) − j r−lθ0 (k))|. + u)||ψ(λ |ψ(λ

(2.97)

Then there exist constants β0 c , β1 c > 0 such that: (1) ∀ β0 ∈ (0, β0 c ) , β1 ∈ (0, β1 c ), the family {ψl jm 0 m 1 } associated to (λ, L 0 , β0 , β1 ) is a frame of L 2 (R2 , d 2 x); (2) ∀ δ > 0, there exist β0 ∈ (β0 c , β0 c + δ) , β1 ∈ (β1 c , β1 c + δ) , such that the family {ψl jm 0 m 1 } associated to (λ, L 0 , β0 , β1 ) is not a frame of L 2 (R2 , d 2 x).

Proof . We want to find the conditions on λ, L 0 , β0 , β1 for which there exists 0 < A, B < ∞, such that:

A f 2

L0

∞

l=0 j,m 0 ,m 1 =−∞

|ψ jlm 0 m 1 | f |2 B f 2 .

(2.98)

60

The 2-D continuous wavelet transform

The central term in these inequalities reads K =

|ψ jlm 0 m 1 | f |2

j,l m 0 ,m 1

=

λ

−2 j

2

R2

j,l m 0 ,m 1

d 2 k ei b jlm0 m1 ·(k−k )

d k

R2

− j r−lθ0 (k)) − j r−lθ0 (k )) ψ(λ f (k ) × ψ(λ f (k) =

λ

−2 j

2

d 2 k ei um0 m1 ·λ

−j

d k R2

j,l m 0 ,m 1

k ) r−lθ0 (k−

R2

− j r−lθ0 (k)) − j r−lθ0 (k )) ψ(λ f (k ) × ψ(λ f (k) =

λ

−2 j

2

d 2 k ei um0 m1 ·(k−k )

d k R2

j,l m 0 ,m 1

R2

k) k ) ψ( f (λ j rlθ0 (k )) × ψ( f (λ− j rlθ0 (k)). Using the Poisson formula, ∞

ei um0 m1 ·(k−k ) =

m 0 ,m 1 =−∞

4π 2 β0 β1

∞

δ(k − k − u m 0 m 1 ),

(2.99)

m 0 ,m 1 =−∞

where 2π 2π u m 0 m 1 = (m 0 , m 1 ), β0 β1

(2.100)

we obtain K =

4π 2 k) k − ψ( d 2 k ψ( u m 0 m 1 ) β0 β1 j,l m 0 ,m 1 R2 f (λ j rlθ0 (k − u m 0 m 1 )) × f (λ j rlθ0 (k))

4π 2 − j r−lθ0 (k) − j r−lθ0 (k)) ψ(λ − d 2 k ψ(λ u m 0 m 1 ) = β0 β1 j,l m 0 ,m 1 R2 f (k − λ j rlθ0 ( u m 0 m 1 )). × f (k)

61

2.4 Discretization, frames

Let us split the double sum as K = |P| + Q, where |P| denotes the term with (m 0 , m 1 ) = (0, 0) and Q the rest: K = |P| + Q 4π 2 − j r−lθ0 (k))| 2 | 2 = d 2 k |ψ(λ f (k)| β0 β1 j,l R2 4π 2 − j r−lθ0 (k)) − j r−lθ0 (k) ψ(λ − + d 2 k ψ(λ u m 0 m 1 ) β0 β1 j,l m 0 ,m 1 ∈Z∗ R2 f (k) f (k − λ j rlθ0 ( u m 0 m 1 )). × Then we obtain the following estimates. (1) For the first term, we get immediately: 4π 2 4π 2 2 2 s(λ, L 0 , ψ) f |P| S(λ, L 0 , ψ) f , β0 β1 β0 β1 where s(λ, L 0 , ψ) and S(λ, L 0 , ψ) are defined in (2.92) and (2.94), respectively. (2) For the second term, we obtain, by the Cauchy–Schwarz inequality: 4π 2 − j r−lθ0 (k))| |Q| d 2 k |ψ(λ β0 β1 j,l m 0 ,m 1 ∈Z∗ R2 − j r−lθ0 (k) − | u m 0 m 1 )| | f (k)| f (k − λ j rlθ0 ( u m 0 m 1 ))| × |ψ(λ 4π 2 β0 β1 j,l m 0 ,m 1 ∈Z∗ !1/2 − j r−lθ0 (k) − j r−lθ0 (k))| − 2 d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ | f (k)| R2 − j r−lθ0 (k) − j r−lθ0 (k))| − × d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ R2

| f (k − λ j rlθ0 ( u m 0 m 1 ))|2

1/2

4π 2 β0 β1 j,l m 0 ,m 1 ∈Z∗ !1/2 − j r−lθ0 (k) − j r−lθ0 (k))| − | 2 d 2 k |ψ(λ u m 0 m 1 )| |ψ(λ f (k)| ×

R2

− j r−lθ0 (k))| − j r−lθ0 (k) |ψ(λ + 2 d k |ψ(λ u m 0 m 1 )| | f (k)| 2

R2

!1/2 .

62

The 2-D continuous wavelet transform

Applying Cauchy–Schwarz a second time, to the summation over j, l, then gives: |Q|

4π 2 β0 β1

" " ×

d 2 k

R2

m 0 ,m 1 ∈Z∗

#1/2

− j r−lθ0 (k) − j r−lθ0 (k))| − 2 |ψ(λ u m 0 m 1 )| |ψ(λ | f (k)|

j,l 2

d k R2

|ψ(λ

−j

|ψ(λ r−lθ0 (k))|

−j

#1/2

+ r−lθ0 (k) u m 0 m 1 )| | f (k)|

2

.

j,l

The terms in braces in the integrals are majorized using (2.97), and we get finally 4π 2 2 α( u m 0 m 1 ) α(− um0m1 ) f . (2.101) |Q| β0 β1 m 0 ,m 1 ∈Z∗ Define the quantity E(λ, L 0 , β0 , β1 , ψ) =

α( u m 0 m 1 ) α(− u m 0 m 1 ) =

m 0 ,m 1 ∈Z∗

%2 α( u m 0 m 1 ) .

$

(2.102)

m 0 ,m 1 ∈Z∗

In virtue of the decay condition (2.96) in assumption (iii), the sum over (m 0 , m 1 ) converges. Then, using the inequality |P| − |Q| K ≡ |P| + Q |P| + |Q|, we obtain a lower bound for the left-hand side: ' 4π 2 & 2 f |P| − |Q| (2.103) s(λ, L 0 , ψ) − E(λ, L 0 , β0 , β1 , ψ) β0 β1 and an upper bound for the right-hand side: ' 4π 2 & 2 f . S(λ, L 0 , ψ) + E(λ, L 0 , β0 , β1 , ψ) |P| + |Q| β0 β1

(2.104)

By condition (2.92) in assumption (i), s(λ, L 0 , ψ) is strictly positive, and the decay condition (2.96) implies that lim

(β0 ,β1 )→(0,0)

E(λ, L 0 , β0 , β1 , ψ) = 0.

(2.105)

Therefore, there exists critical values β0 c , β1 c > 0 such that the first term in (2.103) is dominant for all β0 < β0 c , β1 < β1 c . Thus the left-hand side of (2.103) is strictly positive and yields a lower frame bound. This proves the statement. Notice that, as in [Dau92; Section 3.3.2], we had to use in (2.92) the essential infimum, that is, infimum except on a set of measure zero, instead of the usual infimum, which 0) = 0. This distinction is not necessary in (2.94), because in all is 0 here, since ψ( is continuous. Here also, the infimum in (2.93) has to be practical cases, the function ψ can be brought to this range taken only over a ball of radius λ, since other values of |k| j by dilation with a suitable power λ , except for k = 0, which is a set of measure zero.

63

2.4 Discretization, frames

From the inequalities (2.103) and (2.104), we obtain immediately estimates for the frame bounds: Corollary 2.4.4 . Whenever the conditions of Theorem 2.4.3 are satisfied, and β0 , β1 are such that s(λ, L 0 , ψ) > E(λ, L 0 , β0 , β1 , ψ),

(2.106)

then the bounds of the frame can be estimated by: 4π 2 {s(λ, L 0 , ψ) − E(λ, L 0 , β0 , β1 , ψ)}, β0 β1 4π 2 {S(λ, L 0 , ψ) + E(λ, L 0 , β0 , β1 , ψ)}. B β0 β1 A

(2.107) (2.108)

It is clear from this discussion that we have obtained essentially the same results as in the 1-D case – and they extend to three or more dimensions in the same way [283,Mur90]. As argued in [Dau92; Section 3.3.2], the moral is that, whenever the wavelet ψ is admissible and decays reasonably at infinity, it will yield a frame. This is true, for instance, for all types of Mexican hat or Morlet wavelets. However, the estimates of Corollary 2.4.4 imply, as in 1-D [103], the following inequalities as well, for all k = 0, A

L0 ∞ 4π 2 − j r−lθ0 (k))| 2 B. |ψ(λ β0 β1 j=−∞ l=0

(2.109)

This puts a rather strong limitation on the wavelet, because the frame tends to become loose when λ is not very small, in particular for λ = 2, the preferred value. In 1-D, this defect may be corrected by introducing additional voices. This means that one further subdivides each octave, replacing in (2.109) the exponent j by ν j, ν = 0, 1, . . . , N − 1 (N is called the number of voices). The effect is to “densify” N the lattice #, which improves the ration B/A and thus speeds up convergence of the discrete approximation. For further details, we refer the reader to [Dau92]. The same procedure may be applied in 2-D as well. However, if the speed is the determining criterion, one can do better by using the pseudo-QMF algorithm described in Section 2.6. In addition, other discretization schemes may be considered. For instance, the discretization step in the angular variable θ may be made scale dependent. The idea is that one needs more directions at small scales (high frequencies) than at large scales (low frequencies). This is in the spirit of certain applications of discrete wavelet bases in solid state physics (see [30]), in which additional, smaller, scales are considered in the vicinity of the atomic crystal nodes. We will come back to this point in the next section.

64

The 2-D continuous wavelet transform

2.4.3

Implementation of a 2-D wavelet frame In general, implementing the inverse frame operator is a nontrivial task which requires using the expansion formula (2.84). In the case of a tight (or approximately tight) frame, however, the situation becomes much easier since the reconstruction is provided by the simple sum (2.87). Let us particularize this for the concrete case of the 2-D Morlet wavelet. Using Lee’s parametrization [253], the Fourier transform of this wavelet is simply: κ2 κ2 2 2 2 √ − 2 ((k1 −k0 )2 +4k22 ) κ 2 (k1 +4k2 +k0 ) 2k 2k 1 , k2 ) = ψ(k 8π e 0 −e 0 . (2.110) k0 √ Chosing κ = 3 2 log 2 yields a spatial frequency bandwidth of approximately one octave, which is compatible with the characteristics of the receptive fields of human simple cells [372]. Using the general framework described before, one is able to compute frame bounds for this wavelet. Psychovisual experiments suggesting a density of about 16 to 20 orientations, one can obtain a tight frame by carefully choosing a discretization strategy. Figure 2.3 shows a reconstructed image with a tight frame of Morlet wavelets corresponding to the discretization grid 2π # = 2− j/n , l , b j,l,m 0 ,m 1 , ( j, l, m 0 , m 1 ) ∈ Z × Z16 × Z2 , n = 2 and 4, 16 that is, 2 and 4 voices, respectively, per octave and a uniform translation step β0 = β1 = 0.8.

2.4.4

The dyadic wavelet transform As we saw already in 1-D (Sections 1.3 and 1.6.1), a slightly different construction consists in computing a hybrid wavelet transform in which we sample the scale variable, but leave translations untouched. If we use dyadic scales, the resulting set of coefficients is called the Dyadic Wavelet Transform and offers the advantage of being completely covariant with respect to translations, a very desirable feature [264,265]. In pattern recognition, for instance, translating the object as a whole should affect all wavelet coefficients in the same way, so as not to distort the transform. This is precisely the practical meaning of the so-called “shift invariance” (properly, shift covariance), and the condition is not satisfied in the usual discrete WT. In addition, when the translation parameter is properly discretized, such dyadic wavelet transforms may lead to genuine tight frames. We denote, as usual, the scaled and translated wavelets by , ψb,2 x ) = 2− j ψ (2− j ( x − b)) j (

(2.111)

assume for simplicity that cψ = 1 and use the L 2 -normalization. Then the dyadic WT of f ∈ L 2 is given by

65

2.4 Discretization, frames

(a)

(b)

(c)

Fig. 2.3. Reconstruction of the lena image with a tight frame of Morlet wavelets: (a) the original

image; (b) and (c) the reconstructed image, with 2 and 4 voices, respectively, per octave.

2 j ) = ψ j | f = ψ j# % f , W (b, 2 b,2

(2.112)

x ) = ψa (− x ) and ψa ≡ ψ0,a where ψa# ( . Note that we use a different notation, W(., .) instead of W (., .), in order to emphasize the fact that we are dealing with the dyadic WT instead of the CWT. The superscript in equations (2.111), (2.112) refers to different wavelets, but the practical situation will be that of rotated versions of a single generating function. The following result shows that the dyadic WT is a stable and complete representation of images provided the set of scaled wavelets appropriately pave the frequency plane.

66

The 2-D continuous wavelet transform

Proposition 2.4.1 If there exists two strictly positive constants A and B such that A

L

(2 j k)| 2 B, |ψ

∀ k,

(2.113)

=1 j∈Z

then, A f 2

L

2−2 j W (., 2 j )2 B f 2 .

(2.114)

=1 j∈Z

Moreover, if we define dual or reconstruction wavelets χ through the relation L

(2 j k) χ = 1, (2 j k) ψ

, ∀ k ∈ R2 \ {0}

(2.115)

=1 j∈Z

then, the following reconstruction formula holds strongly in L 2 : f ( x) =

L =1 j∈Z

2−2 j W (., 2 j ) % χ2j ( x ).

(2.116)

2 j ) with respect to the variable Proof . Let d j stand for the Fourier transform of W (b, b: (2 j k) = 2j ψ d j (k) f (k). Condition (2.113) yields: 2 A| f (k)|

L

2 B| 2. 2−2 j |d j (k)| f (k)|

=1 j∈Z

Integrating over k and using the Plancherel formula gives (2.114). Now taking the Fourier transform on both sides of (2.116) gives = f (k)

L

(2 j k) ψ χ f (k) (2 j k).

=1 j∈Z

Finally, applying condition (2.115) gives the final result.

The norm equivalence represented by (2.113) shows that this family of wavelets behaves exactly like a frame, with frame operator 2 j , ) = 2− j W (b, 2 j ). (F f )(b, In particular, if the sum standing in the middle of (2.113) is a constant, we can take A = B and we get the analogue of a tight frame.

67

2.4 Discretization, frames

Among all possible inverses of the frame operator F, the particular choice = ψ (k) χ (k) j 2 j∈Z |ψ (2 k)| leads to an attractive solution. In this case, indeed, the inverse frame operator given by (2.116) is the pseudo-inverse of F as shown by the following result: Proposition 2.4.2 Let F be the normalized frame operator u , 2 j , ) = 2− j W ( u , 2 j ). [F f ] ( Then the following left inverse is the pseudo-inverse of F: ( FW u, 2 j ) =

L =1 j∈Z

u ). 2− j W (., 2 j ) % χ2j (

(2.117)

Proof . Let 2 (L 2 ) be the Hilbert space of square integrable sequences of functions d, j =1...L , j∈Z , d, j ∈ L 2 and d, j 2 < +∞.

j∈Z

We are going to show that, ∀ g ∈ L 2 and ∀ d ∈ Ran{F}⊥ , d = 0. g, F

(2.118)

Let d = Fg and d ∈ Ran{F}⊥ . Using (2.117), we compute ( j (k) d = d ψ (2 k) g (k) d 2 k . g, F , j ( j 2 2 j∈Z R j ∈Z |ψ (2 k)| Now, writing d, j = Fg, we also have ( (2 j k) ψ = d g (k) , j (k) , and using equation (2.113) we find 1 g 1 d, d 2 (L 2 ) g, F 2 (L 2 ) . d, d B A Now, since d ∈ Ran{F} and d ∈ Ran{F}⊥ , (2.119) yields the result.

(2.119)

Another interesting case arises when we define a new set of wavelets through the relation (k) ψ =) ϕ (k) ,

j 2 j |ψ (2 k)|

(2.120)

68

The 2-D continuous wavelet transform

and use the same set of functions to implement the reconstruction: f ( x) =

L =1 j∈Z

2−2 j W (., 2 j ) % ϕ2 j ( x ).

(2.121)

The reader can easily check that the wavelet transform operator is an isometry between L 2 and 2 (L 2 ). The range of this operator is a closed subspace V of 2 (L 2 ) characterized by a reproducing kernel, exactly as for the continuous transform. Indeed, any d, j ∈ V satisfies j, j d, j ( x) = d , j % K, ( x) , , j

where the reproducing kernel is given by

j, j x ). x) = ϕ2 j % ϕ2 j ( K, (

(2.122)

This is easily checked by writing explicitly % $ u ). u ) = F F −1 d , j ( d, j ( Taking the Fourier transform with respect to u on both sides and using (2.120) yields the result. This property is a sign of the redundancy of the representation, exactly as in the continuous case. Section 2.6 gives an efficient and automatic way for building such dyadic wavelet transforms starting from the continuous wavelet transform. This offers more flexibility in the design of directional dyadic wavelets.

2.5

Comparison with the 2-D discrete wavelet transform Before analyzing the recent approach to fast algorithms (Section 2.6), we have to sketch briefly the 2-D discrete wavelet transform, in its various forms and generalizations. As mentioned in Chapter 1, a key step in the success of the 1-D discrete WT was the discovery that almost all examples of orthonormal bases of wavelets may be derived from a multiresolution analysis, and furthermore that the whole construction may be translated into the language of digital filters. In the 2-D case, the situation is exactly the same, as we shall see in this section. Further information may be found in [Dau92] or [Mey94].

2.5.1

Multiresolution analysis in 2-D and the 2-D DWT The simplest approach consists in building a 2-D multiresolution analysis simply by taking the direct (tensor) product of two such structures in 1-D, one for the x direction, one for the y direction. If {V j , j ∈ Z} is a multiresolution analysis of L 2 (R), (2) then { V j = V j ⊗ V j , j ∈ Z} is a multiresolution analysis of L 2 (R2 ). Writing again

69

2.5 Comparison with the 2-D discrete wavelet transform (2)

(2)

(2)

V j ⊕ W j = V j+1 , it is easy to see that this 2-D analysis requires one scaling function: (x, y) = φ(x) φ(y), but three wavelets: h (x, y) = φ(x) ψ(y) v (x, y) = ψ(x) φ(y)

(2.123)

(x, y) = ψ(x) ψ(y). d

As the notation suggests, h detects preferentially horizontal edges, that is, discontinuities in the vertical direction, whereas v and d detect vertical and oblique edges, respectively. Indeed, for j = 1, the relation V1 = V0 ⊕ W0 yields: (2)

(y)

V1 = V1(x) ⊗ V1

(y)

(y)

= (V0(x) ⊕ W0(x) ) ⊗ (V0 ⊕ W0 ) (y)

(y)

(y)

(y)

= (V0(x) ⊗ V0 ) ⊕ (V0(x) ⊗ W0 ) ⊕ (W0(x) ⊗ V0 ) ⊕ (W0(x) ⊗ W0 ) (2)

(2)

= V0 ⊕ W0 , (y)

where V0 = V0(x) ⊗ V0 & φ(x) φ(y) and W0 is the direct sum of the three other products, generated by the three wavelets given in (2.123), respectively. (2) From these three wavelets, one gets an orthonormal basis of V j by defining (2) j { kl (x, y) = φ j,k (x) φ j,l (y), k, l ∈ Z}, and one for W j in the same way, namely α, j {kl (x, y), α = h, v, d and k, l ∈ Z}. Clearly this construction enforces a Cartesian geometry, with the horizontal and the vertical directions playing a preferential role. This is natural for certain types of images, such as in television, but is poorly adapted for detecting edges in arbitrary directions. Other solutions are possible, however (see below). As in the 1-D case, the implementation of this construction rests on a pyramidal algorithm introduced by Mallat [259,260]. The technique consists in translating the multiresolution structure into the language of QMFs, and putting suitable constraints on the filter coefficients. For instance, ψ has compact support if only finitely many coefficients differ from zero. In the 2-D case, obviously one gets a low-pass filter h and three high-pass filters g α , α = h, v, d. Thus, a signal f ∈ L 2 (R2 ) is represented at

resolution 2 j by the function f j = a j + α d αj , α = h, v, d, where (2) j cj = c j,kl kl ∈ V j (2)

(2)

k,l∈Z

d αj

=

α, j

α

d αj,kl kl ∈ W j , α = h, v, d. (2)

k,l∈Z

In this scheme, the decomposition of an image into a low resolution approximation, plus three types of details (h, v, d), at successive finer scales, takes the familiar form of nested boxes, with the low resolution part in the upper left corner. Figure 2.4 presents a schematic three-level decomposition of an image into a low resolution approximation, with coefficients c−3 , plus increasingly finer details, of the three types, with coefficients

70

The 2-D continuous wavelet transform

c3

d3vv

d3h

d3d

d2vv d1vv

h

d

d2

d2

d1h

d1d

Fig. 2.4. Schematic three-level decomposition of an image into a low resolution approximation,

plus increasingly finer details, of the three types (h, v, d).

d αj , j = 1, 2, 3, α = h, v, d. Then we show a (standard) real example in Figure 2.5, which is a three-level orthonormal basis decomposition, with a Daubechies wavelet (compact support, three vanishing moments).

2.5.2

Generalizations As in one dimension, the scheme based on orthonormal wavelet bases is too rigid for most applications and various generalizations have been proposed. We discuss some of them here, for two reasons. First, for the sake of completeness. But also in order to demonstrate that some features which are natural in the continuous transform, such as covariance, are not easy to enforce in the discrete case, however desirable they may be.

2.5.2.1

More isotropic 2-D wavelets The tensor product scheme privileges the horizontal and the vertical directions; more isotropic wavelets may be obtained, either by superposition of wavelets with specific orientation tuning [Mar82], as we did above with the CWT, or by choosing a different way of dilating, using a nondiagonal 2-D dilation matrix, which amounts to dilating by a noninteger factor [Dau92]. Consider, for instance, the following dilation matrices: 2 0 1 1 1 1 D0 = , D1 = , D2 = . (2.124) 0 2 1 −1 −1 1 The matrix D0 corresponds to the usual dilation scheme by powers of 2, whereas D1 and D2 lead to the so-called “quincunx” scheme [Fea90]. In the standard scheme, a unit square is dilated, in the transition j → j + 1, to another square, twice bigger, with the same orientation. This means that three kinds of additional details have to be supplied,

71

2.5 Comparison with the 2-D discrete wavelet transform

Fig. 2.5. Typical three-level decomposition of an image into a low resolution approximation, plus

increasingly finer details, of the three types (h, v, d). The basic wavelet is a Daubechies wavelet with compact support and three vanishing moments.

horizontal, vertical and oblique (see Figure 2.6, left). By contrast, the same operation in the “quincunx” scheme leads to a square circumscribed to the original one, that is, √ rotated by 45◦ and larger by a factor 2, so that only one kind of additional detail is necessary (Figure 2.6, right). Indeed only one wavelet is needed in this scheme, instead of three. This is consistent with a result of Meyer, according to which the number of independent wavelets needed in a given multiresolution scheme equals (| det D| − 1), where D is the dilation matrix used.

2.5.2.2

Biorthogonal wavelet bases In the case of the continuous transform, the wavelet used for reconstruction need not be the same as that used for decomposition, they have only to satisfy a cross-compatibility

72

The 2-D continuous wavelet transform

(v)

❅

(d)

❅

j+2

❅

❅

scale j

(h)

scale j

❅

❅

j+1❅

❅

❅

j+1

❅

❅ ❅

Fig. 2.6. Unit cell at successive resolutions: (left) for the “Cartesian” scheme; (right) for the

“quincunx” scheme.

condition, as shown in (2.30), and this has led us to a host of different, more flexible reconstruction formulas, such as (2.31)–(2.34). The same idea in the discrete case leads to biorthogonal bases [108], i.e., one has two hierarchies of approximation spaces, {V j } j }, with cross-orthogonality relations. and {V In 1-D, the construction goes as follows, and the extension to 2-D proceeds as above. Start with a scale of closed subspaces {V j }, assuming only the existence of a scaling function φ ∈ V0 such that its integer translates {φk (x) ≡ φ(x − k), k ∈ Z} is a Riesz (or unconditional) basis of V0 (see Section 1.5). Then, instead of orthogonalizing this basis, which would lead to the construction of an orthonormal wavelet basis, one takes 0 k }, that is, the vectors defined by the relation φk |φ l = δkl . Let V the dual basis {φ denote the closed subspace generated by {φk , k ∈ Z}. Then the same construction is repeated for each j, using the dilation invariance of the scale {V j }. The outcome is j }, with exactly the same properties. Next, for each a second multiresolution scale {V j , and j ∈ Z, one defines a subspace W j by the two conditions W j ⊂ V j+1 and W j ⊥ V j ⊂ V j+1 and W j ⊥ V j . In this way one obtains two sequences of subspaces similarly W j,k , j, k ∈ Z}, respectively, which are {W j } and {W j }, with bases {ψ j,k , j, k ∈ Z}, {ψ mutually orthogonal: j ,k = δ j j δkk . ψ j,k |ψ

(2.125)

In terms of these bases, one gets two types of expansion formulas, for any f ∈ L 2 (R): j,k | f ψ j,k f = ψ j,k∈ Z

=

j,k ψ j,k | f ψ

(2.126)

j,k∈ Z

in the case of a frame and its dual. Translating the whole structure into the language of digital filters, a 1-D biorthogonal scheme corresponds to four filters, (h, h, g, g ), where the first two are low-pass and the last two are high-pass. A corresponding characterization may be given in the 2-D case.

73

2.5 Comparison with the 2-D discrete wavelet transform

The resulting scheme is much more flexible and is probably the most efficient one in practical applications, both in 1-D and in 2-D (and it is widely used). For instance, it gives a better control on the regularity or decrease properties of the wavelets [108].

2.5.2.3

Wavelet packets and the Best Basis Algorithm Electrical engineers are familiar with the notion of subband coding scheme. In a few words, this means that a discrete signal {cn0 , n ∈ Z} is first subdivided into two subsignals {cn1 , dn1 , n ∈ Z}, obtained by convolving the original signal with two filters, one low-pass and one high-pass filter, respectively. Next, each subsignal is subsampled by a factor of 2, that is, one keeps only the even or odd components, respectively (thus the total number of coefficients is unchanged). Then the operation is iterated a number of times. A reconstruction of the original signal is obtained, more or less exactly, by inverting all the operations. In the construction of 1-D orthonormal wavelet bases, each approximation space V j gets further decomposed into V j−1 and W j−1 , whereas the detail space W j is left unmodified. Expanding the signals in the respective bases {φ j−1,k , ψ j−1,k } shows that the construction is in fact a subband coding scheme, but a very special one, rather asymmetrical. In order to get more flexibility, more general subband schemes have been considered, called wavelet packets, where both subspaces V j and W j are decomposed at each step [Mey94,Wic94,110,111]. Such a scheme provides rich collections (“libraries”) of orthonormal bases, each one corresponding to a given decomposition of L 2 (R) into a sequence of mutually orthogonal subspaces, chosen at successive scales j ∈ Z (an example is given below). Enumerating all possible such decompositions or orthonormal bases is a nontrivial combinatorial problem, whose solution stems from graph theory [Wic94,111]. Facing such a plethora of orthonormal bases, one needs a strategy for determining the best basis to use in a given situation. An efficient solution, based on entropic criteria, has been proposed by Coifman et al. [Wic94,111], under the name of the Best Basis Algorithm, and it has become a standard tool in signal analysis. Of course, all this applies verbatim in two dimensions, although the labeling of basis vectors becomes even more intricate. This holds, in particular, for the comments made at the end of Section 1.5 concerning the numerical implementation of finite reconstruction formulas. In order to fix ideas, we show in Figure 2.7 the subband subdivision scheme of the standard wavelet 1-D multiresolution analysis, in the case of a three-level decomposition. This corresponds to the following decomposition into orthogonal subspaces: V0 = W−1 ⊕ W−2 ⊕ W−3 ⊕ V−3 , or, in the notation of Section 1.5 [compare (1.58)], the representation of a signal s ≡ s0 ∈ V0 by its wavelet coefficients s0 = (d−1 , d−2 , d−3 , c−3 ). By comparison, Figure 2.8 shows the modified subdivision scheme used in the wavelet packets formalism, together

74

The 2-D continuous wavelet transform

V0 V−1 V−2 V−3

W−3

W−1 W−2

W−1

W−2

W−1

Fig. 2.7. The 1-D wavelet subband scheme, with a three-level decomposition.

V0 V−1 V−2 V−3

00 W−3

W−1 0 W−2

01 W−3

02 W−3

1 W−2 11 W−3

12 W−3

2 W−2 21 W−3

22 W−3

Fig. 2.8. The 1-D wavelet packet subband scheme, with a particular choice of three-level

decomposition.

with a particular choice of three-level decomposition, namely [compare Figure 1.11 (b) and (c)]: 2 11 12 V0 = V−1 ⊕ W−2 ⊕ W−3 ⊕ W−3 .

In the corresponding orthonormal basis, the signal is represented as s ≡ s0 = (c−1 , dd−2 , ccd−3 , dcd−3 ). The notation proceeds as follows. For j = −1, the coefficients are (c−1 , d−1 ). Then at each step, the coefficient x j is replaced by the pair (cx j−1 , d x j−1 ). Thus, for j = −2, one has, from left to right in Figure 2.7, (cc−2 , dc−2 , cd−2 , dd−2 ); for j = −3, (ccc−3 , dcc−3 , cdc−3 , ddc−3 , ccd−3 , dcd−3 , cdd−3 , ddd−3 ); and so on. For more details on the basis labeling, and wavelet packets in general, we refer to [Wic94] or [110,111].

2.5.2.4

The lifting scheme: second-generation wavelets One can go one step beyond, and abandon the regular dyadic scheme and the Fourier transform altogether. Using the “lifting scheme” leads to the so-called secondgeneration wavelets [349,350], which are essentially custom-designed for any given

75

2.5 Comparison with the 2-D discrete wavelet transform

problem. The starting point is that, in a biorthogonal scheme, one scale {V j } does not j } uniquely, but the freedom left in the generating wavelet determine its counterpart {V is known explicitly, and reduces to an arbitrary trigonometric polynomial. Thus the idea is to start from a given biorthogonal scheme with filters (h, h, g, g ), then tranform (1) (1) (1) (1) it using that freedom into a new one (h , h , g , g ), and so on, by a succession of “lifting steps.” But one must generalize first the very notion of biorthogonal scheme, in order to get a more flexible scheme. To that effect, one weakens Definition 1.5.1 of a multiresolution analysis {V j , j ∈ Z} by replacing the two conditions (1) and (2) (which enforce the scale invariance and thus the dyadic scheme) by the single condition (3) for each j ∈ Z, V j has a Riesz basis {ϕ j,k , k ∈ K( j)}, the elements of which are called scaling functions. Here K( j) is a general index set, which allows irregular sampling (no translation invariance!). One assumes only that K( j) ⊂ K( j + 1), without the dilation relation j }, with dual given by condition (1). In the same way, one considers a dual scale {V scaling functions { ϕ j,k , k ∈ K( j)}, biorthogonal to the previous ones: ϕ j,k | ϕ j,k = δkk , k, k ∈ K( j).

(2.127)

Then the filter h ≡ h j,k,l enters through a refinement equation ϕ j,k = h j,k,l ϕ j+1,l ,

(2.128)

l∈K( j+1)

h are and similarly for h ≡ h j,k,l (all filters are assumed to be finite). The two filters h, then biorthogonal (see (2.131) below). Next one defines wavelets in the usual way. A family of functions {ψ j,m , m ∈ M( j)}, where M( j) = K( j + 1) \ K( j), is a set of wavelet functions if: (i) the space W j = j ; (ii) the set span {ψ j,m , m ∈ M( j)} is a complement of V j in V j+1 , and W j ⊥ V 2 {ψ j,m /ψ j,m , j ∈ Z, m ∈ M( j), } is a Riesz basis for L (R). Dual wavelets are vectors of the biorthogonal basis, j ,m = δ j j δmm . ψ j,m |ψ

(2.129)

j , which complement V j in V j+1 , and W j ⊥ V j . By construction, They span spaces W the wavelets satisfy refinement relations: ψ j,m = g j,m,l ϕ j+1,l , (2.130) l∈K( j+1)

which thus define the filter g ≡ {g j,m,l }, and similarly for the dual filter g. Altogether, the four filters h, h, g, g satisfy biorthogonality relations: g j,m ,l = δmm , g j,m,l = 0 g j,m,l h j,k,l l

l

l

h j,k ,l = δkk , h j,k,l

l

h j,k,l = 0. g j,m,l

(2.131)

76

The 2-D continuous wavelet transform

Similar relations may be written for scaling and wavelet functions. Now, of course, since the index sets K( j) are general, some extra care is required to guarantee the convergence of all the expansions (hence the finiteness condition on the filters), and also the rapidity of the algorithm. We refer to [350] for technical details. This scheme becomes simpler if one introduces an operator notation (familiar in the signal processing literature), as follows. 2 2 r The filter h j,k,l is embodied in the operator H j : (K( j + 1)) → (K( j)), defined 2 2 by b = H j a, where a ≡ (al ) ∈ (K( j + 1)), b ≡ (bk ) ∈ (K( j)), and bk =

h j,k,l al .

l∈K( j+1)

r

The filter g j,m,l is embodied in the operator G j : 2 (K( j + 1)) → 2 (M( j)), and j , G j. similarly for the operators H In this notation, the conditions for exact reconstruction can be written in matrix form as: j j H 1 0 H = and = 1. H j∗ G ∗j H j∗ G ∗j j j 0 1 G G (2.132) Now we may describe the lifting scheme. As already mentioned, the idea is to exploit j biorthogonal to a given one H j , G j . The j , G the freedom in designing a set of filters H freedom is that of an arbitrary operator S j : 2 (M( j)) → 2 (K( j)). In components, this operator is represented by a set of coefficients, S j ≡ s j,k,m . In the usual, first generation wavelet scheme, this in turn would be given by a trigonometric polynomial s(ω). The technique proceeds in two steps. (i) First a lifting step, which consist in passing from a given biorthogonal filter set j , G j , G j } to a new one, {H j , H (1) , G (1) , G j }, where {H j , H j j (1) = H j, j + S j G H j

∗ G (1) j = G j − Sj Hj,

j being unchanged. Hence, the original scaling functions ϕ j,k the two filters H j , G j,m are modified. do not change, but all the other functions ϕ j,k , ψ j,m , ψ (1) , G (1) , G j , } to the set (ii) A dual lifting step, leading from the set {H j , H j j (1) (1) (1) (1) {H j , H j , G j , G j , }, where H j(1) = H j + S j G (1) j ,

(1) . (1) = G j − G S ∗j H j j

(1) , G (1) and S j : 2 (M( j)) → 2 (K( j)) is another operator. Here the two filters H j j remain unchanged.

77

2.5 Comparison with the 2-D discrete wavelet transform

Each of the two steps preserves the biorthogonality of the filter sets, as can be checked easily on the conditions (2.132). Of course, one has to verify along the way that the new scaling and wavelet functions belong correctly to the appropriate spaces (in particular, that they are square integrable). Then it may be shown that any biorthogonal filter set may be obtained by this procedure after a finite number of steps, starting from a trivial set, called the Lazy wavelet, because it does nothing but split the sequences into two subsets of components. j = E, G j = More precisely, the Lazy wavelet corresponds to the filter set H j = H j = D, where E : 2 (K( j + 1)) → 2 (K( j)) and D : 2 (K( j + 1)) → 2 (M( j)) are G simply the restriction or subsampling operators. [In the engineering literature, the lazy wavelet is also called a polyphase filter (of size 2) [Vet95].] The resulting scheme is fast, and independent from translation invariance and from the Fourier transform. Thus it applies to wavelets on intervals or on curves, and in higher dimensions as well, for instance, to wavelets on two-dimensional manifolds [Mal99,349,350]. Here the idea is to start from a succession of finer and finer grids. Consider a 2-D manifold (such as the 2-sphere) and choose a certain grid G( j) on it. Then refine the latter to a grid G( j + 1), of the form G( j + 1) = G( j) ∪ C( j), where C( j) denotes the complement. A typical example is to start from a triangulation of the manifold and refine it by bisecting each side, as illustrated in Figure 2.9 in the case of the sphere. Then the multiresolution spaces are defined as V j+1 = 2 (G( j + 1)), V j = 2 (G( j)), W j = 2 (C( j)), with appropriate bases {ϕ j,k } ∈ V j , {ψ j,m } ∈ W j , and the whole machinery is put into operation. The resulting tool has proven to be extremely versatile and efficient. For instance, Schr¨oder and Sweldens [336] have applied it to the design of wavelets on the sphere, with a very convincing application to the reproduction of coastlines on a terrestrial globe (we will see in Chapter 9, Section 9.2, another approach to the same problem, this one directly based on the CWT). As a final remark, we may point out that the lifting scheme opens the door to nonlinear multiresolution decompositions, such as the median transform of Bijaoui [Sta98] or the

Fig. 2.9. Typical grid refining for applying the 2-D lifting scheme: the geodesic sphere construction, starting with the icosahedron on the left (subdivision level 0) and the next two subdivision levels (from [336]).

78

The 2-D continuous wavelet transform

morphological wavelets introduced by Goutsias and Heijmans [198,219]. The latter are explicitly based on the lifting scheme and make the connection with the standard field of mathematical morphology. We note finally that the lifting scheme by itself offers a very pedagogical entrty into the wavelet world, as examplified in the little volume of Jensen and la Cour-Harbo [Jen01].

2.5.2.5

Integer wavelet transforms In their standard numerical implementation, the classical (discrete) WT converts floating point numbers into floating point numbers. However, in many applications (data transmision from satellites, multimedia), the input data consists of integer values only and one cannot afford to lose information: only lossless compression schemes are allowed. Recent developments, based on the lifting scheme, have produced new methods that allow one to perform all calculations in integer arithmetic [2,92]. In addition, such methods also improve the performances of lossy compression techniques [324].

2.6

Bridging the gap: continuous wavelet packets and fast algorithms

2.6.1

Custom design of dyadic frames Besides the full discretization described in Section 1.3, and the discrete WT just discussed, there is an intermediate procedure, introduced in [159], under the name of infinitesimal multiresolution analysis. It consists in discretizing the scale variable alone, on an arbitrary sequence of values (not necessarily powers of a fixed ratio). This leads to fast algorithms that could put the CWT on the same footing as the DWT in terms of speed and efficiency, by extending the advantages of the latter to cases where no exact QMF is available. We describe the method in 2-D, the 1-D case (already sketched in Section 1.6.1) being easily derived on this basis. Interested readers should refer to [Tor95] for further details. Instead of the standard L 2 -normalization used in (2.13), it is more convenient to Note that, for simchoose the L 1 -normalization and use ψ(b,a) = a −2 ψ(a −1 ( x − b)). plicity, we consider here only isotropic wavelets, but the extension to the general case is straightforward (see Section 2.6.3). Let us start with the L 1 -reconstruction formula associated to the CWT in two dimensions: ∞ da ˘ b, a) ψ(b,a) d 2 b S( x ). (2.133) f ( x) = ( 2 a R 0 The basic idea behind the proposed construction is now to segment the integral over scales in (2.133) and replace it by a sum over dyadic intervals. This is done first by rewriting the reconstruction formula as ∞ da f ( x) = x ), da ( a 0

79

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

where we have defined the infinitesimal detail ˘ b, a) ψ(b,a) da ( x) = d 2 b S( x ). ( R2

By virtue of Young’s convolution inequality, da ∈ L 2 and, taking its Fourier transform, we obtain k)| . = |ψ(a 2 da (k) f (k)

(2.134)

These equations show that da represents the amount of information captured by the wavelet between scales a and a + da, hence the name “infinitesimal details.” Summing all these details, that is, integrating over the scale variable, reproduces the original signal. In the same vein, we can synthesize a low resolution approximation of f by integrating up to a given resolution, say a0 : ∞ da da ( x) = x) . f a0 ( a a0 Taking Fourier transforms on both sides suggests to introduce the following Fourier multiplier: ∞ da 2 2 |φ(k)| = |ψ(a k)| . (2.135) a 1 It is then shown in [159] that the approximation f a can be written f a ( x) = d 2 b φ(b,a) x ), | f φ(b,a) (

(2.136)

R2

and that the following limit holds in the strong sense in L 2 : lim f a = f.

a→0

(2.137)

Thus, following [Tor95], we speak of the bilinear formalism. Remark also that (2.135) implies k)| 2=0 lim |φ(

|k|→∞

and thus defines a scaling function. Now, starting from an approximation of f at scale a0 = 2− jo , we can refine up to an arbitrary resolution by adding up details. For this purpose, we introduce slices of details 2− j da D j ( da ( x) = x) . 2−( j+1) a Taking Fourier transforms on both sides, we have 2− j da 2 ( = |ψ(a k)| f (k) D j (k) −( j+1) a 2

80

The 2-D continuous wavelet transform

and this leads us to define the integrated wavelet packets: 2= | (k)|

1

1/2

da 2 |ψ(a k)| . a

(2.138)

This function satisfies a two-scale relation, the analog of (1.62): 1 k)| k)| 2 − |φ( 2. 2 = |φ( | (k)| 2

(2.139)

Finally, putting equations (2.136) and (2.138) together, we obtain the following dyadic decomposition: f =

R2

2

d b φ(b,2 − jo ) | f φ(b,2 − jo ) +

∞ j= jo

R2

d 2 b (b,2 − j ) | f (b,2 − j ),

(2.140)

which holds in L 2 norm. In order to simplify our notations, we will write the wavelet 2− j ) = (b,2 coefficients of f as W f (b, − j ) | f and introduce the approximation coef−j ficients S f (b, 2 ) = φ(b,2 − j ) | f . Equation (2.140) now reads x) + f ( x ) = S f (·, 2− jo )|φ(·,2− jo ) (

∞

W f (·, 2− j )|(·,2− j ) ( x)

(2.141)

j= jo

(scalar product over b in each term). We have thus built a dyadic wavelet transform starting from the CWT. It is important to realize that the scaling function φ and the integrated wavelet packet inherit the localization and smoothness properties of ψ. In view of the considerable freedom we have in the choice of ψ, we are now able to easily design custom, translation invariant dyadic frames. Remark: More flexibility is obtained if one subdivides the scale interval [1/2, 1] into n subbands, by a0 = 1/2 < a1 < . . . < an = 1. In that case, one ends with one scaling function ( x ) and n integrated wavelets i ( x ), i = 0, . . . n − 1, corresponding to integration from ai−1 to ai . This more efficient version allows one to compute explicitly the characteristics of the wavelet packet, such as its central frequency, its standard deviation, etc., following (1.14)–(1.15), and these in turn may be expressed in terms of the corresponding quantities of the mother wavelet ψ. An application to sound analysis is given in [233]. A simpler decomposition formula, called the linear scheme in [Tor95], arises when one starts from the so-called Morlet reconstruction formula (2.34) (this is the scheme we have sketched in the 1-D case in Section 1.5). The same reasoning as before leads us to introduce, as in (1.61), a scaling function ∞ da k) = φ( ψ(a k) , (2.142) a 1

81

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

and integrated wavelet packets 1 da = ψ(a k) (k) 1/2 a k) 1 k) − φ( . = φ( 2

(2.143) (2.144)

One notices that these wavelets are simply expressed as a difference of smoothing functions, as in 1-D, (1.62). In this case, the reconstruction formula is much simpler, since it involves a straight sum of approximation and wavelet coefficients: f ( x ) = S f ( x , 2− jo ) +

∞

W f ( x , 2− j ) .

(2.145)

j= jo

One of the main advantages of this construction is that it allows to build wavelets and scaling functions that have fast decay both in the spatial and frequency domains. This is very useful in applications where one wants to use wavelets that have sharp prescribed localization properties in the Fourier domain and are also of fast decay in the spatial domain, as it is the case with Gabor functions. This is very difficult to achieve in practice. For example, if one wants to use spline-based wavelet frames, it appears that, although the spatial localization is very good, splines are not sharply localized in Fourier variables (they have an algebraic decay, see [14] for a review) and can even show disturbing sidelobes. A concrete example is given by texture analysis where the latter are distinguished on the basis of the statistics of frequency subbands as measured using Gabor wavelets [234]. If one wants to use a dyadic frame, special care has to be given to the frequency localization of the wavelets and this can be easily done using the technique described above.

2.6.2

Example of a typical design We will now apply the previous formalism to a concrete example. Our aim is to construct isotropic dyadic wavelets and scaling functions belonging to the class of C ∞ functions with fast decay. Our starting point is a family of wavelets associated to n/2 pseudo-differential operators defined by multiplication by k n = k12 + k22 in the Fourier domain: k) = k n e−k 2 . ψ( The associated scaling functions are computed using (2.135): ∞ da n −a 2 a e , (n 2) φn (k) = a k and it satisfies the following recurrence relation: n (k) = 1 k n−2 e−k 2 + (n − 2) φ n−2 (k) , φ 2 2

(2.146)

82

The 2-D continuous wavelet transform

with 2 (k) = 1 e−k 2 , φ 2

3 (k) = 1 ke−k 2 + φ 2

*

π erfc(k). 4

Here the error function is defined by ∞ 4 2 dk e−k . erfc(k) = √ π k Normalizing this family of functions by

R

d 2 x φn ( x ) = 1 leads us to define

n (k) = αn−1 φ(k), φ where the constant αn is defined by n−2

2 (0) αn = φ

2 n − 2i

i=1

2

, n even,

n > 2,

n−3

3 (0) αn = φ

2 2i + 1 , n odd, 2 i=1

n > 3.

2 (0) and α3 = φ 3 (0). Scaling functions of the lowest The recursion starts with α2 = φ orders are listed in Table 2.1. Using (2.138) or (2.144), we obtain the desired family of isotropic wavelets. An example of such a scaling function of order 4 is given in Figure 2.12, in the next section. Note that the parameter n also controls the number of vanishing moments of the associated wavelet. As an example of this technique, Figure 2.10 shows a three-level decomposition of the lena image into isotropic wavelet packets.

2.6.3

Designing directional dyadic frames The previous construction applied only to isotropic wavelets and yielded frames of isotropic elements. In the following we will add to this scheme sensitivity to local orientation. A simple and straightforward way to achieve this is to start from an isotropic integrated wavelet and segment it into directional ones (a precise definition of this Table 2.1. Scaling functions of lowest orders Order

Scaling function

n=2 n=3 n=4

2 (k) = e−k 2 φ √ 3 (k) = 1 ke−k 2 + π erfc(k) φ 2 4 2 4 (k) = e−k (1 + k 2 ) φ 5 (k) = α5−1 1 k 3 e−k 2 + 3 ke−k 2 + φ 2 4

n=5

√ 3 π 8

erfc(k)

83

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

Fig. 2.10. Three levels of decomposition of the lena image using an isotropic wavelet packet frame. The lower right image is the low resolution approximation.

notion will be given in Section 3.3). This can be done by introducing an angular window η(ϕ), ϕ ∈ [0, 2π ), in the Fourier domain and then defining a new wavelet ϕ) = (k, (k)η(ϕ) . Note that this construction amounts to work with wavelets that are separable in polar coordinates. The choice of the angular window is restricted by the need for an exact, linear or bilinear, reconstruction formula. More precisely, if one makes use of (2.141), η has to satisfy + L−1 ++ + 2 +η ϕ − 2π + = 1, + L + =0

(2.147)

84

The 2-D continuous wavelet transform

while the simpler formula (2.145) requires L−1 2π η ϕ− = 1, L =0

(2.148)

where we have assumed L orientations. An additional requirement, further explored in Chapter 3, is that the support of η be strictly less than π in order to have some directional sensitivity. In order to preserve the frequency localization of , it is also important that η be regular enough. The optimal choice is thus to build a partition of the circle using a suitable compactly supported C ∞ function. Altogether, we introduce the L angular windows η (ϕ) ≡ η ϕ − 2π ,= L 0, 1, . . . , L − 1, and the corresponding directional wavelets (k, ϕ) = (k) η (ϕ), = 0, 1, . . . , L − 1 .

(2.149)

An example of a dyadic directional wavelet built using this technique is depicted on Figures 2.11 and 2.12.

2.6.4

Implementation using approximate QMFs One of the main drawbacks of these oriented frames is that they are not designed to be implemented using a fast pyramidal algorithm. Nevertheless we will now show that one can design special QMF pairs that allow for a very good approximation of the discrete WT and provide a substantial gain in computational speed. This technique is mainly an extension to 2-D of the original work of Muschietti and Torr´esani [291]. −3.14

−3.14

0

0

3.14 −3.14

0

(a)

3.14

3.14 −3.14

0

3.14

(b)

Fig. 2.11. Fourier transform of a directional dyadic wavelet of order 4 and angular resolution of π/5 for two values of the rotation parameter: (a) = 0; and (b) = 1.

85

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

1 0.025

0.6

0.02 0.015

0.8

0.01

0.4 0.005

0.2

0 −0.005

0.0

−0.01

−0.2 1

−0.015 1

1

0.5

1

0.5

0.5

0 −0.5

0

−0.5

−0.5 −1

0.5

0

0 −1

−0.5 −1

(a)

−1

(b)

0.025 0.02 0.015

0.02

0.01

0.015

0.005 0

0.01

−0.005 −0.01

0.005

−0.015

0 1

−0.02 1 1

0.5 0.5

0

1

0.5

0

−0.5

0

−0.5

−0.5 −1

0.5

0 −0.5 −1

−1

−1

(d)

(c)

Fig. 2.12. (a) Scaling function of order 4; (b–d) the associated wavelet with angular resolution of π/5: (b) real part; (c) imaginary part; and (d) modulus.

Since we work with translation invariant frames, we will now sample the position parameter of the DWT over a regular grid, that is, we will consider wavelets and scaling functions indexed by the integer grid (we use here the L 1 -normalization): j;m,n (x, y) = 22 j (2 j (x − m, y − n)) ,

j, m, n ∈ Z .

We then obtain a discrete dyadic wavelet transform (compare Section 2.4.4) by just restricting the DWT to this particular grid: W j f (m, n) = j;m,n | f ,

S j f (m, n) = φ j;m,n | f .

(2.150)

86

The 2-D continuous wavelet transform

Let us assume there exists 2-D discrete filters h and g , = 0 . . . L − 1, such that one can compute these quantities using a pyramidal algorithm [260], that is, S j f (m, n) = h p,q S j+1 f (m − 2 j+1 p, n − 2 j+1 q) (2.151) p,q∈Z

W j f (m, n) =

g p,q S j+1 f (m − 2 j+1 p, n − 2 j+1 q) .

(2.152)

p,q∈Z

This is equivalent to asking that the related wavelets and scaling function satisfy a two-scale equation of the form: k) k), = h(k) φ( φ(2 (2k) (k), = g (k)

(2.153) (2.154)

≡ g (k x , k y ), = 0 . . . L − 1, are the Fourier series ≡ h(k x , k y ) and g (k) where h(k) of the filters h and g respectively. As already stressed in Chapter 1, we know that the same filters can be used to reconstruct the signal provided they satisfy a QMF relation: 2+ |h(k)|

L−1

2 = 1. |g (k)|

=0

This is the exact discrete equivalent of (2.141). Similarly, one can design filters that implement the weaker formula (2.145), provided they satisfy a simpler constraint: + h(k)

L−1

= 1. g (k)

=0

The main problem at this point is that the integrated directional wavelets we have designed in Section 2.6.1 do not in general satisfy any two-scale equation. Another way to formulate the problem is to remark that, although there usually exists regular multipliers hˇ and gˇ satisfying k) k) ˇ k) = h( φ( , φ(2

k) (2k) = gˇ (k) φ( ,

these are in general not 2π × 2π periodic and hence cannot be used to compute successive approximations and details as in (2.151) and (2.152). Nevertheless it is crucial to notice that these multipliers, exactly as the filters h and g , are always multiplied by Now, if the latter is very well localized in [−π, π] × [−π, π], the lack of periodicity φ. and it seems then reasonable to of hˇ and gˇ is compensated by the localization of φ ˇ gˇ using periodic filters. For this purpose, let us look for good approximations of h, 2 assume that our signal lives in L (R2 ) and introduce the subspace of L 2 (R2 ) spanned by the integer translates of the scaling function defined in (2.135): V0 = { f ∈ L 2 | f = cm,n φ(x − m, y − n) , cm,n ∈ 2 } m,n∈Z

87

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

and suppose that the family {φ(x − m, y − n) , m, n ∈ Z} forms a Riesz basis of V0 . Finding a best approximant for hˇ can be formulated as finding an element of V0 whose distance to 14 φ(x/2, y/2) is minimal in L 2 . Similarly, finding best approximants for the gˇ is equivalent to finding those elements of V0 that minimize the distance to 1 (x/2, y/2). In other words, the problem is to minimize the L 2 distances 4 !1/2 a 2 ˇ a 2 ˇ d k |h(k) − h (k)| |φ(k)| , ν(h, h ) =

ν(gˇ , g

,a

)=

R2

k)| − g ,a (k)| |φ( 2 d k |gˇ (k) 2

R2

!1/2 .

Now since V0 is a vector subspace of a Hilbert space, the projection theorem applies and guarantees the existence and uniqueness of such solutions. Suppose that h a ∈ V0 ˇ It can be expanded as is the solution for h. ha = h am,n φm,n , m,n∈Z

where the h am,n are the Fourier coefficients of the approximate filter ≡ h a (k x , k y ) = h a (k)

1 a −i(mkx +nk y ) h e . 4π 2 m,n∈Z m,n

The projection theorem gives also the following characterization of these coefficients:

# x y 1 d x d y φ(x + m, y + n) h ap,q φ(x + p, y + q) − φ( , ) = 0 , 4 2 2 R2 p,q∈Z "

∀ m, n ∈ Z .

(2.155)

,a Similarly, for the directional wavelets, one obtains L approximate filters gm,n characterized by " # 1 x y ,a d x d y φ(x + m, y + n) g p,q φ(x + p, y + q) − ( , ) = 0 , 4 2 2 R2 p,q∈Z

∀ m, n ∈ Z .

(2.156)

Finally, for j −1, the approximate low resolution and detail coefficients read S aj f (m, n) = h ap,q S aj+1 f (m − 2 j+1 p, n − 2 j+1 q) (2.157) p,q∈Z

and W ,a j f (m, n) =

p,q∈Z

a j+1 g ,a p, n − 2 j+1 q) . p,q S j+1 f (m − 2

(2.158)

88

The 2-D continuous wavelet transform

The following theorem, due to Gobbers and Vandergheynst [360], extends to 2-D the result obtained in 1-D by Muschietti and B. Torr´esani [291]. It gives explicit formulas for the best approximants h a and g ,a , as well as an estimation of the error with respect to the original coefficients S j f (m, n) and W j f (m, n): Theorem 2.6.1 (i) The optimal filters h a and g ,a , solutions of (2.155) and (2.156), are given by

p,q

h (k x , k y ) = a

2(k x + 2π p), 2(k y + 2πq) φ(k x + 2π p, k y + 2πq) φ ,

2 p,q |φ(k x + 2π p, k y + 2πq)| (2.159)

g

,a

(k x , k y ) =

p,q

2(k x + 2π p), 2(k y + 2πq) φ(k x + 2π p, k y + 2πq) .

2 p,q |φ(k x + 2π p, k y + 2πq)| (2.160)

and C = 2 ess sup g ,a (k), then, for all j −1, (ii) If we write† C0 = 2 ess sup h a (k) R2 k∈

R2 k∈

| j|

ˇ h a ) 1 − C 0 f 2 , S aj f − S j f ∞ 2 ν(h, 1 − C0 and

W ,a j f − W j f ∞

| j|−1

ˇ h a ) 1 − C0 2 ν(gˇ , g ,a ) + C ν(h, 1 − C0

f 2 . (2.161)

Proof . We will essentially follow the proof given in [291]. The first part is obtained by taking the Fourier transform of (2.155) and (2.156) and using the periodicity of h a and g ,a . We then have 2π 2π + +2 ik x m+ik y n + dk x dk y e φ(k x + 2π k, k y + 2πl)+ 2 h a (k x , k y ) 0

0

k,l∈Z

−

, x + 2πk, k y + 2πl) φ 2(k x + 2π k), 2(k y + 2πl) = 0 , 2 φ(k

k,l∈Z

(2.162) which gives the result for h a and similarly for g ,a . As for the second part of the theorem, the inequality f ∞ f 1 /4π 2 allows us to work directly in the Fourier domain. For the purpose of the calculation, let us introduce the intermediate quantities †

Since this quantity is orientation independent, we drop the corresponding superscript for ease of notation.

89

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

Sj (m, n) =

h ak,l S j+1 m − 2 j+1 k, n − 2 j+1l ,

k,l∈Z

(m, n) = W j

j+1 g k,a p, n − 2 j+1 q . p,q S j+1 m − 2

p,q∈Z

We have the inequality -(a -(a - + -S − S - , S − S − S S - j - j - j jjj1

1

(2.163)

1

- ,a and similarly for -W j − W j 1 . Let us now compute the second term in the right-hand side of (2.163): - dk x dk y 2 j -S j − S j - = 1 S 1 ×S 1 + + + 2 j (k x + 2π k), 2 j (k y + 2πl) f (k x + 2πk, k y + 2πl)φ + + k,l∈Z f (k x + 2πk, k y + 2πl) h a 2 j−1 k x , 2 j−1 k y − k,l∈Z

+ + j−1 2 (k x + 2π k), 2 j−1 (k y + 2πl) ++ . ×φ +

Using the Cauchy–Schwarz inequality and the periodicity of h a , we find . + + - k) k) φ( − φ(2 +2 d 2 k + h a (k) -S j − Sj - 8π 2 f 2 1

R2

ˇ h a ). 8π f 2 ν(h, 2

For the first term on the right-hand side of (2.163), we find + + + + a -(a + d 2 k +h a 2 j+1 k + +S -S j − S j - 2 j+1 (k) − S j+1 (k) 1 R2 - a 2 ess sup |h a | -S − S - . j+1 j+1 R2 k∈

1

Combining these estimations, we have -(a a ˇ h a ) + C0 − S -S - . -S j − S j - 8π 2 f 2 ν(h, j+1 j+1 1

1

(2.164)

By iteratively bounding the last term of (2.164) in the same way, we finally obtain -(a ˇ h a ) 1 + C0 + C02 + . . . + C | j|−1 -S j − S j - 8π 2 f 2 ν(h, 0 1

| j|

ˇ ha ) = 8π 2 f 2 ν(h,

1 − C0 . 1 − C0

An equivalent processing of the wavelet coefficients concludes the proof.

(2.165)

90

The 2-D continuous wavelet transform

There remain several open questions related to the design of these approximate QMFs. In particular, in contradiction to the orthogonal case discussed in the previous chapter, the convergence of the cascade algorithm is not ensured. It is thus impossible to tell precisely if these filters really correspond to a frame and what would be the properties of such a frame. This scheme is thus really a handy numerical shortcut that allows one to efficiently compute an approximation of the original frame expansion.

2.6.5

Some implementation issues When implementing (2.157) and (2.158), the first fact to consider is that, generally, we possess only a finite number M × N of samples f (m, n) (with (m, n) ∈ [0 . . . M − 1, 0 . . . N − 1] ) of the signal to be analyzed. A decision has thus to be made about the nature of the signal outside this range. Furthermore, the filters will practically be computed on a finite grid. That is, we will only use a finite number P × Q of filter coefficients h ap,q and g k,a p,q , with ( p, q) ∈ [−P/2 . . . P/2 − 1, −Q/2 . . . Q/2 − 1]. In a first approach, the signal is considered to be zero outside this range. Unfortunately, this decision leads to impractical and inefficient algorithms, as one has to compute (M + 2| j|−1 (P − 1)) × (N + 2| j|−1 (Q − 1)) coefficients for each j < 0 to avoid side effects. A second approach is to consider the signal as being periodic of period M × N . With this approach, fast circular convolution algorithms can be used and we are led to the following algorithm structure: (i) compute h a and g ,a and their associated impulse responses using Theorem 2.6.1; (ii) compute a first approximation of the analyzed signal f using (2.150); this step is traditionally skipped and the signal is considered as its first approximation; (iii) for each j < 0, iteratively compute details and approximations using (2.157) and (2.158). The cost of this algorithm is: C(M, N ) = c · M · N where the constant c depends on the size of the impulse responses of the filters. This dependency strongly limits the power of this algorithm as experiments show that, even for small sizes, this algorithm is always slower than the FFT-based algorithm used to compute the CWT. As those sizes may not be arbitrarily chosen, another algorithm should be used in order to get valuable results with small computation times. The obvious way to handle this problem is to replace convolutions in direct space by products of Fourier transforms in frequency space. The main advantage of this technique is that we no longer need to restrict ourselves to use small filters, as we may now use impulse responses that are the same size as the signal, thus giving much

91

2.6 Bridging the gap: continuous wavelet packets and fast algorithms

more precise results than with the above algorithm. For the sake of simplicity, we will develop this new algorithm in the 1-D case, its extension to the 2-D directional case being straightforward. For the first step, we have to compute two periodic convolution products: c−1,k =

N −1

c0,k−n h n

n=0

d−1,k =

N −1

c0,k−n gn .

n=0

By introducing {C j,n } = F F TN {c j,k } , {D j,n } = F F TN {d j,k } , {Hn } = F F TN ({h k }) and {G n } = F F TN ({gk }) where the notation F F TN means the FFT algorithm ap√ √ plied to a sequence of length N , we get: C−1,n = N C0,n Hn and D−1,n = N C0,n G n . We may then get the details back in the real space with {d j,k } = I F F TN {D j,n } . For the second step, we have to compute: c−2,k =

N −1

c−1,k−2n h n

n=0

d−2,k =

N −1

c−1,k−2n gn .

n=0

Given the periodicity of the signal, we may rewrite this as:

0 c−2,k ≡ c−2,2k =

N /2−1

c−1,2(k−n) (h n + h n+N /2 )

n=0 1 c−2,k

≡ c−2,2k+1 =

N /2−1

c−1,2(k−n)+1 (h n + h n+N /2 )

n=0 0 d−2,k ≡ d−2,2k =

N /2−1

c−1,2(k−n) (gn + gn+N /2 )

n=0 1 ≡ d−2,2k+1 = d−2,k

N /2−1

c−1,2(k−n)+1 (gn + gn+N /2 )

n=0

for k = 0, . . . , N /2 − 1. Let us now define h 1n = h n + h n+N /2 for n = 0, . . . , N /2 − 1, we get: 0 c−2,k =

N /2−1

0 1 c−1,k−n h 1n c−2,k =

n=0 0 d−2,k =

N /2−1 n=0

N /2−1

1 c−1,k−n h 1n

(2.166)

1 c−1,k−n gn1

(2.167)

n=0 0 1 c−1,k−n gn1 d−2,k =

N /2−1 n=0

92

The 2-D continuous wavelet transform 0 1 0 1 where c−1,k = c−1,2k , c−1,k = c−1,2k+1 , d−1,k = d−1,2k and d−1,k = d−1,2k+1 . In the Fourier space, (2.166) and (2.167) may be rewritten as: / / 0 0 1 1 C−2,n = N /2C−1,n Hn1 C−2,n = N /2C−1,n Hn1 / / 0 0 1 1 D−2,n = N /2C−1,n G 1n D−2,n = N /2C−1,n G 1n

0 1 0 1 0 where {C−2,n } = F F TN /2 {c−2,k } , {C−2,n } = F F TN /2 {c−2,k } , {D−2,n }= 0 1 1 F F TN /2 {d−2,k } and {D−2,n } = F F TN /2 {d−2,k } . A straightforward calculation gives: √ Hn1 = 2H2n for n = 0, . . . , N /2 − 1. Furthermore: 1 0 C−1,n = √ (C−1,n + C−1,n+N /2 ) 2 2iπ Nn e 1 C−1,n = √ (C−1,n − C−1,n+N /2 ) 2 for n = 0, . . . , N /2 − 1. The last equation is particularly interesting as it introduces the same twiddle factors as those already present in the traditional FFT implementations. Extending the above results to the following steps, one comes to the conclusion that the complexity of this new algorithm is also of order N log2 (N ), but with a hidden constant that is exactly half the one encountered in the FFT-based algorithm, this constant being associated with the pyramidal structure. The same ideas apply to the 2-D case, thus giving an algorithm of complexity C(M, N ) = M · N · log2 (M.N ), with a hidden constant also exactly half that of the FFT-based algorithm. As a matter of comparison, Figure 2.13 shows timings of this algorithm for different image sizes. Timings of the usual pyramidal algorithm and standard implementation in the Fourier domain are also displayed. Putting it all together, we have a new fast algorithm, perfectly suited to compute the 2-D CWT, faster than the traditional “pseudo-pyramidal” algorithm, and sharper. Furthermore, it is essential to note that the whole construction is equivalent to that leading to the FFT algorithm in the Fourier transform theory. It should be noted that a Fourier implementation of the pyramidal algorithm is quite natural when one addresses the problem of designing maximally regular wavelets. That is why the algorithm described above shares common features with the Fourier implementation of the Meyer wavelet decomposition [Kol97]. We refer the interested reader to the work of Rioul and Duhamel [327] for more general considerations on implementing pyramidal algorithms in the frequency domain.

93

2.7 Steerable filters

6

5 pyr (32)

4

3

2 pyr (16)

std_F

1

pyr (8) pyr_F 0 128

256

512

1024

Fig. 2.13. Computation time for different implementations of the 2-D pyramidal algorithm as a function of the image width (square images assumed): pyr(n) is the standard pyramidal algorithm with n by n filters, pyr F is the modified Fourier pyramidal algorithm explained in the text. The graph is normalized with respect to the standard algorithm with convolution in the Fourier domain (std F).

2.7

Steerable filters While looking for a flexible tool for processing oriented data, Freeman and Adelson [170] introduced some time ago the concept of steerable filters. These were further developed by Perona [311] and Simoncelli et al. [342]. Here again one obtains a multiscale pyramid decomposition, which is quite efficient in a number of problems, mostly related to machine vision. Similar techniques have been used with the Gabor transform [234]. We will briefly describe this scheme and compare it to the directional wavelet packets of Section 2.6. The basic idea is quite simple, and best illustrated on the example of a Gaussian kernel 2 2 2 2 G(x, y). From the partial derivatives G x (x, y) = −2x e x +y , G y (x, y) = −2y e x +y , one computes the derivative in the direction θ: G θ (x, y) = cos θ G x (x, y) + sin θ G y (x, y). Since convolution is a linear operation, one may use G θ for filtering an image f in the direction θ by superposing the filterings in directions x and y: = (G θ % f )(b) = cos θ Fx (b) + sin θ Fy (b). Fθ ( f )(b)

94

The 2-D continuous wavelet transform

This is the property of orientability. More generally, a filter f is orientable or steerable if any oriented version of it may be obtained from a finite number of basic orientations: x )) = f (rθ (

M

km (θ) f (rθm ( x )).

(2.168)

m=1

The weights {km (θ), m = 1, . . . , M} are called interpolation functions. (The notion of orientability may be extended to other transformations, such as scaling [311], but we will not consider these generalizations here.) When the filter f admits a finite Fourier series (i.e., it is a real trigonometric polynomial), f (r, ϕ) =

N

an (r ) einϕ ,

x ≡ (r, ϕ),

(2.169)

n=−N

Freeman and Adelson have shown that the interpolation functions must satisfy the relation 1 ... 1 1 eiθ eiθ1 . . . eiθ M k1 (θ) . (2.170) .. = .. .. .. . . . k M (θ) ei N θ e i N θ1 . . . e i N θ M and that the minimal number of interpolation functions is always larger than the number of nonzero coefficients in the angular Fourier series (2.169) of the filter f . These steerable filters obviously generalize the interpolation properties associated to the partial derivatives G x , G y , which are thus prototypes of such filters. Moreover, the steerability property (2.168) is independent of the radial part of the filter, so that one may use functions that generate a dyadic pyramid. The main virtue of steerable filters is to allow the computation of filtering in any direction from the interpolation functions and the basic filters. This explains their intensive use in vision for studying oriented features in images [311,342]. From the algorithmic point of view, the complexity of steerable filters is that of the associated dyadic pyramid, thus comparable to that of directional wavelet packets. Note, however, that separable filters may sometimes be obtained, which is an additional bonus. So the natural question arises, can directional wavelets be made steerable? Unfortunately, the answer is no. Indeed, the Fourier transform of a steerable filter is also steerable. On the other hand, the angular support of a directional wavelet (2.149) necessarily has a width smaller than π , so that the relation (2.169) can never be satisfied. More intuitively, the steerability condition (2.168) requires that the basic filters overlap in a fixed way, in order to ensure the existence of interpolation functions, whereas the angular overlap of directional wavelet packets may be taken to be arbitrarily small. What about angular resolution? That of steerable filters may be made as good as one wishes, simply by taking more basic filters. However, because of the substantial overlap

95

2.7 Steerable filters

(d)

(b)

(c)

(a)

Fig. 2.14. Four reconstructions of the image barbara: (a) 2-D wavelet orthonormal basis; (b) redundant frame (3 bits/coefficient); (c) and (d) the same with 2 bits/coefficient.

between the latter, the steerable scheme will always require more basic filters than the directional wavelet packets in order to achieve a high angular resolution, so that, in the end, the computing cost may become prohibitive. In conclusion, steerable filters and directional wavelet packets are two comparable, yet incompatible, tools for decomposing an image into an oriented pyramid. In the former case, however, a large number of basic filters, with fixed overlap, is required for achieving a high angular resolution. This precludes their use in applications that require a maximal decorrelation of orientations, such as watermarking of images or texture segmentation. Here directional wavelet packets are the best choice. We will explore these applications in the next chapters (Sections 4.7.2 and 5.5, respectively).

96

The 2-D continuous wavelet transform

2.8

Redundancy: plus and minus Exactly as in 1-D, redundancy has many advantages, that more than compensate the higher computational cost it implies. In a nutshell, redundant decompositions lead to better quality reconstructions and are more robust to noise. Actually, the whole discussion of Section 1.6.2 could be repeated here almost verbatim, in particular concerning the robustness issue. We highlight the reconstruction aspect with one striking example. We consider the standard barbara image and decompose it in two ways, first with an orthonormal wavelet basis (using 2-D Daubechies DB4 wavelets), then with a redundant frame of directional wavelets. The images reconstructed by the two methods are presented in Figure 2.14. Panels (a) and (b) show the reconstruction using 3 bits per coefficient, while (c) and (d) show the result obtained with 2 bits per coefficient. In either case, the resulting image is visually better when the redundant frame is used, (b) or (d). The orthonormal basis gives more artifacts and distortions. Of course, the effect is more marked in the 2 bit case. Although the two results are poor, we show them for emphasizing the point.

3

Some 2-D wavelets and their performance

3.1

Which wavelets? The next step is to choose an analyzing wavelet ψ. At this point, there are two possibilities, depending whether one is interested or not in detecting oriented features in an image, i.e., regions where the amplitude is regular along one direction and has a sharp variation along the perpendicular direction. (i) Isotropic wavelets If one wants to perform a pointwise analysis, i.e., when no oriented features are present or relevant in the signal, one may choose an analyzing wavelet ψ which is invariant under rotation. Then the θ dependence drops out, for instance, in the reconstruction formula (2.26). The most familiar example is the isotropic 2-D Mexican hat wavelet (2.21). (ii) Anisotropic wavelets When the aim is to detect oriented features in an image (for instance, in the classical problem of edge detection or in directional filtering), one has to use a wavelet which is not rotation invariant. The best angular selectivity will be in obtained if ψ is directional, which means that the (essential) support of ψ spatial frequency space is contained in a convex cone with apex at the origin (by which we mean that the wavelet is numerically negligible outside the cone). Typical directional wavelets are the 2-D Morlet wavelet (2.22) or the conical wavelets. There are many ways of designing wavelets of either kind, but in fact almost all of those available on the market may be obtained by a general procedure, outlined in the proposition below. The starting point is a scaling function, that is, a function φ( x ), whose integral over the plane does not vanish. Then wavelets can be built by taking derivatives of the scaling function or the difference of two scaling functions. The most obvious example of a scaling function is a Gaussian, which is very easy to use and essentially localized in a disk. Another method, that we shall describe in Chapter 6, is to impose the saturation of the uncertainty product of two or more operators corresponding to infinitesimal generators

97

98

Some 2-D wavelets and their performance

of elementary operators, namely the dilation, orientation and translation operators, as given in Section 2.1 above. Proposition 3.1.1 Let φ be a (sufficiently smooth) scaling function, that is, a function satisfying: d 2 x φ( x ) = 0. (3.1) R2

Then the functions ψ1 and ψ2 defined below are wavelets: N M

∂n ∂m φ( x ), where x ≡ (x, y), N , M 1, ∂ x n ∂ ym n=1 m=1 1 1 x) = x) − x ), U (b1 , a1 , θ1 )φ ( U (b2 , a2 , θ2 )φ ( ψ2 ( a1 a2 (b1 , a1 , θ1 ) = (b2 , a2 , θ2 ), ψ1 ( x) =

cnm

(3.2)

(3.3)

a, θ ) is the unitary operator already defined in (2.13): where U (b, a, θ )s ( x − b)). U (b, x ) = a −1 s(a −1 r−θ ( The proof is straightforward. We shall now examine in detail several examples of wavelets of each kind. As will be seen in the sequel, almost all of them fall into the general types described in the proposition, possibly combining the two operations of derivation and difference. It turns out also that in many cases, the basic scaling function is the Gaussian exp(−| x |2 ). This is not a coincidence. In order to understand this, assume the wavelet ψ to be, as in (3.2), some derivative of a scaling function φ, ψ( x) =

∂n ∂m φ( x ). ∂ x n ∂ ym

Then we may rewrite the basic formula (2.19) as a genuine convolution: ∂n ∂m # a, θ ) = φa,θ ∗ n m s (b) S(b, ∂x ∂ y ∂ n ∂ m # = φa,θ ∗ s (b), ∂bx ∂b y

(3.4) (3.5)

where # φa,θ ( x ) = a −1 φ(−r−θ ( x )/a).

a, θ) is simply the derivative of the signal rotated and Equation (3.5) shows that S(b, blurred at resolution (scale) a. Thus we expect large wavelet coefficients to occur at locations of sharp variation of s. Since the Gaussian is a very common kernel for blurring, it is not a surprise that many wavelets will be related to derivatives of a Gaussian. As for the populatity of the latter, it is due to several factors: it is very well

99

3.2 Isotropic wavelets

localized, both in position and in Fourier space, it is indeed optimal for the joint space– frequency localization, as argued by Gabor [180], and finally it is rotation invariant, thus it does not privilege any particular direction.

3.2

Isotropic wavelets

3.2.1

The 2-D Mexican hat and its generalizations In its isotropic version, this is simply the Laplacian of a Gaussian: x ) = − exp(− 12 | x |2 ) = (2 − | x |2 ) exp(− 12 | x |2 ). ψH (

(3.6)

This is a real, rotation invariant wavelet, with vanishing moments of order up to 1, also known in the literature as the LOG wavelet. It is shown, in position domain and in spatial frequency space, in Figures 3.1 and 3.2, respectively. It was originally introduced by Marr and Hildreth [Mar82,266], in their pioneering work on vision, precisely because it

Fig. 3.1. The 2-D Mexican hat wavelet in position domain, seen in 3-D perspective (left) and gray

levels (right): (a) the isotropic wavelet; (b) the anisotropic wavelet with $ = 2.

100

Some 2-D wavelets and their performance

20 0.8

40

0.6 60

0.4 0.2

80 5

0 −5

100 0

0

120

5 −5

(a)

20

40

60

80

100

120

20

40

60

80

100

120

20 2

40

1.5 60

1 0.5 0 −15

80 5 −10

100 −5

0

0 5

10

120

15 −5

(b) Fig. 3.2. The same wavelets as in Figure 3.1, seen in spatial frequency space.

is obtained by applying an isotropic differential operator of second order to the Gaussian (this was in fact the original motivation of [266]). The Mexican hat will be efficient for a fine pointwise analysis, but not for detecting directions. This will be confirmed by quantitative calibration tests in Section 3.4 below. In addition, and for the same reasons, one may also use higher order Laplacians of the Gaussian, ψH(n) ( x ) = (− )n exp(− 12 | x |2 ).

(3.7)

For increasing n, these wavelets have more and more vanishing moments, and are thus sensitive to increasingly sharper details. An interesting technique, pioneered in 1-D by A. Arn´eodo [49], is to analyze the same signal with several wavelets ψH(n) , for different n. The features common to all the transforms surely belong to the signal, they are not artifacts of the analysis. In several applications, it is useful to introduce an additional parameter, namely, the width σ of the Gaussian (in k-space). Although σ is redundant, since the Gaussian

101

3.2 Isotropic wavelets

can be dilated to an arbitrary width, nevertheless it helps to fix explicitly the central frequency. Thus one uses, instead of (3.6), the wavelet x ) = − exp(− ψH (

σ 2 | x |2 ), 2

2 2 H (k) = |k| exp(− |k| ) . ψ 2 σ 2σ 2

(3.8)

An approximate version of the Mexican hat has been introduced by Arn´eodo et al. [Arn95,44,171], under the name of Bessel filter, namely: B (k) = 1 , for R/γ |k| R, ψ = 0 , otherwise,

(3.9)

where R and R/γ are the radii of the external and internal disks, respectively. The inverse Fourier transform ψB ( x ) is a Bessel function (hence the name): x) ψB (

1 J1 (r/γ ) J1 (r ) − 2 , r γ r/γ

r = | x |.

(3.10)

This filter was designed and used systematically for the so-called optical wavelet transform, which consists in a hardware (optical) realization of the CWT. We will come back to this application in Section 5.4.1 [Arn95,46]. Another isotropic wavelet, very similar to the previous one, has been introduced in [175] under the rather funny name of Pet hat. It is defined in Fourier space as k) = cos2 π log2 |k| , for π < |k| < 4π, ψ( 2 2π < π and |k| > 4π. = 0 , for |k|

(3.11)

This wavelet has a better resolving power in scale than the Mexican hat, hence it is more efficient in sorting objects in astrophysical images according to their characteristic scale, which is precisely the aim of the authors of [175]. Yet another one, which has the advantage of being both isotropic and continuous, is the Halo wavelet [117], defined as 2 −|ko |2 ) k) = c e−(|k| ψ( .

(3.12)

|ko |. This is a real wavelet, that selects precisely the annular region |k|

3.2.2

Difference wavelets Among the many wavelets (or filters) proposed in the literature, an interesting class consists of wavelets obtained as the difference of two positive functions, according to (3.3) in Proposition 3.1.1. In order to get an isotropic wavelet in this way, the only possibility is to take the difference between a single isotropic function φ and a contracted version of the latter, that is, the particular case where only the scale factors differ in (3.3). Indeed, if φ is a smooth non-negative function, integrable and square integrable,

102

Some 2-D wavelets and their performance

with all moments of order one vanishing at the origin, then the function ψ given by the relations ψ( x ) = α −2 φ(α −1 x) − φ( x ),

k) k) k) = φ(α − φ( ψ(

(0 < α < 1)

(3.13)

is easily seen to be a wavelet satisfying the admissibility condition (2.17). Since φ is typically a smoothing function, the wavelet ψ is called the “Difference-of-Smoothings” or DOS wavelet [Duv91]. A typical example is the “Difference-of-Gaussians” or DOG wavelet, obtained by taking for φ an isotropic Gaussian: ψD ( x) =

1 2α 2

exp(− 2α1 2 | x |2 ) − exp(− 12 | x |2 ) (0 < α < 1).

(3.14)

The DOG filter is a good substitute for the isotropic Mexican hat (for α −1 = 1.6, their shapes are extremely similar), frequently used in psychophysics works [DeV88,Duv91,124]. It was also considered by Grossmann [209] for signal analysis, together with more general linear combinations of Gaussians. An immediate application is the construction of wavelets on the 2-sphere, simply by lifting a DOG wavelet in the tangent plane to the sphere by inverse stereographic projection (Section 9.2). Another example is the “Difference-of-Mesas” or DOM filter, corresponding to a function φ which is a smoothed version of the characteristic function of a disk (a ‘mesa’ function) [369,370]. The resulting annular wavelet has been used, together with the Halo wavelet (3.12), in the detection of Einstein gravitational arcs in cosmological pictures [82] (see Section 5.1). The principle behind this application is again the filtering property: the wavelet detects preferentially objects that resemble it. More generally, the concept of the difference wavelet is useful for reducing noise in images. Take an image, consisting of an object to be identified, embedded in noise or clutter. Let φ( x ) be an averaged version of the image. Then the corresponding difference function ψ( x ) given by (3.13) is a wavelet ideally suited for the analysis of the object in question [16]. Indeed the difference operation substantially reduces the background noise, and ψ incorporates a maximal amount of resemblance with the object (a priori information). A similar subtraction technique, known in astronomy as “unsharp masking,” is commonly used in the treatment of astrophysical images for enhancing the relevant information (such as galaxies or stars), while reducing the noise. Here one computes first the high-frequency content of the image as the difference Ih = Io − Is between the original image Io and a smoothed version Is of it, and adds it to the original image, with a multiplicative factor λ. Then, in the corrected image Icorr = Io + λIh = Io + λ(Io − Is ), the high-frequency details are enhanced over the background (but severe distortions may result if λ is chosen too big) [258]. An example of unsharp masking may be found at the address http://www.chapman.edu/oca/gallery2/demo.htm. A similar procedure has been introduced for the problem of automatic target recognition (ATR). We will discuss the corresponding algorithm in Section 4.2.1 and the general problem of image denoising in Section 4.6.

103

3.3 Directional wavelets

An additional advantage of these difference wavelets is that they lead to interesting and fast algorithms, for instance, in the formalism of continuous wavelet packets [159]. Indeed, in 1-D, we have seen in (1.62), that the integrated wavelet (x) associated to the scaling function (x) is precisely (x) = 2 (2x) − (t x), and this property extends to 2-D, equation (2.144) (the L 1 normalization is used here, contrary to (3.13)). Notice, finally, that φ, and thus also ψ, need not be isotropic. The directional wavelet packets constructed in Section 2.6 are a striking example.

3.3

Directional wavelets

3.3.1

Oriented wavelets and edge detection Detecting oriented features (segments, edges, vector field, . . . ) poses a major challenge in computer vision, and many techniques have been proposed in the literature to meet it. In the context of wavelet analysis, one needs a wavelet which is directionally selective. A natural way of designing such a wavelet is to modify an isotropic one, such as the Mexican hat, simply by stretching it. Mathematically, this amounts to replacing in (3.6) x by A x, where A = diag[$ −1/2 , 1], $ 1, is a 2 × 2 anisotropy matrix. However, such a wavelet still acts as a second-order operator and detects singularities in all directions and it is of little use in practice. Indeed, on one hand, the calibration tests that we will discuss in Section 3.4 below show that it performs poorly [13]. On the other hand, there is a theorem due to Daugman [125], according to which no real wavelet rendered anisotropic by a mere stretching in one direction can have a good directional selectivity, no matter how large the anisotropy $ is taken.

3.3.1.1

Some precursors Stretching an isotropic wavelet being inefficient for inducing directional selectivity, the next step is to take a directional derivative. As we have seen in Section 3.2.1, Marr and Hildreth [266] apply the Laplacian to the Gaussian, thus getting the Mexican hat wavelet, precisely because it is an isotropic differential operator of second order. As a consequence, the Mexican hat is inefficient at detecting directions. In order to get a good edge detector, Canny [98] designed a filter which is optimal for several criteria (detection, localization, uniqueness of answer). However, this filter is numerically heavy to implement and it is advantageously replaced, to a very good approximation, by the first derivative of a Gaussian. The wavelet d −|x |2 /σ 2 ψ (1) ( e x) = (3.15) dx detects edges oriented in the y-direction, and it suffices to rotate it to get an edge detector that is sensitive to an arbitrary direction. In addition, Canny considered many different values of the width σ of the Gaussian G(x), which amounts to varying the

104

Some 2-D wavelets and their performance

scale. Furthermore, given an image s, his technique consists in locating the maxima of ψ (1) % s, which are the zero-crossings of ∂x2 G % s, i.e., precisely the technique of Mallat [261,264]. In other words, Canny was very close to a primitive version of wavelet analysis! However, for computational reasons, one is forced to sample the orientation angles, keeping only a few values. But, as remarked by Perona in 1992 [311], . . . this practice has the strong drawback of introducing anisotropies and algorithmic difficulties in the computational implementations. It would be preferable to keep thinking in terms of a continuum, of angles for example, and to be able to localize the orientation of an edge with the maximum accuracy allowed by the filter one has chosen.

Perona’s answer to this objection is to advocate the use of steerable filters, described in Section 2.7. However, taking together Canny’s approach and Perona’s remark leads directly to the continuous wavelet transform, which thus appears as a very natural tool for edge detection, and more generally in computer vision. Coming back to Canny’s work, it is instructive to compare the first, second and third derivatives of the Gaussian (see [Bha99]; Chapter 4) and test their performance, exactly as in 1-D [49]. The Canny edge detector was later improved by Deriche [135,136] and Bourennane et al. [83], still keeping the same philosophy. The former gets for the optimal filter the function f (x) = −c e−α|x| sin ωx −c x e−α|x|

for

α/ω ! 1.

(3.16)

An alternative is to consider a mixed derivative, such as ∂x ∂ y exp(−| x |2 ). This wavelet has good capabilities for detecting corners in a contour, but we will describe in Section 3.3.3 another one that performs even better, the so-called end-stopped wavelet of [Bha99,76].

3.3.1.2

The concept of directional wavelets Although the directional derivative wavelets just described do have some capabilities of directional filtering, they are by far not sufficient, because their angular selectivity is rather poor. In order to go beyond, we introduce the concept of directional wavelets [18,19,24]: Definition 3.3.1 . A wavelet ψ is said to be directional if the effective support of its is contained in a convex cone in spatial frequency space {k}, with Fourier transform ψ apex at the origin, or a finite union of disjoint such cones (in that case, one will usually call ψ multidirectional). Since it may sound counter-intuitive, this definition requires a word of justification. According to (2.20), the wavelet acts as a filter in k-space (multiplication by the function Suppose the signal s( ψ). x ) is strongly oriented, for instance a long segment along the

105

3.3 Directional wavelets

is a long segment along the k y -axis. In order to s(k) x-axis. Then its Fourier transform detect such a signal, with a good directional selectivity, one needs a wavelet ψ supported k) is essentially aligned in a narrow cone in k-space. Then the WT is negligible unless ψ( not ψ. The onto s(k): directional selectivity demands restriction of the support of ψ, corresponding standard practice in signal processing is to design an adequate filter in the frequency domain (high pass, band pass, . . . ). In addition, there are cases (magnetic resonance imaging, for instance) where data are acquired in k-space (then called the measurement space) and the image space is obtained after a Fourier transform: here again directional filtering takes place in k-space. According to this definition, the anisotropic Mexican hat is not directional, since the H is centered at the origin, no matter how big its anisotropy is. Indeed, the support of ψ detailed tests described in [13] confirm its poor performances in selecting directions. We will come back to this point, with quantitative results, in Section 3.4. A review of directional wavelets and their use may be found in [19].

3.3.2

The 2-D Morlet wavelet This is the prototype of a directional wavelet: ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) − exp(− 12 |A−1 ko |2 ) exp(− 12 |A x|2 ), √ M (k) 2 )). = $ (exp(− 1 |A−1 (k − ko )|2 ) − exp(− 1 |A−1 ko |2 ) exp(− 1 |A−1 k| ψ 2 2 2

(3.17) (3.18)

The parameter ko is the wave vector, and A = diag[$ −1/2 , 1], $ 1, is a 2 × 2 anisotropy matrix. The correction term in (3.17) and (3.18) enforces the admissibility M (0) = 0. However, since it is numerically negligible for |ko | 5.6, one condition ψ usually drops it altogether (but not always, see Section 3.4). In that case, putting $ = 1, we obtain the function ψG ( x ) = exp(i ko · x) exp(− 12 | x |2 ).

(3.19)

This function is well-known in the image processing literature under the name of Gabor function [126]. One reason of its popularity is its computational simplicity. Another one, in particular in the modeling of human vision, is that a large fraction of cells in the primary visual cortex of primates (including man, presumably) have a receptive field that resembles a Gabor function [DeV88,369,370,372]. An example is shown, for $ = 2, ko = (0, 6), in Figures 3.3 and 3.4. The Gabor function (3.19) has the qualitative behavior expected from a wavelet. It is well localized, both in position space, around the origin, and in spatial frequency space, around k = ko = 0, but, strictly speaking, it is not admissible. On the other hand, the full Morlet function given in (3.17)–(3.18) is always admissible but, for small |ko |, it is M consists then essentially of two disjoint pieces. Thus the useless as a wavelet, since ψ Morlet function is interesting for practical wavelet analysis only for |ko | large enough.

106

Some 2-D wavelets and their performance

0.08 0.06 0.04 0.02 0 −0.02 −0.04 −0.06 −0.08 0 10 20 30 40

0

5

10

15

20

25

30

35

40

(a) 1 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 0 10 20 30 40

0

5

10

15

20

25

30

35

40

(b) Fig. 3.3. The 2-D Morlet wavelet ψM ( x ) (or rather ψG ( x )), with $ = 2, ko = (0, 6) in 3-D

perspective, in position domain: (a) real part; (b) imaginary part.

The Morlet wavelet is complex. The modulus of the truncated wavelet ψG is a Gaussian, elongated in the x direction if $ > 1, and its phase is constant along the direction orthogonal to ko and linear in x, mod(2π/|ko |), along the direction of ko . Thus, plotting the phase of ψG ( x ) as a function of x, we get a succession of straight lines, perpendicular to ko , and with intensity varying periodically and linearly from 0 to 2π . As compared to the 1-D case, the additional feature here is the inherent directivity of the wavelet ψG , entirely contained in its phase. This turns out to be a crucial ingredient in the study of directional features of objects (Chapters 4 and 5). Indeed, from the fact that the WT is a convolution of the signal with the dilated wavelet, we see that the wavelet ψG smoothes the signal in all directions, but detects the sharp transitions in the direction

107

3.3 Directional wavelets

100

100

200

200

300

300

400

400

500

500 100

200

300

400

500

100

200

(a)

300

400

500

400

500

(b)

100

100

200

200

300

300

400

400 500

500 100

200

300

400

500

100

200

(c)

300

(d)

Fig. 3.4. The 2-D Morlet wavelet ψM ( x ) with $ = 2, k0 = 6 in gray levels, in position domain: (a) real part; (b) imaginary part; (c) phase; and (d) modulus.

perpendicular to ko . In Fourier space, the effective support (“footprint”) of the function M is an ellipse centered at ko and elongated in the k y direction, thus contained in a ψ √ convex cone. Since the ratio of the axes is equal to $, the cone becomes narrower as $ increases. Clearly this wavelet will detect preferentially singularities (edges) in the x direction, and its angular selectivity increases with |ko | and with the anisotropy $. The best selectivity will be obtained by taking ko parallel to the long axis of the ellipse in k-space, that is, ko = (0, ko ). The Morlet wavelet ψM (or rather ψG ) then becomes (see Figures 3.3 and 3.4): ψG ( x ) = exp(iko y) exp[− 12 ( 1$ x 2 + y 2 )],

x = (x, y).

(3.20)

Many variants of the basic wavelets may be designed for specific problems. For instance, we know that the Mexican hat is very good at detecting discontinuities (e.g., edges) in an image, but it is not directional. On the other hand, the Morlet wavelet is directional, but mostly selective in spatial frequency (these statements will be proved in Section 3.4). Both properties may be combined in a single wavelet, the Gabor (or modulated) Mexican hat wavelet, defined as follows:

108

Some 2-D wavelets and their performance

ψGM ( x ) = −($

∂2 ∂2 2 1 1 2 + ) exp(ik y) exp[− ( x + y )] , $ 1, o 2 $ ∂x2 ∂ y2

= (2 − 1$ x 2 − (y − iko )2 ) exp(iko y) exp[− 12 ( 1$ x 2 + y 2 )], GM (k) = ψ

√

$($k x2 + k 2y ) exp[− 12 ($k x2 + (k y − ko )2 )].

(3.21) (3.22) (3.23)

Notice that no correction term is needed here, the function ψGM is admissible as it stands. This wavelet, introduced in [289], is extremely efficient in detecting edges, even in the presence of heavy noise. One of its possible applications may be character recognition (see Section 4.1). Going back to the Gabor or (truncated) Morlet wavelet (3.20), we notice that it has the general form ψD ( x ) = φ(x) ψ(y),

x = (x, y),

(3.24)

where φ is a 1-D scaling function (a bump, typically a Gaussian) and ψ is a 1-D wavelet. Functions of this type provide an easy way to design a separable, yet directional wavelet; in the present case, a horizontal one. This technique is due to Bournay Bouchereau [Bou97] (see Section 5.2.2). A related example is the Gabor-like wavelet of Unser [357], obtained by replacing the Gaussian in (3.20) by another window function, typically a B-spline.

3.3.3

End-stopped wavelets A basic problem in the characterization of an image, for instance, in the comparison of two images, is the identification of specific features. In human vision, this is achieved by the so-called saccadic movement of the eyes: the eyes scan the scene freely, momentarily focus on some point of interest, and quickly move on to the next target point in the scene [Yar67]. The visual jump from one target to another is called a saccade, the target points of consecutive saccades being points of interest which stand out against the general background of the scene. This process has been analyzed in detail and modeled with wavelets by Bhattacharjee [Bha99], that we now quote. . . . There is evidence that such features of interest are identified by the lower levels of the visual system, and are not a result of conscious reasoning. Psychovisual studies on several mammals show the presence of certain cells, called hypercomplex cells or end-stopped cells, in the primary visual cortex. End-stopping behavior is related to oriented linear stimuli, that is, end-stopped cells are activated, under certain conditions, by linear stimuli having a particular orientation. Two kinds of end-stopping behavior have been identified. The single end-stopped cells respond strongly if a line in a particular orientation ends within the receptive field of the cell. For real-world scenes, these cells respond strongly to corners, or points of high curvature in general. Some other cells respond strongly only to short, oriented, linear stimuli. Such cells are called double or complete end-stopped cells. For cells of this kind, the response is strong as long as the stimulus has a specific orientation (the characteristic

3.3 Directional wavelets

orientation varies from cell to cell), and both ends of the stimulus lie within the receptive field of the cell. Thus, double end-stopped cells respond to short linear segments in images.

In order to model faithfully this physiological behavior, Bhattacharjee [76,Bha99] introduces two specific wavelets, called end-stopped wavelets, that we now describe. According to the discussion above, the responses of end-stopped cells are related to end points of linear structures lying in a specific orientation. Thus, end-stopping can be simulated by isolating linear structures in the image that have a particular orientation, and then processing these structures further to locate their end points, or to determine their lengths. The first stage, to detect lines having a specific orientation, can be achieved with a Morlet (more properly, a Gabor) wavelet. Then, the end-points of a line can be detected by applying the first derivative of a Gaussian filter along the line. Combining the two operations yields the end-stopped wavelet ψE1 , namely ψE1 ( x) =

1 4

x exp[− 14 {(x 2 + y 2 ) + ko (ko − 2i y)}],

(3.25)

in position space, and E1 (k) = exp[− 1 (k x2 + (k y − ko )2 )] −ik x exp[− 1 (k x2 + k 2y )] ψ 2 2

(3.26)

E1 is the product of in spatial frequency space. Equation (3.26) clearly shows that ψ two components. The first factor is a Morlet wavelet, with wave vector ko = (0, ko ) oriented along the k y -axis. The second factor is a first derivative of a Gaussian oriented along the frequency axis k x , that is, in the direction perpendicular to the orientation of the Morlet wavelet. On the other hand, if one regroups the two Gaussian factors in (3.26), the result √ is then simply the derivative in the k x -direction of a Morlet wavelet of width σ = 1/ 2 and wave vector 12 ko (up to a multiplicative constant). An example is shown, in spatial frequency space, in Figure 3.5.

−3 −2 −1

ky

109

0

Ky

1

2

3

−3

−2

−1

0

1

2

3

Kx

kx

(a)

(b)

E1 (k) in spatial frequency space: (a) in gray levels; (b) in 3-D Fig. 3.5. The end-stopped wavelet ψ perspective.

110

Some 2-D wavelets and their performance

As we will see in Section 4.1, the end-stopped wavelet ψE1 performs extremely well for detecting corners in an image. This can be understood intuitively as follows. AccordE1 may be interpreted as the difference between two Morlet wavelets. ing to the figure, ψ Thus an infinite rod oriented along the k y -axis will not be seen, but if we take a finite segment, the end-points will appear. On the other hand, if the rod or the segment are slightly misaligned, they will be seen. Thus the end-stopped wavelet ψE1 detects direction by a zero-crossing, hence it has a better resolution compared to the plain Morlet wavelet. If one wants to detect, in addition, the size l of a linear structure, oriented in a certain direction θ , one may use the double end-stopped wavelet, ψE2 . This wavelet indeed produces a strong response at the center of any linear structure that is oriented at θ and has a length close to l. Compared with ψE1 , the new wavelet is obtained by replacing the first derivative of a Gaussian by a second derivative, that is, a Mexican hat. The latter is used to determine the range of lengths to which the wavelet should respond. Its efficiency is easy to understand. Quoting [Bha99] again, . . . The Mexican hat is characterized by an excitatory central region surrounded by an inhibitory region. In 1-D, consider a linear stimulus of length l = l1 , and a Mexican hat filter which has an excitatory region of length l > l1 . The filter will produce the strongest response when it is exactly centered over the stimulus. Moreover, as l increases, the maximum response of the filter will initially increase. This response will increase as long as the length l l . After l exceeds l , a part of the stimulus will fall in the inhibitory region of the filter, thus damping the response of the filter. Thus, given several linear stimuli of different lengths, the Mexican hat filter responds most strongly to the stimulus which is closest to (but not greater than) the width of the excitatory portion of the filter.

Thus we obtain the required wavelet by taking, as before, a Morlet wavelet of unit width, with wave-vector ko = [0, ko ] oriented along the k y -axis, multiplied by a Mexican hat of width σ , oriented along the k x -axis, i.e., perpendicular to the orientation of the Morlet factor. This yields for the double end-stopped wavelet ψE2 the function 2σ 2 (x 2 + σ 2 + 1) σ 2 (x 2 + y 2 ) + ko (ko − 2iσ 2 y) ψE2 ( , (3.27) x) = exp − (σ 2 + 1)3 2(σ 2 + 1) E2 (k) = − 1 (k x2 − σ 2 ) exp[− 1 2 (k x2 + k 2y )] exp[− 1 (k x2 + (k y − ko )2 )]. ψ 2 2σ 8σ4 As for the pure Mexican hat, the width parameter σ allows us to control the resolving power of the wavelet. The two end-stopped wavelets, ψE1 and ψE2 , have been used successfully in [Bha99] for the detection of characteristic features in images, in the general context of image retrieval. We will discuss this application in more details in Section 4.3.

3.3.4

Conical wavelets In order to achieve a genuinely oriented wavelet, it suffices to consider a smooth function (C) (k) with support in a strictly convex cone C in spatial frequency space and behaving ψ inside C as P(k1 , . . . , kn )e−ζ ·k , with ζ ∈ C and P(.) denotes a polynomial in n variables.

111

3.3 Directional wavelets

Alternatively one may replace the exponential by a Gaussian, which gives a better localization in spatial frequency. In both cases, the resulting wavelets will be called conical. We begin with the former case, thus obtaining the class of Cauchy wavelets [18,19,24]. For simplicity, we consider a strictly convex cone, symmetric with respect to the positive k x -axis, namely C ≡ C(−α, α) = {k ∈ R2 | − α arg k α, α < π/2}, that is, the convex cone determined by the unit vectors e−α , eα . The dual cone, with sides perpendicular to those of the first one, is also convex and reads: C = C(−α, ˜ α) ˜ = {k ∈ R2 | k · k > 0, ∀ k ∈ C(−α, α)}, where α˜ = −α + π/2. Therefore e−α · eα˜ = eα · e−α˜ = 0. Given the fixed vector η = (η, 0), η > 0, we first define the (symmetric) Cauchy wavelet in spatial frequency variables: η m m −k· , k ∈ C(−α, α) m(α) (k) = (k · eα˜ ) (k · e−α˜ ) e (3.28) ψ 0, otherwise. m(C) (k) is strictly supported in the cone C(−α, α) and the parameter The Cauchy wavelet ψ ∗ on the edges of the cone, m ∈ N , m 1, gives the number of vanishing moments of ψ and thus controls the regularity of the wavelet. An interesting aspect of this wavelet is that its inverse Fourier transform may be calculated exactly [24], with the result: ψm(α) ( x) =

(−1)m+1 (sin 2α)2m+1 (m!)2 $ %m+1 , 2π z · σ (α)z

(3.29)

where we have introduced the complex variable z = x + i η ∈ R2 + i C and the 2 × 2 matrix cos2 α 0 σ (α) = . 0 − sin2 α Indeed, from the definition (3.28), we get: 1 (α) x) = d 2 k ei k·x (k · eα˜ )m (k · e−α˜ )m e−k·η ψm ( 2π C(−α,α) (−1)m x ]m [e−α˜ · ∇ x ]m = d 2 k e−k·(η−i x) . [eα˜ · ∇ 2π C(−α,α) The integral on the right-hand side is convergent, since k · η > 0. Write ξ = η − i x = −i( x + i η) and let A be the matrix that maps the unit vectors e1 , e2 onto e−α , eα , respectively: (eν )i = Aν j (e j )i , ν = ±α (where we use the usual summation convention), so that k j = Aν j k ν (contravariant coordinates) and ξν = Aν j ξ j (covariant coordinates). Explicitly, we have:

112

Some 2-D wavelets and their performance

A=

cos α cos α − sin α sin α

,

so that

det A = sin 2α.

In the new (nonorthogonal) coordinates, the cone becomes C(−α, α) = {k ∈ R2 : k ν 0, ν = ±α}, and the integral may be evaluated immediately: ξ 2 −k· d ke = dk 1 dk 2 exp(−Aν j k ν ξ j ) C(−α,α) C(−α,α) ∞ ∞ dk α dk −α exp(−k ν ξν ), = det A 0

0

det A = ξα ξ−α sin 2α = (eα · ξ ) (e−α · ξ ) − sin 2α . = [( x + i η) · eα ] [( x + i η) · e−α ] Then the result follows by differentiation, if one remembers that eα˜ · e−α = e−α˜ · eα = 0, eα˜ · eα = e−α˜ · e−α = sin 2α. Indeed: x ) (eα˜ · ∇

eα˜ · e−α 1 = = 0, [( x + i η) · e−α ] [( x + i η) · e−α ]2

x )m (eα˜ · ∇

(−1)m m! (eα˜ · eα )m 1 = [( x + i η) · eα ] [( x + i η) · eα ]m+1 =

(−1)m m! (sin 2α)m , [( x + i η) · eα ]m+1

and similarly for the other factor. Thus one obtains as the final result the function ψm(α) given in (3.29). Clearly, this function is square integrable, and admissible in the sense of (6.15), in other words, it is a wavelet. We show in Figures 3.6 and 3.7 various aspects of the Cauchy wavelet ψ4(10) . This is manifestly a highly directional filter, strictly supported in the cone C = C(−10◦ , 10◦ ). Notice the slow decay in x-space of the conical wavelet ψm(α) , independent of α. This is the price to pay for forcing the wavelet to be strictly supported in a cone, and is to be expected in the light of standard results on the localization properties of wavelets (theorems of Balian–Low and Battle) [Fei98,36]. However, as we have already stressed in Section 3.3.1.2, this is irrelevant for the analysis, only the behavior in k-space counts. The construction generalizes in a straightforward way [24] to an arbitrary convex ˜ = {k ∈ R2 , k · cone C ≡ C(α, β) = {k ∈ R2 | α arg k β}, with dual C ≡ C(α, ˜ β)

k > 0, ∀ k ∈ C(α, β)}, where β˜ = α + π/2, α˜ = β − π/2, and arbitrary vanishing

113

3.3 Directional wavelets

(10)

Fig. 3.6. The Cauchy wavelet ψ4 : (a) real part; (b) imaginary part; (c) modulus, all in position

domain; and (d) modulus in frequency domain.

the resulting wavelet reads moments l, m > 0 on the cone edges. For any fixed η ∈ C, in spatial frequency space as (k · eα˜ )l (k · eβ˜ )m e−k·η , k ∈ C(α, β); (C) (3.30) ψlm (k) = 0, otherwise, and in position space,

114

Some 2-D wavelets and their performance

(10)

Fig. 3.7. The Cauchy wavelet ψ4 , in gray levels, in position domain: (a) real part; (b) imaginary

part; (c) phase; and (d) modulus. (C) ψlm (x) = const. (z · eα )−l−1 (z · eβ )−m−1 ,

(3.31)

where again z denotes the complex variable z = x + i η ∈ R2 + i C. Actually the origin of the name “Cauchy” is the following example. For α = 0, β = π/2, η = eπ/4 and m = 1, one gets: ψ1(C) (x) =

1 (1 − i x)−2 (1 − i y)−2 , 2π

(3.32)

i.e., the product of two 1-D Cauchy wavelets [Hol95]; that is, derivatives of the Cauchy kernel (z − t)−1 . Of course, this example is of little use in practice. Indeed, the main interest of Cauchy wavelets is their good angular selectivity, which requires a narrow

115

3.3 Directional wavelets

20

20

15

15

10

10

5

5

0

0

−5

−5

−10

−10

−15

−15

−20 −20

−15

−10

−5

0

5

10

15

20

−20 −20

−15

−10

−5

0

5

10

15

20

(b)

(a)

M Fig. 3.8. Two directional 2-D wavelets, in k space, seen in level curves: (a) the Morlet wavelet ψ ◦

4(10 ) with supporting cone C10 = C(−10◦ , 10◦ ), (ko = (0, 6), $ = 5); (b) the Cauchy wavelet ψ ◦ rotated by 90 for the sake of comparison.

cone. For applications, it turns out that the wavelet ψ4(10) , with support in the cone C10 = C(−10◦ , 10◦ ) has properties very similar to those of the Morlet wavelet (3.17) with |ko | = 5.6, except that here the opening angle of the cone is totally controllable. For a Morlet wavelet, on the contrary, the cone gets narrower for increasing |ko |, but then the amplitude decreases as exp(−|ko |2 ). In that sense, Cauchy wavelets are better adapted. M with ko = (0, 6), $ = 5 (left) We show side by side in Figure 3.8 the Morlet wavelet ψ (10) ◦ (k), rotated by 90 for the sake of comparison (right). and the Cauchy wavelet ψ 4 Quantitative comparisons will be made in Section 3.4. In 1-D, a wavelet ψ is called progressive or a Hardy function [Hol95,205], if ψ(ω) = 0 for ω < 0. This in turn may be expressed in terms of the Hilbert transform, defined by H f (ω) = −i sign ω f (ω), namely ψ = (1 + i H )φ, φ ∈ L 2 (R, dt) (that is, ψ is the analytic signal associated to φ). Equivalently, ψ belongs to the Hardy space H+2 (R) of square integrable functions which extend analytically into the upper half-plane. We claim that the conical wavelets are the 2-D analogs of this concept; that is, the genuine 2-D progressive wavelets. In order to prove that statement, we first notice that the convex cone C(α, β) may ν = α, β: also be expressed in terms of the covariant coordinates kν˜ = (eν˜ · k), C(α, β) = {k ∈ R2 : kα˜ 0, kβ˜ 0}. Consider the directional Hilbert transforms:

(3.33)

116

Some 2-D wavelets and their performance

H ν˜ f (k) = −i sign kν˜ f (k).

(3.34)

Given φ ∈ L 2 (R2 , d 2 x), define the function ψ = (1 + i Hα˜ + i Hβ˜ − Hα˜ Hβ˜ )φ = (1 + i Hα˜ )(1 + i Hβ˜ )φ.

(3.35)

k) vanishes outside the cone C(α, β), and Then it is easy to see, as in [366], that ψ( indeed: k) = 4φ(k), k ∈ C(α, β), ψ( (3.36) 0, otherwise. Therefore the inverse Fourier transform ψ( x ) is the boundary value of a function ψ(z ) 2 holomorphic in the tube R + i C; i.e., a 2-D Hardy function. For a fixed convex cone 2 C(α, β), the set of all such functions constitutes a Hilbert space, naturally denoted H(α,β) , which is unitary equivalent, via the complex Fourier transform, to the space β), d 2 k) [Ste71; Theorem VI.3.1]. In that sense, conical wavelets are a genuine L 2 (C(α, multidimensional generalization of the 1-D Hardy functions, much more so than the so-called 2-D Hardy functions defined by Dallard and Spedding (in particular, their “Arc” wavelet) [117]. Among them Cauchy wavelets are particularly simple (they occupy a special niche, as we will see in Section 8.2). In conclusion, notice that we are 2 talking here of a Hardy space H(α,β) , but similar considerations may be made for Hardy 1 spaces H(α,β) , in terms of the Riesz operators, which are a natural multidimensional generalization of the Hilbert transform [366]. Cauchy wavelets have a good angular selectivity, provided one chooses a narrow cone. However their radial selectivity is not terribly good, because the exponential decays → ∞. In order to obtain a better radial localization, one may replace rather slowly as |k| the exponential by a Gaussian in√k x [Vdg98]. This has the effect to concentrate the wavelet on its central frequency ( 2m, 0). The resulting wavelet is called the Gaussian conical wavelet. The angular selectivity of ψ is specified by the angular aperture of the cone and is well controlled by the parameter α. However, the radial selectivity is only roughly fixed by the moment number m. The resulting wavelet is very similar to the Cauchy wavelet, except that it is more concentrated in spatial frequency space, since it is also localized in scale, around the central scale ao . However, although the pure Gaussian is well peaked, the addition of a large number of vanishing moments tends to spread it. Thus, one can achieve an even better scale localization by using an appropriate width σ > 0 for the Gaussian. We may still improve on this by adding another parameter χ(σ ) > 0, whose sole rˆole is to control the radial support of ψ [26,27]. This leads to the following formula for our conical wavelet: σ 2 (k · e−α˜ )m (k · eα˜ )m e− 2 (kx −χ(σ )) , k ∈ C(−α, α), ψC (k) = (3.37) 0, otherwise,

117

3.3 Directional wavelets

Fig. 3.9. The Gaussian conical wavelet (3.37), in frequency space, for m = 4, α = 10◦ , σ = 1.

√ −1 where χ(σ ) = 2m σ√ is called the center correction term. Notice that the central σ frequency is the point ( 2m, 0) for any σ , and for σ = 1, one recovers the pure Gaussian conical wavelet. This is the wavelet we will mostly use in the sequel. It is shown in frequency domain, in Figure 3.9. It is clear on this picture why this wavelet is sometimes called the shark wavelet! Another alternative is the conical Mexican hat, introduced by Murenzi et al. [289]. and l, m ∈ N ∗ , as The wavelet with support in C = C(β) is defined, for any η ∈ C (C) (k) ψ lm

=

1 2 2 (k · e−β˜ )m (k · eβ˜ )l ($k x2 + k 2y )e− 2 ($kx +k y ) , 0, otherwise.

k ∈ C(β)

(3.38)

Besides these directional wavelets, there exist two other tools especially designed for the detection of lines or curves, called ridgelets and curvelets, respectively. These actually define new transforms, that we will discuss later, in Section 11.1.

3.3.5

Multidirectional wavelets Given a directional wavelet ψ, as above, it is easy to build a multidirectional one, with n-fold symmetry simply by superposing n suitably rotated copies of ψ: x) = ψn (

n−1 1 2π ψ(r−θk ( x )), θk = k , k = 0, 1, . . . , n − 1. n k=0 n

(3.39)

118

Some 2-D wavelets and their performance

Taking, for instance, n = 4 and for ψ a Gabor (truncated Morlet) wavelet, we get the following real wavelet with four-fold symmetry: ψ4M (x, y) =

1 1 2 2 (cos ko x + cos ko y) e− 2 (x +y ) . 2

(3.40)

This wavelet filters out all features which are not primarily horizontal or vertical. In the same way, one gets wavelets with symmetry 6 or 10, which may find applications, respectively, in biological problems or the analysis of quasicrystals. In general, multidirectional wavelets should be useful for pattern recognition. Notice that a similar construction was proposed by Watson [369,370]. His fan filters are obtained by taking first the difference between two “mesa” functions, which yields an annular wavelet, and then repeatedly bisecting the spatial frequency space and taking only one side (i.e., the associated analytic signal). The allowed directions θ are thus restricted to a fan-shaped region: 0 2θ

2π (n = 2, 3, . . . ). 2n−1

(3.41)

This construction may then be generalized to arbitrary angles [306]. Actually, a comparison with Section 2.6 immediately shows that the construction of directional wavelet packets is based on the very same idea. These fan filters have all the properties of directional wavelets, including admissibility in the form (2.17). Applying on these filters discrete rotations and scaling, Watson builds a pyramid of oriented filters as a tool for data compression and signal reconstruction after coding, in a model of human vision. This is in fact a discretized version (in polar geometry) of the CWT. Another example, very similar to the previous one, is that of the steerable filters, described in Section 2.7.

3.4

Wavelet calibration: evaluating the performances of the CWT Given a wavelet, what is its angular and scale selectivity (resolving power)? What is the minimal sampling grid for the reconstruction formula (2.26) that guarantees that no information is lost? The answer to both questions resides in a quantitative knowledge of the properties of the wavelet, that is, the tool must be calibrated. To that effect, one takes the WT of particular, standard signals. Three such tests are useful, and in each case the outcome may be viewed either at fixed (a, θ) (position representation) or at fixed b (scale-angle representation). r Point signal: for a snapshot at the wavelet itself, one takes as the signal a delta function, i.e., one evaluates the impulse response of the filter: ψa,θ,b |δ = a −1 ψ(a −1r−θ (−b)).

(3.42)

119

3.4.1

3.4 Wavelet calibration: evaluating the performances of the CWT r

Reproducing kernel: taking as the signal the wavelet ψ itself, one obtains the reproducing kernel K , which measures the correlation length in each variable a, θ, b : −1 ψ( cψ K (a, θ, b|1, 0, 0) = ψa,θ,b |ψ = a ψ(a −1r−θ ( x − b)) x ) d 2 x. (3.43)

r

Benchmark signals: for testing particular properties of the wavelet, such as its ability to detect a discontinuity or its angular selectivity in detecting a particular direction, one may use appropriate “benchmark” signals.

The scale and angle resolving power Suppose the wavelet ψ has its effective support in spatial frequency in a vertical cone in the x and y directions of aperture ϕ, corresponding to ko = (0, ko ). The width of ψ is given by 2wx and 2w y , respectively: !1/2 !1/2 1 1 k)| k)| 2 2 d 2 k k x2 |ψ( d 2 k (k y − ko )2 |ψ( , wy = . wx = ψ ψ (3.44) is concentrated in an ellipse of semi-axes wx , w y , and its radial Then the wavelet ψ ko + w y . Thus the scale width or scale resolving power support is ko − w y |k| (SRP) of ψ is defined as: S R P(ψ) =

ko + w y . ko − w y

(3.45)

In the same way, one defines the angular width or angular resolving power (ARP) by considering the tangents to that ellipse. Then a straightforward calculation yields: ) ko2 − w 2y −1 A R P(ψ) = 2 cot ϕ. (3.46) wx For instance, if ψ is the (truncated) Morlet wavelet (3.17), one obtains: √ ) ko 2 + 1 , A R P(ψM ) = 2 cot−1 $(ko2 − 1), S R P(ψM ) = √ ko 2 − 1

(3.47)

and, for ko ! 1: √ A R P(ψM ) = 2 cot−1 (ko $).

(3.48)

This last expression coincides with the empirical result of [13]: the angular sensitivity √ of ψM depends only on the product ko $. Notice also that the SRP is independent of the anisotropy factor $. If ψ is the Cauchy wavelet (3.28) with support in the cone C(−α, α), the ARP is simply the opening angle 2α of the supporting cone.

120

Some 2-D wavelets and their performance

Fig. 3.10. Filter bank obtained with the Morlet wavelet (5 scales, 16 orientations).

3.4.2

The reproducing kernel and the resolving power of the wavelet A natural way of testing the correlation length of the wavelet is to analyze systematically its reproducing kernel. Let the effective support of the wavelet ψ in spatial frequency be, in polar coordinates, ρ and ϕ. Then an easy calculation [18] shows that the effective support of K is given by amin = ( ρ)−1 a amax = ρ for the scale variable, and − ϕ θ ϕ for the angular variable. Thus we may define the wavelet parameters (or resolving power) ρ, ϕ in terms of the parameters a, θ of K , as: √ r scale resolving power (SRP): ρ = a = √a max /amin ; r angular resolving power (ARP):

ϕ = 12 θ. a j ,θ (k)}, which yields a complete In this way, one may design a wavelet filter bank {ψ tiling of the spatial frequency plane, in polar coordinates [15,18]. An example is given in Figure 3.10, for the case of the Morlet wavelet. Clearly this analysis is only possible within the scale-angle representation. Thus it requires the use of the CWT, and it is outside of the scope of the DWT, which is essentially limited to a Cartesian geometry (see Section 2.6).

3.4.3

Calibration of a wavelet with benchmark signals The capacity of the wavelet at detecting a discontinuity may be measured on the (benchmark) signal consisting of an infinite rod (see [13] for the full discussion). The result is that both the Mexican hat and the Morlet wavelet are efficient in this respect. For testing the angular selectivity of a wavelet, one computes the WT of a semi-infinite rod, sitting along the positive x-axis, and modeled as usual with a delta function: s( x ) = ϑ(x) δ(y),

(3.49)

121

3.4 Wavelet calibration: evaluating the performances of the CWT

1 θ = 0° 0.9

0.8

θ = 5°

0.7

0.6

0.5

0.4

0.3

0.2

θ = 10°

0.1 θ = 90° 0 −5

0

θ = 20° 5

10

Fig. 3.11. Angular selectivity of the Morlet wavelet for different values of the orientation angle: θ = 0◦ , 5◦ , 10◦ , 20◦ , 45◦ and 90◦ . The graph shows the modulus |S((bx , 0), 1, θ)| as a function of bx .

where ϑ(x) is the step function. Plugging this expression into the definition of the transform yields (we take a = 1 for simplicity): +∞ 1, θ ) = d x ψ r−θ (x − bx , −b y ) . (3.50) S(b, 0

Let us take first a Morlet wavelet with $ = 5, oriented at an angle θ, and compute the CWT of s as a function of bx . As illustrated by Figure 3.11, the result is that ψM detects the orientation of the rod with a precision of the order of 5◦ . Indeed, for θ < 5◦ , the WT is a “wall,” increasing smoothly from 0, for x −5, to its asymptotic value (normalized to 1) for x 5. Then, for increasing misorientation θ, the wall gradually collapses, and essentially disappears for θ > 10◦ . Only the tip of the rod remains visible, and for large θ (θ > 45◦ ), it gives a sharp peak. Essentially the same result is obtained with a Cauchy wavelet supported in the cone C(−10◦ , 10◦ ), of opening angle A R P = 20◦ , as shown in Figure 3.12(a). Conversely, one sees in panel (b) that, for a fixed misorientation angle θ = 20◦ , the Cauchy wavelet yields the same selectivity for A R P 20◦ . On the contrary, as observed on Figure 3.13, the same test performed with an anisotropic Mexican hat gives a result almost independent of θ. Even varying the anisotropy factor $ doesn’t really change the result: the discontinuity is detected by a sharper variation, but the sensitivity to its orientation is not greatly improved. The conclusion is that the Morlet and the Cauchy wavelets are highly sensitive to orientation, but the anisotropic Mexican hat is not.

122

Some 2-D wavelets and their performance

| S(b ,1,20°) |

| S(bx ,1,θ) | 1

x

misorientation angle

θ = 20

°

1 Φ = 30 °

θ = 0°

0.9

0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

Φ = 25 °

0.2

0.2

θ = 20 °

0

Φ = 10 °

0.1

0.1

−5

θ = 90 0

°

Φ = 20 °

θ = 45 ° 5

bx

(a)

10

0

−5

0

5

(b)

bx

10

Fig. 3.12. Angular selectivity of the Cauchy wavelet ψ1,1 : (a) for a cone of fixed width = 2α = 20◦ and for different values of the orientation angle: θ = 0◦ , 5◦ , 10◦ , 20◦ , 45◦ and 90◦ ; (b) for a fixed value of the misorientation angle θ = 20◦ and various values of the ARP = 2α. The graph shows the modulus |S((bx , 0), 1, θ )|, respectively |S((bx , 0), 1, 20◦ )|, as a function of bx . 1

0.9

1 ε=5

θ = 0°

θ = 90°

0.9

ε = 2.5

θ = 45°

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0 −5

0

5

(a)

ε=1

10

0 −5

0

5

10

(b)

Fig. 3.13. Angular selectivity of the anisotropic Mexican hat wavelet with: (a) anisotropy $ = 5 and

for different values of the orientation angle (θ = 0◦ , 45◦ and 90◦ ) and (b) for a fixed orientation (θ = 0◦ ) but various anisotropy factors ($ = 1, 2.5 and 5).

Let now the signal be a segment. If one uses a Morlet or a Cauchy wavelet as above, the WT reproduces the segment if the misorientation φ between the signal and the wavelet is smaller than 5◦ , but the segment becomes essentially invisible for

φ > 15◦ , except for the tips (these are point singularities). In the end, the image of the segment reduces to two peaks corresponding to the two endpoints (see Figure 3.14).

123

3.4 Wavelet calibration: evaluating the performances of the CWT

1 θ = 0° 0.9

0.8 θ = 90°

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 −5

−4

−3

−2

−1

0

1

2

3

4

5

Fig. 3.14. Modulus of the wavelet transform of a segment positioned at θ = 0◦ , using a Morlet wavelet for several values of the orientation angle θ.

This is exactly the property used crucially in the measurement of the velocity field of a turbulent fluid (see Section 5.3.3 below). One may remark that the precision mentioned here is obtained with the modulus of the WT. In fact, if the wavelet is complex (like ψM ), one may also exploit the phase of the WT, and it gives a higher precision yet [13]. But this is practical only on academic signals, real data are in general too noisy and only the modulus is useful. Another way of comparing the angular selectivity of the two wavelets is to analyze a directional signal in the angle–angle representation (α, θ ) described above. The result confirms the previous one [17]. In order to illustrate the difference in angular selectivity between the anisotropic Mexican hat and the Morlet wavelet, we analyze a directional signal with both of them and view the transform in the angle–angle representation described in Section 2.3.3. The result is shown in Figure 3.15. The signal is a rectangular slab of size 3 × 2, positioned radially at π/2 and it is analyzed with an anisotropic Mexican hat with $ = 2 (left) and a Morlet wavelet (right). The figures show the modulus of the CWT, at range = 3 and scale a = 1, in the angle–angle representation (α, θ ). For the Mexican hat, |b| the transform exhibits a maximum located at α = π/2, θ = π/2, as it should, and the graph is periodic both in α and in θ. For the Morlet wavelet, several maxima can be distinguished, all located around α = π/2. A careful inspection shows that the strongest peaks at θ = π/2 and 3π/2 correspond to the two longest edges of the rectangle while the smaller peaks at θ = 0, π, 2π correspond to the smaller edges. Notice however that, for the Mexican hat, the peak is sharp in α and quite broad in θ , as expected, since this wavelet has a very good resolution in position (α), but a rather poor one in directional

124

Some 2-D wavelets and their performance

6

0.4

5 0.3 4

3

0.2

2 0.1 1

0

6

0

6

6 6

4

4

4

4

0

(a)

0

(b) 6

5

5

4

4

α

α

6

3

3

2

2

1

1

0

θ

2

2

θ

0

0

α

α

2

2

0

1

2

3

θ (c)

4

5

6

0

0

1

2

3

θ (d)

4

5

6

Fig. 3.15. CWT of a slab of size 3 × 2 positioned radially at 90◦ , analyzed with an anisotropic

Mexican hat (left) with $ = 2 and a Morlet wavelet (right), both at scale a = 1. The figures in the top row give a 3-D perspective view of the transform, in the angle–angle representation (α, θ), at = 3. The same result is shown in level curves in the figures of the bottom row. range |b|

selectivity (θ). In addition, the contour of the slab can be precisely detected by tracking the zero crossings of this representation. On the opposite, the Morlet wavelet gives a sharp peak in both variables, since it has good selectivity in both [13], but the contour is less visible. This example illustrates Daugman’s theorem [126], and at the same time the usefulness of the angle–angle representation.

4

Applications of the 2-D CWT. I: image processing

The 2-D CWT has been used by a number of authors, in a wide variety of problems [Com89,Mey91,Mey93]. In all cases, its main use is for the analysis of images, since image synthesis or compression problems are rather treated with the DWT. In particular, the CWT can be used for the detection or determination of specific features, such as a hierarchical structure, edges, filaments, contours, boundaries between areas of different luminosity, etc. Of course, the type of wavelet chosen depends on the precise aim. An isotropic wavelet (e.g. a Mexican hat) often suffices for pointwise analysis, but a directional wavelet (e.g. a Morlet or a conical wavelet) is necessary for the detection of oriented features in the signal. Somewhat surprisingly, a directional wavelet is often more efficient in the presence of noise. In the next two chapters, we will review a number of such applications, including some nonlinear extensions of the CWT. First, in the present chapter, we consider various aspects of image processing. Then, in Chapter 5, we will turn to several fields of physics where the CWT has made an impact. Some of the applications are rather technical and use specific jargon. We apologize for that and refer the reader to the original papers for additional information.

4.1

Contour detection, character recognition

4.1.1

The detection principle Exactly as in the 1-D case, the WT is especially useful to detect discontinuities in images, for instance the contour [Mur90,13] or the edges of an object (which are discontinuities in the luminosity), and in particular its corners [204,266]. For that purpose, one may first ignore the directions and perform a pointwise analysis. Then, the simplest choice is an isotropic wavelet, such as the radial Mexican hat ψH given in (3.6) or (3.8). In that particular case, the effect of the WT consists in smoothing the signal with a Gaussian and taking the Laplacian of the result. Thus large values of the amplitude will appear at the location of the discontinuities, in particular the contour of objects. In particular, the corners of the contour, which are point singularities, will be highlighted. In addition, if

125

126

Applications of the 2-D CWT. I: image processing

the wavelet is real, the CWT detects the convexity of each corner: a convex corner gives rise to a sharp positive peak, whereas a concave one yields a negative peak. However, it turns out that other wavelets are useful too. For instance, in the presence of heavy noise, directional wavelets outperform the Mexican hat. On the other hand, if only the corners of the contour are needed, as in character recognition, mixed derivatives of the Gaussian or, even better, the end-stopped wavelets become the first choice. In order to test these properties on a concrete example, we begin by analyzing a simple geometric object, namely, a set with the shape of the letter L, represented by its characteristic function. Thus our test image is the white L-shaped region against a black background, the signal presented in Figure 4.1(a) or, equivalently, in Figure 4.2(a). First, we analyze the effect of scaling on the WT, i.e., going to finer and finer scales. For a pointwise analysis, we choose the isotropic Mexican hat ψH given in (3.6) [13]. The CWT is plotted in Figure 4.1 for three values of the scale parameter a (conveniently taken as powers of 2), a = 2 j , j = 3, 2, 1. Each WT is plotted both in 3-D perspective and in level curves. From these pictures, the following observations can be made. For a large value of a, the WT sees only the object as a whole, thus allowing the determination of its position in the plane. When a decreases, increasingly finer details appear. In this simple case, the WT vanishes both inside and outside the contour, since the signal is constant there. Eventually, only the contour remains and it is perfectly seen at a = 2. This is the analog of the precise localization of discontinuities in 1-D [204]. Of course, if one takes values of a that are too small, numerical artifacts (aliasing) appear and spoil the result. This is only a numerical limitation, however, that could be improved by a finer discretization (but with a longer computing time). We notice that the exterior contour is a sharp negative “wall,” whereas the interior contour is a positive one. The same effect would appear in 1-D if one would consider, for instance, the full WT of a square pulse. The jump from 0 to 1 gives a negative minimum followed by a sharp positive maximum, and the jump from 1 to 0 gives the opposite pattern. Note also that the corners of the figure are highlighted in the WT by sharp peaks. The amplitude is larger at these points, since the signal is singular there in two directions, as opposed to the edges. In addition the WT detects the convexity of each corner. The six convex corners give rise to positive peaks, whereas the concave one yields a negative peak. Here we see again the advantage of using a real wavelet and plotting the WT itself, not its modulus, which is a frequent practice. The conclusion is that the CWT is an efficient edge detector, provided it is evaluated at a scale that is sufficiently small (i.e., high spatial frequency), but still avoiding numerical artifacts (aliasing). The next step is to compare the performances of various wavelets on the same signal, and the L-shape is an ideal benchmark for making the comparison. The result is shown in Figure 4.2 (taken from [Bha99]). Panel (a) gives the signal. Panel (b) is the response of the isotropic Mexican hat wavelet (this is in fact the negative of Figure 4.1(d)!). As noted there already, the response of the wavelet is strongest at the corners, but is also

127

4.1 Contour detection, character recognition

(a)

4.1(a)

4.1(b)

(b)

4.1(b)

4.1(c)

(c)

4.1(c)

(d) Fig. 4.1. The L-shape and its CWT, obtained with an isotropic Mexican hat, presented in 3-D

perspective (left column) and in level curves (right column), at three successive scales: (a) the signal; (b) a = 8; (c) a = 4; (d) a = 2.

128

Applications of the 2-D CWT. I: image processing

(a)

(b)

(c)

(d)

Fig. 4.2. Comparison between the different wavelets applied to the L, all at scale a = 2: (a) the test

image; (b) response of the isotropic Mexican hat wavelet; (c) response of the 2-D Morlet wavelet with |k0 | = 6, θ = 0◦ ; (d) response of the end-stopped wavelet ψE1 at θ = 0◦ (from [Bha99]).

fairly strong at other points along the edges. Panel (c) is the result of applying a 2-D Morlet wavelet, oriented horizontally (θ = 0◦ ), which detects only the vertical edges, as it should (it is an efficient directional filter, as we will see below). Notice the horizontal “leaking” of the response: this is an artifact due to the use of a scale that is too small for the signal (so that the wavelet gets too wide in spatial frequency space). Finally, panel (d) is the response of the end-stopped wavelet ψE1 , again oriented horizontally. This wavelet responds only to endpoints of vertical edges, which in this case are all the corners of the contour. Unlike the previous wavelets, however, it shows no response to other points along the edges.

129

4.1 Contour detection, character recognition

(a)

(b)

(c)

(d)

Fig. 4.3. Analysis of the L-shape with the double end-stopped wavelet ψE2 with |k0 | = 6, θ = 90◦ :

(a) response of the wavelet of width σ = 12; (b) local maxima of (a), thresholded at 50%, superimposed (white crosses) on the input image. The wavelet identifies the short horizontal edges; (c) and (d) the same analysis for the same wavelet of width σ = 36. Now long horizontal edges are selected (from [Bha99]).

In a last case, we analyze the L-shape with the double end-stopped wavelet ψE2 , and the result is shown in Figure 4.3, again taken from [Bha99]. This wavelet is sensitive both to the orientation and to the size of the edges, and its behavior may be understood as follows. The Morlet component of the filter selects linear structures which are perpendicular to the orientation of the wavelet, i.e., here horizontal segments. Then the Mexican hat component, which operates parallel to the linear structures detected by the Morlet component, produces strong responses only for those structures that approach

130

Applications of the 2-D CWT. I: image processing

the characteristic length of the Mexican hat (which is determined by σ ). Thus, the ψE2 wavelet detects linear structures of given size and orientation. In our example, the ψE2 wavelet is oriented vertically, using two different values of σ . Figures 4.3(a) and (b) show the response of the wavelet with σ = 12, first the raw response, then the local maxima of (a), thresholded at 50% and superimposed (white crosses) on the input image. The wavelet identifies the short horizontal edges. Similarly, Figures 4.3(c) and (d) show the raw response and the corresponding thresholded local maxima, respectively, for the wavelet with σ = 36. For this value of σ , the strongest response of the wavelet corresponds to the longer horizontal edge in the input image.

4.1.2

Application to character recognition We will now apply the technique developed in the previous section to the problem of character recognition. Our signal will be a set of simplified characters, modeled by the corresponding characteristic function. At this stage of our investigation, we only consider characters composed of segments or union of rectangles. Let us consider for instance a few simple characters (Figure 4.4). We see that r the L has six corners: five convex and one concave, r the A has 12 corners: six convex and six concave, r the E has 12 corners: eight convex and four concave, r the H has 12 corners: eight convex and four concave. The interesting point is that, in this case, the number of concave corners and convex corners completely characterizes these letters. Therefore, an automatic recognition of these characters requires a fast algorithm for extracting this particular information and encoding it. In what follows we will propose a first step for designing such an algorithm, based on the 2-D CWT. In order to follow the construction in detail, we focus on a thick letter A, represented by its characteristic function. Of course, this object behaves exactly as the academic

Fig. 4.4. A set of simplified characters: the letters L, A, E, H.

131

4.1 Contour detection, character recognition

10

5

0

−5

−10 0 0

200

200

400

400 600

Y

600 X

(a)

1

1

1

-1

-1

-1 -1

-1 -1 1 1

1 (c)

(b) Fig. 4.5. Detecting the contour of the letter A with the radial Mexican hat: (a) the CWT at a = 0.075, in 3-D perspective; (b) the same, in level curves; (c) coding of the same by the signs of the respective corners.

image of the preceding section. If one works again with a radial Mexican hat, and goes down to a sufficiently small scale, the CWT reveals the contour of the letter. Moreover, the corners of the figure are highlighted in the WT by sharp peaks, the sign of which is determined by the convexity of the corresponding corner, since the wavelet is real. The result is shown in Figure 4.5. In panel (a), we show the WT of the letter at scale a = 0.075, in 3-D perspective. Here we see clearly the twelve peaks corresponding to each corner, some positive (for the six convex corners), some negative (for the six concave corners). Panel (b) presents the same result in level curves, and panel (c) shows the coding of the corners by a logical flag (± 1 for concavity or convexity).

132

Applications of the 2-D CWT. I: image processing

This exercise leads to an algorithm for automatic character recognition [17]. The basic idea of the method is to treat only the significant parts of the signal, focusing on the information needed for unambiguous recognition. In the case of simple letters, this information is entirely contained in the high-frequency components, namely corners and edges. Take, for example, the letter A of Figure 4.5. It can be entirely characterized by the succession of its 12 corners and the additional information that consists in deciding whether a corner is concave or convex. That is, twelve points and a logical flag (concavity or convexity) for each point. The following simple algorithm achieves this treatment. It consists in locating the local maxima of the CWT and eliminating everything else by thresholding, and it is able to detect an A unambiguously. r Compute the CWT S(b, a f ) in position representation with the Mexican hat wavelet at the finest relevant scale a = a f . The transform exhibits local extrema at the corners and is positive (negative) for a convex (concave) corner. Compute the absolute extrema of the transform a f )}, m(a f ) = min{S(b, b

r

a f )}. M(a f ) = max{S(b, b

To get rid of the other high-frequency components, threshold the transform using a negative value T− > m(a f ) and a positive value T+ < M(a f ), both directly computed from the CWT. All the values between T− and T+ are set to zero (in the terminology of Donoho [146,147], this is a hard thresholding, see Section 4.6). r We are left with an image, denoted by T S(b, a f ), composed of positive and negative peaks at the position of the corners, which we encode as a vector with components +1 or −1 depending on the local sign of the (thresholded) transform. Using this simple technique we are able to deal with simple shapes, especially with characters that are not corrupted with noise. In the case where we have additive noise or if we need to be more accurate, we use the same treatment at a different small scale, $ % a j )}( j=1,...,N ) . that is a ∈ amin , (amax /amin )1/2 . We obtain a sequence of images {T S(b, Adding these images together gives an image from which one encodes again the local maxima and minima, which are now enhanced against the background noise. The scheme of this algorithm is displayed in Figure 4.6 and applied for recognition in Figure 4.7, for the (noiseless) case of the four letters of Figure 4.4. Panel (a) shows the CWT of the signal with a Mexican hat wavelet, at scale a = 1.5. Panel (b) gives the skeleton of that CWT, thresholded at 90% of the maximal values; as expected, only the top of the peaks survive, and they are shown with crosses for the positive peaks (maxima, convex corners) and circles for negative ones (minima, concave corners). Needless to say, this algorithm works only for a real wavelet and with the values of the CWT itself, not its absolute value. Actually, since only the corners are needed, we may as well use a wavelet that sees only the corners, not the edges. Typically, a directional wavelet (when it is misaligned),

133

4.1 Contour detection, character recognition

Signal

Wavelet ❅ ❘ ❅

✠

aj) S(b,

$ % a j ∈ amin , (amax /amin )1/2

❄

Thresholding

❄

aj) T S(b,

❄

j

aj) T S(b,

Fig. 4.6. Strategy for character recognition.

or a real wavelet such as the gradient wavelets ∂x exp(−| x |2 ) or ∂x ∂ y exp(−| x |2 ), or even better, the end-stopped wavelet ψE1 described in the previous section. The latter has indeed been designed specifically for that purpose. This simple technique may be further improved by adding some denoising and inclusion of a second wavelet capable of dealing with letters of arbitrary shape (for instance, a ring-shaped wavelet sensitive to circular shapes). In addition, the automatic recognition device will need some training. An elegant solution would then be to use the simple wavelet treatment as a preprocessing for some sort of “intelligent” device, such as a neural network. However, when noise is present, a different strategy works better, namely, to use a directional wavelet instead of an isotropic one. The reason is that the detection capability of the former is more robust to noise. This feature may be understood as follows: to specify a direction is an additional element of information, that is present in the signal, but not in the noise, in general, so that the SNR ratio improves. In order to show this, we will analyze below another set of simple letters, namely, A, B and C, in the presence of increasingly strong noise (additive Gaussian noise). Now the technique used here is

134

Applications of the 2-D CWT. I: image processing

Fig. 4.7. Application of the algorithm to character recognition: (a) wavelet tranform of the signal in

Figure 4.4; (b) the same after thresholding: maxima (convex corners) are indicated with crosses, minima (concave corners) with circles.

a particular instance of the general problem of feature detection and recognition, so we postpone the details of the algorithm to the next section. We may also notice that directional wavelets, namely directional derivatives of a smoothing function, have been applied to the same problem of character recognition by Hwang and Chang [230], implementing the wavelet maxima technique of Mallat and Zhong [264,265].

4.2

Object detection and recognition in noisy images Suppose we have an image containing a certain number of targets, embedded in a cluttered environment: how can one detect and identify the various targets in an automated way? Stated explicitly [16], the purpose of automatic target detection and recognition (ATR) is the use of computer processing to detect and recognize signatures in sensor data, especially targets embedded in a cluttered environment, with the aim of neutralizing potential threats to military and civilian populations while minimizing the required resources and the risk to human life (this problem has an obvious military connotation, which explains the jargon used!). Such targets can be tanks, planes, other vehicles, missiles, ground troops, etc. Clutter can be grass, trees, topographical features, atmospheric phenomena (i.e., clouds, smoke, etc.). In general, the situation can be modeled using the following equation: s( x ) = n( x) +

L l=1

Tl ( x ).

(4.1)

135

4.2 Object detection and recognition in noisy images

where n( x ) represents an additive noise (clutter plus measurement noise), Tl ( x ) are targets to be detected and recognized, and s( x ) represents the accessible measured signal. Automatic or assisted target detection and identification requires the ability to extract the essential features of an object from (usually) cluttered environments. However, detection and identification lead to different requirements. Typically, to provide detection and identification of difficult targets while maintaining full surveillance coverage, a coarse resolution sensor is required for detection, while a fine resolution sensor is necessary for recognition (identification). This suggests the use of multiscale techniques, which provide the flexibility to utilize only the resolution required at each level and, perhaps, allow optimal processing for each of the required operations. Many methods are used for the ATR problem: classical pattern matching, modelbased schemes, dyadic wavelets, subband coding, even some attempts using neural networks [155,247]. Here we will describe an approach based on the 2-D continuous wavelet transform, which could offer a great improvement over traditional pattern matching methods. The rationale for using the CWT for ATR is the following. Typical features to be extracted from the image of a target are its position, its spatial extent, and its shape, including its orientation and symmetry. Thus the relevant parameters are position, scale and orientation, that is, exactly those considered in the 2-D continuous wavelet transform. The scale dependence allows sensitivity to variations in sensor resolution, as well as determination of target size, or equivalently the target distance, for instance in optical or infrared imagery. Rotation dependence leads to robust behavior in identifying the orientation of the target. So, unlike other methods, the 2-D wavelet transform incorporates several parameters directly relevant to the essential features of an object. Projection of the transform can thus provide a useful set of image representations for fully automated discrimination. In addition, wavelet methods yield a consistent and efficient image reconstruction algorithm. Indeed, the CWT has been used successfully in a number of situations, notably in infrared imagery. We will discuss this application in the next two sections.

4.2.1

Principle of the ATR wavelet algorithm A simple ATR algorithm based on the 2-D CWT has been proposed in [16]. It consists in a two-stage strategy and relies in an essential way on the successive use of the two basic representations described in Section 2.3.3, the position and the scale-angle representations. A similar technique has been used previously in the analysis of acoustic wave trains in water, see Section 5.3.4. The algorithm reads as follows (Figure 4.8). At the first stage, we compute the CWT in the position representation at all relevant a j , θ j ), j = 1, 2, . . . . For the detection, scales a = a j and angles θ = θ j , that is, S(b, we take the image obtained for each fixed pair (a j , θ j ), threshold it, and add all the images together. Thresholding is performed in an adaptive way, becoming more severe

136

Applications of the 2-D CWT. I: image processing

allow false alarm Input Data

✲

Detection of potential targets: position representation

❄

(a, θ) = (a j , θ j ) j = 1, . . . , N

bi , i = 1, . . . , L

Elimination of ambiguous targets bk , k = 1, . . . , K (K < L) ❄

Target recognition: scale-angle representation

b1 → target T1 (a1 , θ1 ) b2 → target T2 (a2 , θ2 ) ...

Fig. 4.8. A diagram illustrating the two-step strategy for ATR.

for smaller a. Note that other nonlinear transformations (e.g., enhancement, morphological operators) may also be applied. The effect of this procedure is to suppress the clutter information, while preserving the target information. As a result, the latter is reinforced and becomes visually enhanced. Next, we compute the centroids b = bi , i = 1, . . . , L in the resulting composite image. These centroids correspond to the positions of potential targets. False alarms are of course possible, but one may control the false-alarm rate by adjusting the thresholds in order to eliminate spurious false detection. Then, at the second stage, one switches to the scale-angle representation and computes the wavelet transform of the composite image at each remaining centroid b = bk , k = 1, . . . , K (K L). If the centroid bk corresponds to a genuine target Tk , the corresponding wavelet transform will exhibit a unique maximum (ak , θk ), which gives the size and the orientation of the target Tk . Moreover, the signature of each target in the scale-angle representation allows the discrimination between different targets. An academic example of application of the algorithm just described, more precisely the first stage of it, is presented in Figure 4.9. We take our favorite L-shape (a), embedded in a Gaussian white noise (b), with a signal-to-noise ratio of 18. The wavelet transform, with a Mexican hat, is taken at six different scales, a = 1, 2, 4, 8, 16, 32, each image is properly thresholded (at 90%/a), and the six images are added together. The resulting composite image (c) shows the reconstructed object, the noise has largely been suppressed.

137

4.2 Object detection and recognition in noisy images

(b)

(a)

(c) Fig. 4.9. Reconstruction of a signal embedded in noise. (a) The signal; (b) the noisy signal; (c)

reconstruction of the signal with six scales.

Of course, this method is rather primitive, although it is fast and robust. An alternative technique for image denoising relies on the use of directional wavelet packets (Section 2.6). We will make a detailed comparison between the two methods in Section 4.6.

4.2.2

Application to infrared radar imagery: position features The real power of this approach is, of course, better appreciated in real life situations, preferably difficult. A prime example is that of infrared imagery (FLIR or Forward Looking Infrared Radar imagery). Automatic target detection and identification for

138

Applications of the 2-D CWT. I: image processing

FLIR imagery requires the ability to extract the essential features of an object from cluttered environments under the condition that the range to the target is unknown. Moreover, the gray-scale of a target in FLIR imagery displays a great variability: a succession of dark and bright areas depending on the temperature radiating from various parts of the target (for example, a target can be hot or cold). Most of the time, the target does not appear in ideal conditions and it is very difficult to estimate the sensor output probability density function (PDF). One therefore uses algorithms that first extract features of interest (such as structural, spatial and frequency features) and then classifies the objects based on those features (a review of these issues may be found in [134]). Multiscale techniques, such as the CWT, are highly desirable, because they can extract and normalize both the unknown scale and orientation of the target. As indicated above, typical features to be extracted from the image of a target in a cluttered environment are the position, the spatial extent, and the shape of the target, including its orientation and symmetry. This justifies the use of the 2-D CWT in the ATR problem, and in particular for infrared imagery. An additional issue is to determine which wavelet will perform best, an isotropic one (Mexican hat) or a directional one (Morlet or Cauchy). Now, in the presence of heavy noise, it is better to apply the ATR algorithm described above not to the CWT amplitude itself, but rather to one of the partial energy densities discussed in Section 2.3.4. We will then speak of CWT features to be extracted from the image. This opens a choice between two solutions. The most obvious one is to take the position energy density of the signal s( x ), given in (2.56), namely, ∞ da 2π = a, θ )|2 , P[s](b) dθ |S(b, (4.2) a3 0 0 and this we shall do in this section. We test the algorithm on FLIR data from the TRIM2 database, namely, a set of images each of which contains four targets (various types of tanks), seen under 21 different aspect angles (we recall that the aspect angle α is the polar α)). Figures 4.10 and 4.11 present two angle in the position representation, b = (|b|, such images, together with the corresponding receiver operator characteristics (ROC) curves. These curves plot the probability of detection versus false alarm rate, for the whole set of images containing a given type of tank at 21 aspect angles, analyzed in turn with the Mexican hat, the Morlet, and the Cauchy wavelet. The first stage of the ATR algorithm is then applied to the image in Figure 4.11(a) and the results are presented in Figure 4.12. The upper row shows the output of the first step of the algorithm, i.e., the CWT position features, evaluated with the three wavelets successively. Clearly, the targets stand out from their background with more definition than the original. The bottom row then gives the output of the second step of the algorithm, that is, the use of thresholding, morphological transformations, and other nonlinear transformations. The particular image presented corresponds to thresholding at gray-scale value 40. Finally, Figure 4.13 gives the Receiver operator characteristics (ROC) curves obtained

4.2 Object detection and recognition in noisy images

m2

prob. of detection

1

0.75

Cauchy Morlet Mexican Hat

0.5

0.25

0 0

10

20

30

40

(a)

50 60 70 false alarms

80

90

100

110

120

(b)

Fig. 4.10. (a) Original image containing four M2 tanks (left); (b) Receiver operator characteristics (ROC) curves for the set of images containing M2 tanks at 21 aspect angles.

m163

1

prob. of detection

139

0.75

Cauchy

0.5

Morlet Mexican Hat 0.25

0

(a)

0

10

20

30

40

50 60 70 false alarms

80

90

100

110

120

(b)

Fig. 4.11. (a) Original image containing four M163 tanks; (b) Receiver operator characteristics (ROC) curves for the set of images containing M163 tanks at 21 aspect angles.

by combining images with all types of tanks, again for the three wavelets, Mexican hat, Morlet, and Cauchy wavelets. The results show that directional wavelets such as the Morlet and the Cauchy wavelets perform better than an isotropic one such as the Mexican hat. Moreover, the Cauchy wavelet performs better than the Morlet wavelet. This can be understood as follows. The selected images contain objects plunged in clutter noise. Detection of an object of interest supposes, for example, the ability to capture the internal structure (for example, wheels, doors, etc . . . ) of the object, the boundaries (edges, corners) between the

Applications of the 2-D CWT. I: image processing

Fig. 4.12. CWT position densities (top row) and detection results (bottom row) of the original image in Figure 4.11 with the Mexican hat (first column), the Morlet (second column), and the Cauchy wavelet (third column), respectively.

1

prob. of detection

140

0.75

Mexican Hat

0.5

Morlet Cauchy

0.25

0 0

10

20

30

40

50 60 70 false alarms

80

90

100

110

120

Fig. 4.13. Receiver operator curves for the detection algorithms on data containing M1, M2, M163, and M60 tanks.

141

4.2 Object detection and recognition in noisy images

object and its surrounding. This justifies the use of directional wavelets in this context, and explains their better performance over nondirectional filters. This technique might be applied with different wavelets, for example, target-adapted wavelets. It might also be compared to or combined with the CWT-based classification algorithm developed in [287].

4.2.3

Scale-angle features and object recognition The detection algorithm described in the previous section was based on CWT position features extracted from the image. An alternative solution is to use the scale-angle energy density (scale-angle spectrum) of the signal s( x ), given in (2.57), namely, a, θ)|2 . M[s](a, θ ) = d 2 b |S(b, (4.3) R2

In this section, we shall present an algorithm based on scale-angle CWT features, then apply it again to automatic character recognition and target recognition in FLIR images. As we will see, this algorithm allows significant reduction of the data needed for an efficient recognition and it is robust against noise. As in the previous case, we will compare the performance of two wavelets, this time two directional wavelets derived from the Mexican hat, the modulated or Gabor Mexican hat (3.22) and the conical Mexican hat (3.38). Detailed comparative tests show that the conical Mexican hat wavelet outperforms both the Gabor Mexican hat and the usual Morlet wavelet, and all of them outperform traditional methods of character recognition such as template matching [289]. As a first application, we analyze another set of simple letters, namely, A, B, and C, in the presence of increasingly strong noise (additive Gaussian noise). The characters are shown in Figure 4.14(a), at various SNR levels, 20, 25, 30, 35, and 40 dB. Then we compute the continuous wavelet transform of these letters, using the conical Mexican hat wavelet (3.38). The result is presented in Figure 4.14(b) in the scale-angle representation, evaluated at the center of each character. In this way, the directional features of the object are enhanced, and indeed are detected despite the noise. Given any letter in Figure 4.14(a), the problem is to recognize it (in an automated way), that is, to determine which of the 26 letters of the alphabet it resembles most – or actually coincides with. This is an instance of object recognition in a noisy environment, more precisely, identification of a noisy object within a preassigned collection of test objects – here the 26 letters of the alphabet. Here again, the standard technique consists of choosing a certain number of characteristic features that suffice to discriminate unambiguously among the test objects. Then one arranges them in a feature vector and measures the distance between the feature vector of the unknown object and those of each test object. The one that yields the smallest distance gives the answer.

Fig. 4.14. (a) Test data for the recognition algorithm: three characters A, B, and C at various SNR levels, 20, 25, 30, 35, and 40 dB. (b) Continuous wavelet transform of the same, using the conical Mexican hat wavelet. The CWT is presented in the scale-angle representation, evaluated at the center of the character.

143

4.2 Object detection and recognition in noisy images

4.2.3.1

The recognition algorithm We will first describe the algorithm in complete generality. Applications to specific problems will be given in the following sections. Consider a training set, consisting of objects divided into q classes, X j , j = j j j 1, 2, . . . , q, the class X j having p j prototypes, X j = (X 1 , . . . , X l , . . . , X p j ), and an unknown object, Y , that we wish to classify as object X j , for some j. The algorithm for classification is described as follows: r For every j = 1, 2, . . . , q, compute the 2-D CWT scale-angle energy density j j M[X l ](a, θ) for each element X l of the class X j . r For every j = 1, 2, . . . , q, compute the mean and the standard deviation of the elements of the class X j , j 1 j M[X l ](a, θ) p j l=1 " # pj & '2 1/2 1 j σ X j (a, θ) = . M[X l ](a, θ) − µ X j (a, θ ) p j l=1

p

µ X j (a, θ) =

(4.4)

(4.5)

r

Compute the scale-angle energy density M[Y ](a, θ ) of the test object. Select the scale-angle feature vectors V[Y ], V[X j ] for the unknown object Y and for each class X j , i.e., M[Y ](m, n) and µ X j (m, n), σ X j (m, n), m = 1, 2, . . . , N1 , n = 1, 2, . . . , N2 , where N1 , N2 are the chosen number of scales and angles, respectively. r Compute the distance between them, r

d j = d(Y, X j ) ≡ d(V[Y ], V[X j ]), j = 1, . . . , q.

(4.6)

r

Then, if dk is the minimum of the set {d j , j = 1, . . . , q}, the object Y is classified as belonging to the class X k . In principle, one can use any distance, such as the Euclidean distance d(Y, X j ) =

N2 & N1 '2 1 M[Y ](m, n) − µ X j (m, n) , N1 N2 m=1 n=1

or the maximum likelihood distance, which is given by " 2 # N1 N2 j (m, n) M[Y ](m, n) − µ 1 X . 2 log σ X j (m, n) + d(Y, X j ) = N1 N2 m=1 n=1 σ X j (m, n) Clearly this algorithm may lead to false recognition, if two or more distances d j are very close to each other.

4.2.3.2

Application to character recognition The algorithm has been tested [289] on the 26 characters of the alphabet (A,B,C, . . . .) at various noise levels (additive Gaussian noise). In this case, one has q = 26, that j is, each X j is one letter of the alphabet, whereas the p j prototypes X l correspond to different copies of the same letter, distorted, rotated, embedded in different noises. For

144

Applications of the 2-D CWT. I: image processing

No. of orientations = 72

No. of scales = 10

100

100 scales = 10

angles = 72

90

90

80 80

Correct classification

Correct classification

70

60

50

scales = 5

40

70

60

angles = 18

50

30 40 20

10

0 20

30

scales = 2

25

30 35 Noise level (in dB)

(a)

40

20 20

angles = 9

25

30 35 Noise level (in dB)

40

(b)

Fig. 4.15. Results of recognition as a function of noise level for various lengths of the feature vector in the scale-angle plane, using the modulated (Gabor) Mexican hat wavelet with k0 = 6 and eccentricity $ = 2; (a) fixed number (72) of orientations, different numbers of scales; (b) fixed number (10) of scales, different numbers of orientations.

instance, the five letters A on the top row of Figure 4.14(a) are five prototypes of an A, with different levels of noise. The results of the analysis are summarized in Figures 4.15 and 4.16. Both give the performance of the recognition algorithm for various lengths of the feature vector as a function of the noise level (SNR), using the modulated (Gabor) Mexican hat and the conical Mexican hat, respectively. In Figure 4.16 the result of the traditional template matching is added for comparison. In this experiment, it turns out that the conical Mexican hat performs better than the modulated Mexican hat, which in turn performs better than the traditional Morlet wavelet. The same behavior is observed in other cases, for instance in the recognition and classification of targets in FLIR imagery [285–289].

4.2.3.3

Application to radar imagery Other angular features of images have been used successfully in recent work by Kaplan and Murenzi on Synthetic Aperture Radar (SAR) images [241]. The problem here is to have a good estimate of the position of a target (in jargon, pose estimation as part of

145

4.3 Image retrieval

110

100 temp cwt15

Correct classification

90

80

70 cwt10

60

cwt8

50

40

22

24

26

28 Noise level (dB)

30

32

34

Fig. 4.16. Results of recognition as a function of noise level for various lengths of the feature vector in the scale-angle plane; the wavelet now is the conical Mexican hat.

ATR). Since the latter is assumed to be at the center of the image, one fixes the position at b = 0 in the CWT and integrates the energy density over scale only, thus defining the angular energy density amax da E[s](θ ) = |S(0, a, θ )|2 . (4.7) a3 amin The pose estimate is then simply the orientation that maximizes the angular energy density, θ = arg maxθ E[s](θ). This technique yields very good results, using the fast circular convolution algorithm described in Section 2.3.1 with an isotropic Mexican hat wavelet (but, surprisingly, not with a Morlet wavelet).

4.3

Image retrieval

4.3.1

The problem of content-based image retrieval Digital images are everywhere, and they often come in huge quantities: remote-sensing agencies, medical imaging facilities, art museums, travel agencies, law-enforcement

146

Applications of the 2-D CWT. I: image processing

agencies, all have to deal with large collections of archived digital images. Depending on the particular situation, these collections may be homogeneous (as in medical imaging or fingerprint archives), or completely heterogeneous. Thus there arises the problem of image retrieval: given an image, one has to identify those images in the collection that resemble it most (a typical example is the identification of a suspect’s fingerprint among those archived in the police records). Beyond a few dozen images, manual browsing is practically impossible, a purely computerized solution is necessary. Traditional methods of object classification, based on external descriptors, are inoperant for images, especially in heterogeneous collections. One has to resort to a description derived automatically from the image itself, and this leads to a fairly new field of research, called content-based image retrieval. Although several methods have been proposed, we will consider here only the wavelet-based scheme designed by Bhattacharjee in his thesis [Bha99], where an overview of other methods may be found. The roots of this approach lie in the theory of vision initiated by Marr [Mar82], namely the process of comparing images must necessarily be based on low-level information extracted from the images, especially in the case of a heterogeneous collection of images, where no assumption is made about the content of an image. Indeed, methods based on segmentation do not make sense in such a situation. Subdividing the image into blocks, and performing block-wise comparisons of images does not work either, because this approach is not invariant to rotation and cannot support arbitrary subimage-queries. Thus image comparison should be based on visually significant structural features detected in the images. Indeed, this seems to be the case when human subjects view a scene. Psychophysical experiments show that even when analyzing a static scene, our eyes do not remain continually focused on a single retinal image, but rather perform the so-called saccadic movements described in Section 3.3.3. As explained there, the target points of consecutive saccades are points of interest which stand out against the general background of the scene. Thus it remains to identify those key points, and we are facing again the problem of identifying relevant features in an image. According to experiments, the crucial ones are low-level features, which can be classified in various ways (see [Bha99] for a complete discussion). For instance, one may distinguish patches of uniform intensity; edges or lines; and corners or line ends. Among these, the last kind seems to encode the maximum amount of image information. For this reason, it is natural to design wavelets that respond precisely to these features, and the end-stopped wavelets described in Section 3.3.3 are Bhattacharjee’s answer to that question. Before going further into their actual implementation, we shall now briefly describe the whole recognition scheme. The first stage is the feature point detection scheme, which forms the core of the image comparison process. However, image comparison based simply on the positions or color attributes of the feature points would not be very robust. Thus, for each feature point, one constructs a description based on the texture of the immediate neighborhood

147

4.3 Image retrieval

of the point. This is obtained by using suitable filters, namely directional derivatives of a Gaussian, of order one, two and three. The responses of these filters are organized in an ordered set and may be represented as a vector. Thus each feature point detected in the image produces one vector of responses, called a token. These tokens carry information about a small region in the image, including directional information. The last step of the process is the actual measure of similarity between images, based on tokens. This requires a rather complex image indexing strategy. The outcome is an efficient technique for content-based image retrieval, which proceeds iteratively. To quote the author [Bha99]: . . . To use the system, the user presents the query in the form of an image. The system then sorts all images in the collection in order of decreasing similarity to the query, and returns the top few images as the answer-set. The size of the answer-set is specified by the user. From the answer-set, the user may mark the images that are relevant to the user’s needs, and provide this information to the image retrieval system as feedback. The system then refines the query automatically, based on the relevance information provided, and subsequently returns another answer-set of images to the user. This process may be iterated till the user is satisfied.

This image recognition scheme turns out to be both efficient and reasonably economical, as attested by several explicit examples given in [Bha99].

4.3.2

Feature point detection using an end-stopped wavelet To conclude this section, we discuss now in more detail the wavelet-based algorithm for feature point detection [Bha99,76,77] (we quote freely from these works). Both endstopped wavelets ψE1 and ψE2 can be used to select meaningful points in images, and this is why we have discussed them both in Section 3.3.3. However, the author chooses to consider the former only, because, as explained above, he is essentially interested in detecting corner-like features, to which ψE1 responds best. In addition, he works at a fixed scale, since an exhaustive search of the scale space is computationally prohibitive. Furthermore, for a heterogeneous collection of images, to which new images may be added as and when they become available, it is impossible to identify a specific set of scales that will be appropriate for analyzing all images in the collection. Thus, the algorithm applies the ψE1 at the same scale to all images in the collection. There are two negative consequences of this compromise. r The feature point detection scheme is not scale invariant. Consider two images I and 1 I2 , where I2 is a subsampled version of I1 by a factor of two along each dimension. If the ψE1 wavelet is applied at the same scale to both images, we are not guaranteed that the set of feature points detected in I2 will have a one-to-one correspondence with the set of feature points detected in I1 . That is, the system is not invariant to scale changes even by a factor of four. However, as the experimental results demonstrate, the final system is quite robust to small variations of scale, presumably because of the redundancy of information between responses at nearby scales.

148

Applications of the 2-D CWT. I: image processing r

Some of the feature points detected in an image may not be well localized, that is, the local maxima of the wavelet response may not coincide with the corresponding image feature. This happens because the scale of analysis is coarser than the most appropriate local scale, so the local maximum of the wavelet response does not fall on the feature in question, due to the interaction with other neighboring features. Such seemingly spurious points would pose serious problems for recognition systems. However, the performance of the proposed image retrieval system is not appreciably impaired by the use of a fixed scale of analysis. This is because the comparison of images is not based on a recognition process, but rather, is based on a comparison of small image-patches surrounding the feature points. This being said, the algorithm reduces essentially to the standard one described in Section 4.1.2. (i) Transform the input image, at a preselected scale, s0 , with the ψE1 wavelet at N different orientations θ = θ0 , θ1 , · · · , θ N −1 . The result is a set of N response images, each showing strong responses near the end-points of linear structures oriented perpendicular to θ. (ii) For each pixel position, retain only the strongest response value among all the orientations. This produces the so-called maxima image. (iii) Detect peaks of significant local maxima in the maxima image. The coordinates of these peaks give the feature points. In practice, the author uses N = 18, which corresponds to an orientation resolution of 10◦ , to cover the entire semicircle evenly. In fact, the angular resolving power of the ψE1 wavelet is 19.2◦ , as measured by the standard benchmarking technique described in Section 3.4.1, but the response is quite weak away from the axis, so that a significant overlap is recommended. Finally, the scale of analysis chosen is a = 8. We conclude by showing some experimental results obtained with this feature point detection scheme. In particular they demonstrate the robustness of the scheme in the face of rotation and cropping. Figure 4.17 shows the feature points detected for two images of very different kinds. In both cases, the detected points are marked by bright crosses superimposed on the input image. Figure 4.17(c) shows the result for a face image, and Figure 4.17(d) shows the points detected in an ornament image. The ornament images depict very complex hand-drawn artwork, which were designed to be trademarks of publishers in the nineteenth century (these images have been scanned from photocopies of old books, as gray level images). Figure 4.18 demonstrates the robustness of the proposed feature detection scheme towards rotation and cropping. The image in Figure 4.18(a) shows the feature points detected for a subimage of the ornament image shown in Figure 4.17(b). Note that most of the feature points marked in Figure 4.18(a) have corresponding feature points in Figure 4.17(d). Figure 4.18(b) shows the feature points detected for a subimage extracted from a rotated version of the image shown in Figure 4.17(b). The image has

149

4.3 Image retrieval

Fig. 4.17. Results of feature point detection: (a) face test image; (b) ornament test image; (c) feature

points detected in (a); (d) feature points detected in (b). The ornament image shown in (b) is a black-on-white pattern that has been scanned as a gray level image. In both (c) and (d), the white crosses mark the positions of the detected feature points (from [Bha99]).

Fig. 4.18. Robustness to cropping and rotation: (a) feature points detected in a subimage of Figure 4.17(b); (b) feature points marking detected in a rotated, cropped version of Figure 4.17(b). Most feature points are detected in both images (from [Bha99]).

150

Applications of the 2-D CWT. I: image processing

been rotated by 17◦ . Again, most of the feature points in the corresponding section of the original image [Figure 4.17(d)] are also detected in the rotated, cropped version.

4.4

Medical imaging Wavelet applications abound in medicine and biology, both in 1-D and in 2-D, often with the discrete WT [Ald96]. For the 2-D case, one may quote image segmentation, mammography, tomography [68,315], and magnetic resonance imaging (MRI) [217]. Here we will mention only one potential application, still under development, namely the technique of spiral reconstruction in MRI. The problem at hand is the so-called spiral acquisition method in MRI, that is, the Fourier transform of the image to be analyzed is sampled along several interleaving spirals [84]. This raises the question of the completeness of the reconstruction of the original image from such data. This is clearly a case of nonuniform sampling, and completion means that the set of sampling points generate a frame. The question was analyzed and solved by Benedetto [67], using powerful mathematical tools, such as Beurling’s theorem. Actually, 2-D wavelets may also be used in this context, the hope being that they may allow to bypass the infamous “gridding” problem, namely the necessity of adjusting the sampling points on a Cartesian grid in order to apply the FFT algorithm.

4.5

Detection of symmetries in patterns

4.5.1

The tools for symmetry detection Wavelets may be used for evaluating the symmetry of a given pattern under discrete rotations and dilations, as was demonstrated in detail in [24], on which this section is based. The method presented here allows one to determine, in a straightforward and economical way, all the (possibly hidden) symmetries of a given pattern. Of course, invariance under separate rotations or dilation is easy and there are various methods for determining it. But the determination of combined dilation–rotation invariances (helicoidal symmetries) of a given pattern is much more delicate and, in fact, we do not know any other method for doing it. The technique uses, in an essential way, the angular selectivity of the directional wavelets. In order to achieve good precision, one needs a directional wavelet with a very good selectivity in the scale-angle variables. The Gaussian conical wavelet (3.37) is an extremely efficient tool in that respect and we will use it systematically. However, it is a fact that most patterns of interest possess only local or approximate symmetries of this type, i.e., without true periodicity or discrete translational invariance. They are thus quasiperiodic sets, such as quasilattices, planar self-similar (Penrose)

151

4.5 Detection of symmetries in patterns

tilings or diffraction patterns of quasicrystals. These objects have indeed a local symmetry only, but the method applies as well, because the local character of the wavelet transform allows us to treat exact and local symmetries precisely on the same footing. For quasiperiodic objects, the useful information is contained, to a first approximation, only in the scale and angle variables, since there is no spatial periodicity as for regular (crystallographic) tilings [Bar94,61,322]. In such a case, one may ignore the dependence of the CWT on the translation degrees of freedom, represented by the One possibility is to use the scale-angle representation, which consists parameter b. of fixing the position parameter b (Section 2.3.3). However, this may lead to ambiguities, because the result, including its symmetries, depends sensitively on the value of b that has been chosen [24] (an example is given in Section 4.5.2.3). The alternative is to average over all values of b and consider the scale-angle spectrum, as defined in (2.57). a, θ)|2 M[s](a, θ) = d 2 b |S(b, R2 2 rθ−1 k)| 2 d 2 k , 2 |ψ(a (4.8) = (2πa) | s(k)| R2

a, θ) is the wavelet transform of the signal s( where S(b, x ) with respect to a directional wavelet. Clearly, M[s] gives the intensity of the spectrum of s, namely, the contents of |s|2 according to the shape of |ψ|2 . Furthermore, if ψ is averaged locally around arθ−1 (k), is supported in a narrow cone, and then (4.8) “probes” the behavior is directional, ψ of the signal in the direction θ, as the beam of a torchlight exploring a target. This intuitively explains all the results that follow. Positions are not considered in the analysis, because only the modulus of s is used. This is why the method may be interesting in a (quasi)crystallographic context, where only amplitudes of the diffraction spectrum are recorded in experiments. In practice, however, the scale-angle spectrum is often not very readable, because some of the (approximate) symmetries may be rather weak. Therefore, exactly as for the wavelet transform itself, one plots instead the skeleton of the scale-angle spectrum, which in this case reduces to the set of local maxima (isolated peaks in the scale-angle spectrum), and then the possible periodicity properties become clearly visible. In addition, the scale-angle measure is well adapted for analyzing a statistical symmetry [321]. This is a weaker concept of symmetry, which corresponds to the invariance under rotation, or dilation, of the frequency of appearance of any given local configuration inside of the pattern. This is clearly the relevant concept when one is dealing with an approximate symmetry. Let us thus assume that a certain substructure of s interferes positively with the wavelet ψbo ,ao ,θo ; that is, the corresponding wavelet coefficient is large. Assume further that this configuration occurs a certain number of times in s, giving to M[s] a local maximum at a point (a0 , θ0 ). Then, if s is statistically symmetric under a certain rotation

152

Applications of the 2-D CWT. I: image processing

Rρ and a certain dilation Dλ , the local maximum of M[s] has a mirror image at (a0 λ, θ0 + ρ) by the covariance of the CWT (Proposition 2.2.3). The important fact is that not only the maxima are relevant, but all the points of the scale-angle spectrum, because, in the definition of statistical symmetry, any local configuration has importance and gives its contribution to M[s]. Of course, it is nice to detect (statistical) symmetries of the image, but one also wants to know whether one has found all of them. An answer to that question is given by the voting algorithm introduced in [Vdg98], extending a technique of Hwang and Mallat [228]. The idea is to make a vote on the more significant symmetries of the image s under consideration. First, one computes the correlation P[s](τ0 , α0 ) between the scale-angle spectrum M[s] and the version obtained under dilation by a factor λ0 and rotation by an angle α0 , that is, M[s](aλ−1 0 , θ − α0 ). Introducing logarithmic coordinates a = t τ0

e , λ0 = e , and defining M [s](t, θ) = M[s](et , θ), one has M[s](aλ−1 0 , θ − α0 ) = M [s](t − τ0 , θ − α0 ). The correlation P[s] is thus given by: tmax 2π −2 dt dθ M [s](t − τ0 , θ − α0 ) M [s](t, θ ), (4.9) P[s](τ0 , α0 ) = s2 tmin

0

where τ0 ranges from 0 to the width of the logarithmic scale interval. In practice, of course, we take a bounded interval [tmin , tmax ] for t and we work with discrete steps and thus the integration is approximated by a summation over a linear grid ⊂ [tmin , tmax ] × [0, 2π ]. Then the algorithm allocates a vote to the point (τ0 , α0 ) if P[s](τ0 , α0 ) exceeds a given constant K > 0 (which specifies the error that is tolerated). Once a vote has been cast for a point (τ0 , α0 ), one identifies all its integer multiples (nτ0 , nα0 ) that lie within , and give all their votes to (τ0 , α0 ) [Vdg98,24].

4.5.2

Detecting symmetries in 2-D patterns We shall now apply the method developed in Section 4.5.1 to the detection of rotation– dilation symmetries of certain classes of 2-D patterns, following essentially [24,26].

4.5.2.1

Geometric patterns We begin with a simplified version and eliminate the scale dependence by integrating over a, thus ending with the angular spectrum α[s](θ) of the object, defined in (2.60). In general, α[s](θ) is 2π-periodic. However, when the analyzed object has rotational symmetry n, i.e., it is invariant under a rotation of angle 2π /n, then α[s] is in fact 2π/n-periodic. To give a simple example, consider three geometrical figures, a square, a regular hexagon and a rectangle [24]. The square and the hexagon have symmetry n = 4 and n = 6, respectively, and thus their angular spectrum show four and six equal peaks, respectively (Figure 4.19). The width of these peaks is simply the aperture of the support

153

0

4.5 Detection of symmetries in patterns

90

180

(a)

270

360

0

90

180

(b)

270

360

0

90

180

270

360

(c)

Fig. 4.19. Angular measure of regular figures obtained with a Cauchy wavelet (A R P = 20◦ ): (a) a square; (b) a regular hexagon; and (c) a rectangle, with side ratio 2 : 1 (from [24]).

cone (i.e., the ARP) of the wavelet. The case of the rectangle is more interesting. It has symmetry n = 2 × 2 (two mirror symmetries, or rotations by π around both O x or O y), and that is reflected on the graph of its angular spectrum: there are two large peaks corresponding to the longer edges and two smaller peaks corresponding to the shorter ones, and the ratio 2 : 1 between the two equals that of the lengths of the corresponding edges. Indeed, the wavelet catches the direction of the edges, not that of the corners, so that indeed the maxima of α[s] are again at θ = 0◦ , 90◦ , 180◦ , 270◦ , just as for the square, but now the amplitudes are different. This also explains why the peak at 90◦ in the case of the hexagon, panel (b), is slightly higher: the vertical sides in the original figure are sharp, the oblique ones are ragged, for numerical reasons, so that the former give a larger response to the wavelet. This explains why one needs a highly directional wavelet in this case. It is remarkable that the scale-angle spectrum technique works in the presence of severe noise. Let us take again a square pattern and compute its CWT with a directional wavelet, first without noise (Figure 4.20), then with moderate additive Gaussian noise (Figure 4.21, top panels), finally with severe additive Gaussian noise (Figure 4.21, bottom panels). In each figure, we show successively: (c) the angular spectrum (2.60), which reveals the fourfold symmetry of the pattern; and (d) the scale spectrum (2.61), which measures the size of the object (from [289]). Next we proceed to patterns with a genuine combined rotation–dilation symmetry. In this case we need the full scale-angle spectrum M[s](a, θ), which will again be 2π/n-periodic in θ if the pattern has rotational symmetry n. In addition, if the object is invariant under dilation by a factor ao , then M[s] is (log ao )-periodic in log a. Thus in the case of an inflation invariance, M[s](a, θ) is a doubly periodic function in log a and θ (note that, like wavelet transforms themselves, a scale-angle spectrum is usually plotted as a function of log a and θ). The first object we analyze is a “twisted snowflake,” that is, a mathematical snowflake [43,44] with the following modified construction rule: upon each downscaling by a

154

Applications of the 2-D CWT. I: image processing

(b)

(a) 0 10

scale

20 30 40 50 60 0

100

1

0.8

0.8

0.6

0.6

0.4 0.2 0

300

(d)

1

energy

energy

(c)

200 angle

0.4 0.2

0

100

200 angle

300

400

0

0

20

40 scale

60

80

Fig. 4.20. Detection of the symmetries of a square pattern: (a) the signal; (b) the scale-angle spectrum; (c) the angular spectrum; and (d) the scale spectrum (from [289]).

factor of 3, the figure is turned by 36◦ . The scale-angle spectrum of this object, given in Figure 4.22(b), shows precisely the combined symmetry. The set of four maxima at a given scale ao is reproduced, at scale ao /3, but translated in θ by 36◦ .

4.5.2.2

Quasiperiodic point sets An interesting class of point sets is that of the quasilattices based on algebraic numbers (see [Bar94,61] for a systematic analysis). All of them possess a rotational symmetry of order n, where n may be crystallographically allowed (n = 1, 2, 3, 4, 6) or not (e.g. n = 5, 8, 10, 12). Moreover, each pattern is invariant under dilation by a characteristic factor, which is an integer equal to 2 cos(2π /n) in the first case, and an irrational number βn in the second case (thus called quasicrystallographic). But in fact there is more. In many cases, there is in addition a combined rotation–dilation symmetry, typically a rotation by π/n together with a dilation by a factor δn related to βn . As an example, we consider the octagonal pattern shown in Figure 4.23 [24]. It has a global √ symmetry n = 8 and is invariant, by construction, under dilation by a factor 1 + 2. But the scale-angle spectrum M[s](a, θ ) (calculated with a Gaussian conical wavelet) reveals two combined rotation–dilation symmetries, namely a rotation of π/8 √ together with a dilation by a factor δ1 = 2 cos(π/8), or δ2 = 2 cos(π/8), respectively. The remarkable fact is that these two additional symmetries were discovered on the graph of the scale-angle spectrum, not on the tiling itself! (Actually both symmetries

4.5 Detection of symmetries in patterns

(a)

(b) 0 10

scale

20 30 40 50 60 0

100

(c) 0.8 energy

energy

300

1

0.95

0.9

0.85

0.8

200 angle

(d)

1

0.6 0.4 0.2

0

100

200 angle

300

0

400

0

20

(a)

40 scale

60

80

(b) 0 10

scale

20 30 40 50 60 0

100

(c)

200 angle

300

(d)

1

1

0.98

0.8 energy

0.96 energy

155

0.94

0.6 0.4

0.92 0.2

0.9 0.88

0

100

200 angle

300

400

0

0

20

40 scale

60

80

Fig. 4.21. The same analysis as in Figure 4.20, in the presence of additive Gaussian noise: (top) moderate noise [signal-to-noise ratio (SNR) = 26 dB]; (bottom) severe noise (SNR = 22 dB).

Applications of the 2-D CWT. I: image processing

−1.5

log(3)

−2

−2.5

log(a)

−3

−3.5

−4 0

20

40

60

80

100

120

140

160

180

θ

36

(a)

(b)

Fig. 4.22. Analysis of a “twisted snowflake”: (a) the pattern; (b) the scale-angle spectrum M[s](a, θ ), computed with a Cauchy wavelet (m = 4, γ = 10◦ ). Corresponding local maxima are shifted by 36◦ and a scaling ratio of 3 (from [24]).

2 1.8 1.6 1.4 1.2 ln(a)

156

1 0.8 0.6 0.4 0.2 0 0

(a)

20

40

60

80

θ

100

120

140

160

180

(b)

Fig. 4.23. Analysis of an octagonal pattern: (a) the pattern; (b) the local maxima of its scale-angle spectrum M[s](a, θ ); this pattern has a rotation symmetry by π/4, and two distinct √ mixed symmetries, consisting of a rotation by π/8 combined with a dilation by δ1 = 2 cos(π/8), δ2 = 2 cos(π/8), respectively. Homologous maxima are linked by a line segment, continuous for δ1 and dashed for δ2 (from [24]).

157

4.5 Detection of symmetries in patterns

(a)

(b)

Fig. 4.24. Two sets of octagons on the octagonal tiling obtained by successive applications of a √ rotation by π/8 combined with a dilation by (a), δ1 = 2 cos(π/8); (b), δ2 = 2 cos(π/8) (from [24]).

were later derived by a geometrical argument.) A nice way of visualizing the symmetries is to draw successive octagons, representing the orbits of successive points under a rotation by π/4 (Figure 4.24). Note that, for a better visualization of these orbits, we have brought back all successive summits into the first sector 0 θ π/8. This operation reveals an additional difference between the two symmetries. Indeed the pattern on the right is invariant under the combined operation δ2 -dilation + rotation by π/8, and this operation generates a semigroup (every point has a successor, not necessarily a predecessor, i.e., the inverse operation is not a symmetry). This semigroup has apparently infinitely many different orbits (on the portion of the tiling visible on the figure, we have detected 10 different orbits). However, the other combined operation, δ1 -dilation + rotation, is not an exact symmetry, it is only approximate. For instance, some orbits stop after a few iterations, or have gaps. Now comes the question, did we detect all symmetries of the octagonal pattern? The answer is in fact yes, as shown by the result of the voting algorithm described in Section 4.5.1, presented in Figure 4.25. The graph indeed shows the two pure operations, rotation √ by π/4 and dilation (ρ0 ) by a factor 1 + 2, and the points ρ1 and ρ2 corresponding to the combined rotation–dilation operations with dilation ratio δ1 and δ2 , respectively. Note that the pure dilation ρ0 is equivalent under the π/4-periodicity to the product ρ1 · ρ2 .

4.5.2.3

Other examples of aperiodic patterns There are many more examples of patterns that exhibit this kind of combined symmetries. A whole class is that of tilings of the plane, some of which are commonly known

Applications of the 2-D CWT. I: image processing

1.2

1

ρ

0

0.8

ln(a)

158

ρ2

0.6

0.4

ρ1

0.2

0

0

10

20

30

40

θ

50

60

70

80

Fig. 4.25. The set of renormalization parameters ρ = (δ, θ) obtained by the voting algorithm on the

scale-angle spectrum of the octagonal tiling given in Figure 4.23. The point on the horizontal axis corresponds to the π/4-periodicity. The points ρ1 and ρ2 correspond to the combined rotation– dilation operations with dilation ratio δ1 and δ2 , respectively, whereas ρ0 is a pure dilation, equivalent to the product ρ1 · ρ2 under the π/4-periodicity. The other, unmarked, points are translations of the previous ones under both periodicities, in a and θ (from [24]).

under the name of Penrose tilings (these are dual to the preceding type, in the sense that they are obtained by drawing the Voronoi cells of the point set). We show a typical example in Figure 4.26. From the scale-angle spectrum, obtained with a Gaussian conical wavelet (3.37), with parameters m = n = 4, σ = 16, we conclude that this pattern √ has 1 a rotation symmetry by π/5, a dilation symmetry by τ = 2 cos(π/5) = 2 (1 + 5), the golden mean, and a mixed symmetry, consisting of a rotation by π/10 combined with a dilation by λ = 1.36. Incidentally, these examples show why it is safer to integrate over all scales in order to isolate the angular behavior, rather than to fix a certain scale a = ao and consider M[s](ao , θ). If ao coincides with one of the characteristic scales, a1 , a2 , . . . , the result is correct, but if ao falls in between, no maximum will be seen, and the symmetry is not detected. The effect is shown in Figure 4.27. Another interesting class of examples may be found in various pattern-forming phenomena in fluids [194]. Typically nonlinear waves at the surface of a fluid generate a regular pattern, via an unstability and a bifurcation. Most of these patterns have a

159

4.5 Detection of symmetries in patterns

(b)

(a)

Fig. 4.26. Symmetry detection with the CWT: (a) a Penrose tiling; (b) the corresponding scale-angle spectrum M[s](a, θ ), obtained with a Gaussian conical wavelet (γ = 10◦ , m = 4, σ = 16). Homologous maxima are linked by a line segment (from [24]). 550

270

500

260 250

450

240 400 230 350 220 300 210 250

200

200

150 0

190

20

40

60

80

100

(a)

120

140

160

180

180 0

20

40

60

80

100

120

140

160

180

(b)

Fig. 4.27. The scale-angle measure of the octagonal pattern from Figure 4.23, for fixed values ao of the scale: (a) for ln ao = 1.20, the periodicity is obvious; (b) for ln ao = 1.13, between two lines of maxima, the symmetry is not seen (from [24]).

rotational invariance of some order, and some of them are quasicrystalline. As a result, they lend themselves quite naturally to a wavelet analysis [86]. The best known case is the instability demonstrated by Faraday in 1831. The resulting pattern was known to have a rotational symmetry of order n = 12. Our standard analysis indeed yields this symmetry, together with √ the corresponding invariance under dilation by the corresponding factor β12 = 2 + 3 3.73. In addition, we find, as before, a combined √ symmetry of a rotation by 2π /24 = 15◦ together with a dilation by δ = 1.89 β12 ,

160

Applications of the 2-D CWT. I: image processing

which was unexpected [26]. It can actually be proved that such a combined rotation– dilation invariance, with half the rotation angle, is always present in all these “algebraic quasicrystals” [Bar94]. This technique permits one to determine, in a straightforward way, the (possibly hidden) symmetries of a given pattern. This applies to a genuine lattice, but also to a quasilattice, for which the symmetry is only local, for instance, the diffraction spectrum of a quasicrystal. Let us remind the reader that quasicrystals are those remarkable alloys discovered in 1984 [338], whose X-ray diffraction patterns show local n-fold point symmetry for n = 5, 8, 10, or 12. The latter are crystallographically forbidden for being incompatible with translational invariance (the only rotational symmetries compatible with lattice periodicity are of order n = 1, 2, 3, 4, or 6). These diffraction patterns display bright Bragg peaks of unequal intensity and they are self-similar with irrational scaling factors. More precisely, the involved irrationals are the following algebraic numbers: √ τ = 12 (1 + (pentagonal or decagonal quasilattices) √ 5) = 2 cos(π/5) β8 = 1 + √2 = 1 + 2 cos(π/4) (octagonal case) β12 = 2 + 3 = 2 + 2 cos(π/12) (dodecagonal case), that is, precisely the dilation factors discussed above. Similarly, some of the wave functions for transport electrons in quasicrystals are critical: they are neither localized (as would be the case in a random amorphous structure), nor spread out (as for perfect periodic crystals). Moreover, they display self-similarity too. For example, in the √ fivefold case, self-similarity ratios are typically powers of the golden mean τ = 12 (1 + 5). Thus we expect that the present wavelet-based method will yield interesting physical applications in the field of crystallography, in three possible directions. The first one concerns the diffraction patterns, where, at a given resolution level, it is necessary to classify and label the Bragg peaks according to their position and intensity. Secondly, two-dimensional wavelets based on the number τ seem particularly appropriate for the scanning analysis of patterns obtained through tunneling or atomic force microscopy of quasicrystalline surfaces. The third application concerns the determination of electronic wave functions in quasicrystals (explicit construction by using a discrete wavelet basis, adapted to the given symmetry type). Actually, these applications may be made easier if one uses a set of wavelets directly adapted to the symmetry. We will describe some of these in Section 11.4.2.

4.5.2.4

Point sets generated from noncrystallographic Coxeter groups A completely different kind of example is given by quasilattices derived from infinite dimensional extensions of noncrystallographic Coxeter groups. The idea is to generate an aperiodic point set by applying successive reflections and translations to the root system. The example we have analyzed is based on an affine extension of the Coxeter group H2 [27,303]. This group is isomorphic to the dihedral group of order 10 and it has a noncrystallographic root system (that contains explicitly the golden mean τ ). Then,

4.5 Detection of symmetries in patterns

0.9624 log(a/amin)

161

σ3

0.7218

σ4

ρ1

0.4812

σ1

ρ4

0.2406

ρ3

ρ5

ρ2

0 0

(a)

σ2

9

18 27 θ (in degrees)

ρ0 36

(b)

Fig. 4.28. (a) A H2aff -induced quasicrystal centered on (0, 0); the picture shows the part of the

quasicrystal located within a radius of 2; (b) result of the voting algorithm with a Gaussian conical wavelet (γ = 3.5◦ , m = 4, σ = 10) (from [27]).

applying a number of reflections and translations in a two-dimensional subspace of the root space, one generates planar aperiodic point set, called an H2aff -induced quasicrystal and shown in Figure 4.28(a). Next we apply the voting algorithm of Section 4.5.1. The result is shown in Figure 4.28(b). We observe six basic symmetries, labeled by ρ j with j = 0, . . . , 5. Among these, the pure operations may be considered as major symmetries, because the intensity of their maxima is dominant, namely the rotation by π/5, denoted ρ0 , and the dilation by τ , denoted ρ1 . All others are weaker symmetries. One of them, denoted ρ2 , is again a pure rotation symmetry, with angle π/10, but since the intensity observed is much weaker than the intensity of ρ0 , it is not possible to view ρ2 as the generator of the latter. In addition, we have combined symmetries, ρ3 , ρ4 , and ρ5 . The remaining symmetries, labeled by σ j , with j = 1, . . . , 4, are compositions of the basic ones. Indeed, σ3 , σ2 , σ4 , σ1 may be obtained by combining the pure dilation ρ1 with ρ4 , ρ3 , ρ5 , and ρ0 , respectively. It remains to provide a geometrical interpretation of all these symmetries. As building blocks of the quasiperiodic tiling, one may take a decagon contained in Figure 4.28(a), just like the octagons of Figure 4.24, and its successive translations. Each translated decagon is reproduced several times during the successive rotations of an angle of kπ /5 with k = 1, . . . , 9. Thus any symmetry found in this construction will be important for the whole tiling. Furthermore, any symmetry which is present already in the nontranslated decagon will appear with an even stronger intensity, because it appears more frequently and thus leads to a more dominant statistical symmetry. Then one finds that each of the symmetries σ2 and ρ3 corresponds to the ratio of two specific segments in the basic decagon, with the angle between them. In addition, the weaker symmetries σ3 , σ4 , ρ4 , and ρ5 can be traced to similar geometric relations in the figure [27].

162

Applications of the 2-D CWT. I: image processing

4.6

Image denoising A very successful application of wavelets in image processing is to denoising. Most of the authors so far have used the DWT for that purpose, and indeed orthogonal or biorthogonal wavelet bases. In some cases, wavelet frames have also been exploited and also the lifting scheme of Section 2.5.2.4 [Jan00,Jan01]. A notable improvement was obtained by E. Cand`es with a new tool, better adapted to images, called curvelets [Can98,96,97,346] (we will describe these in Section 11.1.4). It turns out, however, that the techniques described in the present book may be useful too. We have already seen an example in Section 4.2, namely, the first stage of the ATR algorithm. Although it is rather primitive, this technique has the advantage of being very simple. A much more elaborate approach is to take directional dyadic wavelet frames, introduced in Section 2.6.3. Without going into technical details, it is instructive to compare the two methods, and this we shall do here. The key notion in every denoising method is thresholding. The idea is that only the relevant features of an image yield significant wavelet coefficients, whereas the noise gives many small coefficients, spread more or less everywhere in the parameter space. Thus it suffices to put to zero all the coefficients that lie below a fixed threshold #. The problem, however, is how to choose the latter. A sophisticated and highly successful technique was introduced by Donoho and Johnstone [146,147]. In addition, one has to choose between two versions: r hard thresholding: here the small coefficients, i.e. |c | < #, are replaced by 0 and jk the rest remains untouched. As a consequence, artificial discontinuities are created. r soft thresholding or wavelet shrinkage: in order to remove these discontinuities, all the remaining coefficients are shifted by ±#, so as to make them continuous. This thresholding technique, initially developed in 1-D, extends to 2-D in an obvious way. The ATR method has been described in Section 4.2, and one observes that it uses a hard thresholding, with a threshold level # fixed arbitrarily. As for the dyadic wavelet packets, we start with a dyadic tight frame generated by a wavelet that satisfies the conditions (compare Proposition 2.4.1, in particular, (2.114)) L−1

ϕ − 2π /L)|2 = |ψ(r )|2 , |(r,

(4.10)

=0

) is the Fourier transform of an isotropic where L is the number of orientations, and ψ(r dyadic wavelet, i.e., +∞ j=−∞

j r )|2 = 1 |ψ(2

(4.11)

163

4.7 Nonlinear extensions of the CWT

and J

j r )|2 = |φ(2 J r )|2 , |ψ(2

(4.12)

j=−∞

where φ is the associated 2-D scaling function [265]. Using a directional frame allows us to put more redundancy in the technique and to benefit from the fact that directional wavelets will emphasize oriented features, such as edges, etc. Even though (2.140) [or (2.141)] is a continuous formula, all computations may be carried out in a discrete setting, either by means of the sampling theorem, or by using the approximate QMFs introduced in Section 2.6.4. It suffices then to compute the wavelet coefficients and to threshold them appropriately. The choice of the threshold is a crucial matter and usually requires an estimation of the standard deviation of the contaminating noise. Finally, reconstructing using the thresholded coefficients yields an estimated, denoised image. In order to illustrate the efficiency of the method, we present in Figure 4.29 a comparison between the ATR technique and the present one. We choose again our familiar L-shape, embedded in increasingly severe Gaussian noise, with standard deviation σ = 14 255, 12 255, and 255, and corresponding PSNR 12.07, 6.04, and 0.01 dB, respectively. In the middle column, we show the image denoised with the ATR method. The CWT is computed with an isotropic Mexican hat over five dyadic scales, 2, 4, 8, 16, and 32. Then, a hard thresholding is applied at each scale layer with a relative (with respect to the maximum) threshold decreasing with scale: 50, 25, 12.5, 6.25, and 3.125%, respectively. Finally, all these modified scale layers are summed, giving an image denoised, in the sense that the object has been considerably enhanced over the noise. In the right-hand column of the figure, we show the result obtained with a directional dyadic wavelet packet, using five scales and eight orientations. The denoised images have PSNRs equal to 25.97, 21.63, and 18.16 dB, respectively. The result is obviously better than the one obtained with the ATR method. As a second example, we show in Figure 4.30 the denoising of the lena image with the directional dyadic wavelet packet method, using five scales and 16 orientations. Panel (a) shows the original image, panel (b) the noisy version, obtained by adding to the original picture a white noise of σ = 25.5 (PSNR: 20.02 dB). Panel (c) shows the denoised image, which now has a PSNR of 31.07 dB.

4.7

Nonlinear extensions of the CWT

4.7.1

Local contrast The intensity of light around us varies considerably, in fact by several orders of magnitude. Our visual system is well adapted to this situation. Indeed it analyzes the spatial organization of the luminous field by relying on the contrast of objects and figures contained in the images. Intuitively, contrast is defined as the ratio between a variation

164

Applications of the 2-D CWT. I: image processing

Fig. 4.29. Comparison between the two denoising methods. The left column (a) shows the signal, with noise increasing from top to bottom, the middle one (b) is the result of the ATR algorithm, the right one (c) that obtained with a directional dyadic wavelet packet.

of luminance and a reference level of luminance. It is mathematically expressed using Weber’s law: CW =

L . L

(4.13)

165

4.7 Nonlinear extensions of the CWT

(a)

(b)

(c)

Fig. 4.30. Denoising of lena with a directional dyadic wavelet packet.

This definition is often used for small patches with a luminance offset L on a uniform background of luminance L. In the case of sinusoids or other periodic patterns of symmetrical deviations ranging from L min to L max , which are also very popular in vision experiments, one generally uses the Michelson contrast [Mic27], namely, CM =

L max − L min . L max + L min

(4.14)

While these two definitions are good predictors of perceived contrast for the abovementioned classes of simple stimuli, they fail when the stimuli become more complex and cover a wider frequency range, for example, Gabor patches [307]. It is also evident

166

Applications of the 2-D CWT. I: image processing

that neither of these simple global definitions is appropriate for measuring contrast in natural images, because the brightest and darkest points would determine the contrast of the entire image, whereas actual human contrast perception varies with the local average luminance. In order to address these issues and provide a quantitative definition of contrast, Peli [308] proposed a local band-limited contrast: C Pj (x, y) =

ψ j ∗ I (x, y) , φ j ∗ I (x, y)

(4.15)

where I (x, y) is the input image, ψ j is a band-pass filter at level j of a filter bank, and φ j is the corresponding low-pass filter. The normalization by the low-pass signal takes into account the local luminance variations. Modifications of this contrast definition have been used in a number of vision models [118,257] and are in good agreement with psychophysical experiments on Gabor patches [307]. The particular form of (4.15) suggests use of the wavelet transform for describing the variations of luminance. Now the WT is a space-scale analysis, and the spatial extension of the wavelets is characterized explicitly by their scale factor. Thus it is possible to define at each scale a different normalization, similar to a local average. So, following Duval–Destin [Duv91,12], one is led to the notion of local contrast, defined by combining the wavelet transform with an adaptive normalization. The latter will be obtained by projecting the signal, at a given scale, on a local weight function, chosen with the same localization properties as the wavelets. This local mean value will be called luminous level. This is the background against which luminance variations are measured, and the WT may be interpreted as a representation of these luminance variations within an image. The resulting contrast analysis is nonlinear, but it presents several advantages. It is particularly well adapted to the processing of positive signals. It also yields a multiplicative reconstruction process, which preserves positivity. Let us give some details and an example of application. Let h ∈ L 1 (R2 ) ∩ L 2 (R2 ) be a non-negative, rotation invariant, weight function, normalized to h L 1 = 1. Given an image, represented by a non-negative function f , the luminous level with respect to the weight function h is defined as = h (b,a) . Ma [ f ](b) h (b,a) x ) = a −2 h a −1 ( x − b) (4.16) | f , ( instead of the usual Note that we use throughout the L 1 -normalization, that is, h (b,a) h b,a (see Section 2.2). This is more natural in this context, since all the functions h (b,a) have the same L 1 -norm. Then we define the local contrast as the ratio of the CWT to the corresponding luminous level (the wavelet ψ is assumed to be also rotation invariant): = Ca [ f ](b)

Fa (b) Ma [ f ](b)

=

ψ(b,a) |f h (b,a) |f

=

ψb,a |f h b,a |f

,

(4.17)

167

4.7 Nonlinear extensions of the CWT 1 ˘ b, ≡ F( a) = ψ(b,a) where Fa (b) | f is the CWT of f with the L -normalization (but the local contrast is independent of the normalization). In order to make sense, this definition requires that the support of ψ be contained in the support of h. The local contrast is nonlinear, but its behavior is controlled by an integral condition. Large absolute values of contrast imply the existence of a region where the luminance signal is very small. A typical example, very natural in the study of vision, is to take for h a Gaussian and for ψ a Mexican hat. But one can do better and take for ψ the difference wavelet associated to h, as given in (3.13). Then the local contrast becomes

= Ca [ f ](b)

h (b,aα) |f h (b,a) |f

− 1,

(4.18)

and the existence condition is simply that the support of h be star-shaped. This formula in turn leads to a multiplicative reconstruction scheme. Indeed, estimates of the luminous level at smaller and smaller scale factors a may be considered as smoothened versions of the image with progressively contracted weight functions h. Then, as for the WT, the approximation of a function at a given scale may be written in terms of the approximation at a larger scale and the complementary signal: Maα [ f ] = Ma [ f ] · (Ca [ f ] + 1), Maα2 [ f ] = Maα [ f ] · (Caα [ f ] + 1)

(4.19)

= Maα [ f ] · (Ca [ f ] + 1) · (Caα [ f ] + 1), and, by recurrence: Maαn [ f ] = Maα [ f ] · (Ca [ f ] + 1) . . . (Caαn−1 [ f ] + 1).

(4.20)

Maαn [ f ] is the nth resolution approximation of f ; it is the image as seen through the smoothing function h contracted by a factor aα n (a < 1). One notices the obvious analogy with the usual multiresolution analysis (Section 1.5). The formalism may be generalized further to the so-called infinitesimal contrast analysis developed in [159]. An interesting application of the notion of local contrast is the design of an algorithm for the matching of stereoscopic images [313,314]. The technique consists in using a localized correlation function for comparing the two images, by means of the 2-D CWT. The locality of the latter provides a good estimate of the disparity between images, at each scale. However, in order to prevent the occurrence of artifacts due to the high sensitivity of the localized correlation function to the local mean value of the signals, one must normalize the latter in an adaptive way, and this leads precisely to local contrast described above. Peli and Duval–Destin’s definition of local contrast as defined above measures contrast only as incremental or decremental changes from the local background, which is analogous to the symmetric (in-phase) responses of vision mechanisms. However, a

168

Applications of the 2-D CWT. I: image processing

(b)

(a)

(c) Fig. 4.31. Peli’s local contrast (c) from equation (4.15) computed for the lena image (a) using an isotropic band-pass filter (b).

complete description of contrast for complex stimuli has to include the antisymmetric (quadrature) responses as well [347]. The problem is illustrated in Figure 4.31, which shows the contrast C P computed with an isotropic band-pass filter for the lena image. It can be observed that C P does not predict correctly the perceived contrast, as it varies between positive and negative values of similar amplitude at the border between bright and dark regions and exhibits zero-crossings right where the perceived contrast is actually highest. This behavior can be understood when C P is computed for sinusoids with a constant C M . The contrast computed using only a symmetric filter actually oscillates between ±C M with the same frequency as the underlying sinusoid, which complicates

169

4.7 Nonlinear extensions of the CWT

establishing a correspondence between such a local contrast measure and data from psychophysical experiments. These examples underline the need for taking into account both the in-phase and the quadrature component in order to be able to relate a generalized definition of contrast to the Michelson contrast of a sinusoid grating. Analytic filters represent an elegant way to achieve this. The magnitude of the analytic filter response, which is the sum of the energy responses of in-phase and quadrature components, exhibits the desired behavior, i.e., it gives a constant response to sinusoid gratings. Unfortunately, extending the Hilbert transform to 2-D is not a straightforward task. However, as already stressed in Chapter 3, directional wavelets offer a pleasant alternative, since the Fourier transform of the wavelet is included in a convex cone with apex at the origin and of aperture strictly smaller than π . This means that at least three such wavelets are required to cover all possible orientations uniformly, but otherwise there is no restriction on the number of filters. There are many applications where isotropy is required. In these cases, it is important to combine the analytic responses defined above into an isotropic contrast measure. Working in polar coordinates (r, ϕ) in the Fourier domain, we choose, as in Section ϕ) satisfying the above requirements and the conditions 2.6.3, a directional wavelet (r, (4.10)–(4.12) above. Note that the function φ in (4.12) need not be a scaling function associated to ψ, but it should at least have the same localization properties in order to provide for a meaningful normalization of the luminance level. Now it is possible to construct an isotropic contrast measure from the energy sum of directional filter responses [374]: /

2 | j ∗ I (x, y)|2 I C j (x, y) = , (4.21) φ j ∗ I (x, y) where j denotes the wavelet dilated by 2− j and rotated by 2π /L. If the directional wavelet belongs to L 1 (R2 ) ∩ L 2 (R2 ), the convolution in the numerator of (4.21) is again a square integrable function, and (4.10) shows that its L 2 -norm is exactly what would have been obtained using the isotropic wavelet ψ. C Ij is thus an orientation- and phase-independent quantity, but being defined by means of analytic filters, it behaves as prescribed with respect to sinusoidal gratings (i.e., C Ij (x, y) ≡ C M in this case). Examples for this isotropic contrast are shown in Figure 4.32. It can be seen that the contrast features obtained with C Ij correspond very well to perceived contrast. The combination of the directional analytic filter responses produces a naturally meaningful phase-independent measure of isotropic contrast. This technique may be applied for improving the contrast in any kind of image. An example of application to a photograph was given in [12]. Here we show one with a medical image (Figure 4.33). The image f is decomposed over N contrast levels, as in (4.20), using the couple Gaussian–DOG. For each level j, one defines the contrast chart as the modulus of the local contrast,

170

Applications of the 2-D CWT. I: image processing

(a)

(c)

(b)

Fig. 4.32. Isotropic contrast of the lena image as described by equation (4.21) at three different levels, 0, 1, and 2.

= |C2 j [ f ](b)|, M j (b) j = 1, . . . , N .

(4.22) 6N

as a measure of the = j=1 M j (b) Then one interprets the product of the N charts, S(b) After thresholding, correlation between the successive scales of the image at the point b. one obtains a binary image or mask. The latter is used in medical imagery, for instance, as a preprocessing to more sophisticated algorithms. It is taken as a priori knowledge and helps to reduce the amount of computation.

4.7.2

Watermarking of images Digital image watermarking consists in embedding a digital signature in an image. This operation is usually performed by slightly modifying the visual information in such a way that the perturbation is invisible to human eyes, but can still be recovered by using an appropriate algorithm. This embedding can be performed directly in the spatial domain, but also in the frequency domain (using DCT† coefficients) or, as we shall see now, in the wavelet domain. Finally, one often asks that the watermark should be robust, that is, it should survive common image alterations: geometrical image transformations, addition of noise, lossy compression or even print-scan procedure. We refer to [214] for further details. The generic picture of an image watermarking application is depicted in Figure 4.34. The inputs of the system consist in the original image, the watermark and an optional public or secret key. The watermark, or digital signature, can be of various nature: number, image or text. The key is used to encrypt the watermark and prevents it being read by unauthorized parties. In the sequel, we will mainly focus on the embedding part of the system and, more precisely, we will see how the wavelet transform can be used in conjunction with a vision model for robust and imperceptible image watermarking. †

DCT = discrete cosine transform (see [Mal99]).

171

4.7 Nonlinear extensions of the CWT

Fig. 4.33. Contrast analysis of a medical image: (a) the original image; (b) the CWT with a Mexican

j = −1; (d) the resulting binary image. Many more hat ( j = −1); (c) the contrast chart M j (b), details are seen on the two bottom images than on the ordinary CWT.

Watermark

Host Data

Digital Watermarking

Secret / Public key Fig. 4.34. Typical image watermarking system.

Watermarked Data

logCT

Applications of the 2-D CWT. I: image processing

ε

CT0

172

CM0

logCM

Fig. 4.35. Contrast masking model describing the relationship between the masker contrast and the target contrast at detection threshold.

In order to model the visibility of the watermark in the original image, there are mainly two effects that need to be taken into account, namely, contrast sensitivity and masking. Contrast sensitivity, as we have seen previously, describes the response of the human visual system to the contrast of a stimulus. Masking, on the other hand, describes the phenomenon in which a signal, the masker, is capable of “hiding” a second signal, the target. In other words, the target visibility depends on the presence of a masker. It is possible to combine contrast sensitivity and masking in a model that describes the relation between the masker contrast and the target contrast at detection threshold. Figure 4.35 shows such a model where we have represented on the horizontal axis the logarithm of the masker contrast CM , and on the vertical axis we have the logarithm of the target contrast CT . The curve is divided into a threshold range, where the target detection threshold is independent of the masker contrast, and a masking range, where it grows as a power of the masker contrast. The mathematical description of this model is given by: C T0 if CM < CM0 , ε CT (CM ) = (4.23) CT0 CM /CM0 otherwise. The model contains three parameters, ε, CT0 and CM0 , which specify the size of the threshold and the masking range as well as the slope of the transducer function. They have to be determined by means of subjective experiments.

173

4.7 Nonlinear extensions of the CWT

This model is applied in [359] to a watermarking scheme based on spatial spreadspectrum modulation, as proposed by Kutter [Kut99]. Each bit to be embedded in the image is represented by a two-dimensional pseudo-random pattern. The statistics of the pattern are bimodal with equal probabilities for −1 and 1. The random patterns of all bits are superimposed as follows: w(x, y) = α(x, y) pi (x, y), (4.24) i

where pi (x, y) are the pseudo-random modulation functions for bit i, α(x, y) is the watermark weighting function, and w(x, y) is the resulting watermark which is added to the image. In this watermarking scheme, the pseudo-random patterns pi are sparse, which means that the superposition of all patterns does not necessarily modify all pixels in the image. To quantify the sparseness, we introduce the density D of the watermark, which is given by the modified number of pixels divided by the total number of pixels in the image. The watermark weighting function α(x, y) is computed using the introduced masking model and the local isotropic contrast measure presented in Section 4.7. For computing the local contrast according to (4.21), we use directional wavelet frames as described in Section 2.6, based on the scaling functions of Table 2.1. The minimum number of orientations required by the analytic filter constraint, i.e., an angular support smaller than π , is three. The human visual system emphasizes horizontal and vertical directions, so four orientations should be used as a practical minimum. To give additional weight to diagonal structures, eight orientations are preferred. We only use the highest frequency band of the pyramidal decomposition, because masking is strongest when masker and target have similar frequencies. Furthermore, higher levels tend to smear the local contrast and are thus not suitable for this kind of application. The watermark weighting function α is now computed as follows: α(x, y) = CT (C0I )(x, y) · φ0 ∗ I (x, y),

(4.25)

where C0I is the local isotropic contrast of the masker image at level 0, CT is the corresponding target contrast threshold as given by our masking model, and φ0 is a low-pass filter. The local amplitude of the watermark at the threshold of visibility is thus determined by the multiplication of the isotropic contrast values with the corresponding low-pass filtered image. Finally, the parameters of our vision model (CT0 , CM0 and ε) have been determined by performing subjective tests [359]. Figure 4.36 shows weighting masks for the lena image at watermark densities of 0.4 and 1, respectively. For illustrative purposes, the figures to the right visualize the segmentation into threshold and masking ranges. The dark areas correspond to regions where only contrast sensitivity is exploited, and the bright areas show regions where the masking effect is dominant.

174

Applications of the 2-D CWT. I: image processing

(a) Weights for D = 0.4

(c) Weights for D =1

(b) Segmentation for D = 0.4

(d) Segmentation for D = 1

Fig. 4.36. Watermark weighting function of the lena image for two different density values (left column). The segmented images (right column) illustrate the threshold and masking ranges of the contrast masking model, represented by dark and bright areas, respectively.

In comparison with other watermarking schemes, this weighting mask based on the simple masking model presented above facilitates the insertion of a watermark with higher energy while preserving the visual quality of the image, leading to a watermark that is more robust. It has also been applied successfully to watermarking the blue channel of color images [251].

5

Applications of the 2-D CWT. II: physical applications

In the previous chapter, we have discussed a number of applications of the 2-D CWT that belong essentially to the realm of image processing. Besides these, however, there are plenty of applications to genuine physical problems, in such diverse fields as astrophysics, geophysics, fluid dynamics or fractal analysis. Here the CWT appears as a new analysis tool, that often proves more efficient than traditional methods, which in fact rarely go beyond standard Fourier analysis. We will review some of these applications in the present chapter, without pretention of exhaustivity, of course. Our treatment will often be sketchy, but we have tried to provide always full references to the original papers.

5.1

Astronomy and astrophysics

5.1.1

Wavelets and astronomical images Astronomical imaging has distinct characteristics. First, the Universe has a marked hierarchical structure, almost fractal. Nearby stars, galaxies, quasars, galaxy clusters and superclusters have very different sizes and live at very different distances, which means that the scale variable is essential and a multiscale analysis is in order, instead of the usual Fourier methods. This suggests wavelet analysis. Now, the main problem is that of detecting particular features, relations, groupings, etc., in images, which leads us to prefer the continuous WT over the discrete WT. Finally, there is in general no privileged direction, nor particular oriented features, in astrophysical images. All this leads us to use the CWT with an isotropic 2-D wavelet. In addition, astrophysical images are very noisy. In particular the bright sky and our own galaxy (the Milky Way) represent noise, which must be removed, with a technique similar to that used in 1-D for the subtraction of unwanted lines or noise in spectra [210]. Here statistical techniques play an essential rˆole. All these considerations characterize the type of wavelet applications that have been developed in astronomy and astrophysics. The first attempt to apply the CWT to astrophysical images is due to the group of A. Bijaoui in Nice, in 1990 (see [80] for a review). In their pioneering paper [343], they

175

176

Applications of the 2-D CWT. II: physical applications

used the CWT for the analysis of galaxy clusters, with the 2-D Mexican hat. Similar techniques were exploited by a large number of authors, especially in the last few years. In this section, we will review some of this work, and also present two novel applications, one to solar physics, the other one to the detection of gamma-ray sources in the Universe.

5.1.2

Structure of the Universe, cosmic microwave background (CMB) radiation In several papers [161,162,193,343,344], the authors exploit galaxy counts to identify galaxy groupings, from compact groups (0.5 degrees or tens of kpc† in extent) to clusters (down to 1 degree, from hundreds of kpc to some Mpc), to large-scale structures or superclusters (5 degrees or tens of Mpc or more), including the determination of a possible hierarchy between them. The same technique allows the detection of voids, that is, large regions (up to 60 Mpc) with very few galaxies, and also to a neat definition of each of these notions. The results of such work leads to the analysis of the largescale structure of the Universe, thus to cosmological considerations. For instance, the distribution of groups of various size and of voids points to a possible fractal structure of the Universe. On the other hand, the multiscale approach yields much valuable insight into the inner structure of individual clusters [193]. Here, as in all papers analyzing galaxy maps, the basic data is a bidimensional distribution of Dirac delta functions, possibly weighted according to some statistical criterion. The same type of data will be used in the next two sections. A byproduct of such hierarchical analyses is the multiscale vision model developed by Bijaoui and his group [78,79] in order to detect and characterize structures of different sizes (for numerical reasons, also linked to the necessity of denoising the images, they later switched to a discrete WT, based on spline wavelets). For instance, they propose in [254] a morphological indicator allowing a comparison between various cosmological models (for instance, cold versus hot dark matter). In the same vein, a group from Santander, Spain, has undertaken a systematic analysis, by wavelet methods, of the COBE data on the cosmic microwave background (CMB) radiation. As a first step, they study the local (i.e., in small sky patches) temperature anisotropies in the CMB, including denoising the images [332,333]. In these papers, the authors use both the CWT and the DWT (the latter especially for denoising). As for the former choice, they first consider a 2-D CWT without a rotation parameter, but with two independent scalings in the x and y directions, then the usual isotropic Haar and Mexican hat wavelets. Next [100,361], they use isotropic wavelets to detect and determine the flux of point sources superimposed on the CMB, in conditions simulating the Planck Surveyor mission. As they point out, the advantage of the wavelet method is that no assumption has to be made regarding the statistical properties of the point †

1 pc = 1 parsec = 3.26 light years.

177

5.1 Astronomy and astrophysics

source population or the underlying emission of the CMB. Since the CMB observations are performed with antennas that are best modeled by a Gaussian beam, it turns out that the isotropic Mexican hat wavelet is in fact optimal for detecting point sources. In a further work [362], a detailed comparison is made of the wavelet method with the standard maximum-entropy method. The conclusion is that the two methods are in fact complementary and can be combined to improve the accuracy of the detection. More recently, the Santander group has turned to a global analysis of the CMB, trying to detect potential non-Gaussian CMB temperature fluctuations. This is an important observation for cosmology, for any non-Gaussianity would be evidence for a departure from standard inflationary theories. Since the data used in these experiments is the full sky COBE-DMR data, it is clear that the sphericity of the data has to be taken into account. As a consequence, one has to resort to spherical wavelets. A first attempt was made by Barreiro et al. [62], using discrete spherical Haar wavelets, constructed with the lifting scheme of Schr¨oder and Sweldens [336], described in Section 2.5.2.4. Then, following the same logic that recommends the use of the CWT with an isotropic Mexican hat, the Santander group introduced the spherical Mexican hat (see below), establishing the superior capability of the latter over the Haar wavelets [267]. The net result of these investigations is that the CMB temperature fluctuations are consistent with a Gaussian distribution, thus vindicating the standard theories [101]. Finally [363], the same group has used the same spherical Mexican hat wavelet for extending their previous work [361] on simulated Planck maps, thus achieving a large catalog of potential point sources. Coming back to the CWT on the 2-sphere S 2 , a mathematically precise transform was constructed in [29] and will be discussed at length in Chapter 9, Section 9.2. The idea is simply to take the plane R2 as the tangent plane at the North Pole of S 2 and lift functions on R2 to functions on S 2 by inverse stereographic projection. Introducing polar coordinates both on the plane and on the sphere, the correspondence reads: θ S 2 & (θ, ϕ) ⇐⇒ (r, ϕ) ≡ (2 tan , ϕ). 2 For square integrable functions, this leads to a unitary map between the respective Hilbert spaces, I −1 : L 2 (R2 , d x) → L 2 (S 2 , sin θ dθ dϕ), namely, (I −1 f )(θ, ϕ) =

θ 2 f (2 tan , ϕ). 1 + cos θ 2

(5.1)

In the case of an isotropic wavelet ψ(r ), with r = | x |, the correspondence is simply (I −1 ψ)(θ ) =

2 θ ψ(2 tan ). 1 + cos θ 2

(5.2)

Then, choosing for ψ the isotropic Mexican hat wavelet ψH , one gets the spherical Mexican hat wavelet:

178

Applications of the 2-D CWT. II: physical applications

ψH,sph (θ ) = 4

1 − 2 tan2 θ2 exp(−2 tan2 θ2 ). 1 + cos θ

(5.3)

It should be noted that, in this parametrization, the scaling x → a −1 x in the tangent plane becomes on the sphere: tan θ2 → a −1 tan θ2 . We refer to Chapter 9 for the full analysis. Another research area where the CWT has been used is the detailed analysis of individual galaxies, notably in the group of P. Frick [173,174]. Of particular importance is the cross-correlation between images obtained at different wavelengths. To that effect, the authors of [174] consider the (normalized) wavelet cross-correlation function or wavelet correlation coefficient, obtained by polarization from the wavelet spectrum (2.61) W[s](a): C[s1 , s2 ](a) =

a) S2 (b, a) d 2 b S1 (b,

(W[s1 ](a) W[s2 ](a))1/2

,

(5.4)

and originally introduced by Hudgins et al. [227]. Using this tool, the authors study a particular spiral galaxy, called NGC 6946, comparing the images of total radio emission, red light and mid-infrared dust emission on all scales. Note that in their treatment they use both the Mexican hat wavelet and, for a better separation of scales, their own isotropic wavelet, called Pet hat and defined in (3.11). In a later work from Frick’s group [175], spherical wavelets (in a somewhat primitive form) are used for isolating coherent structures in the distribution of the Faraday rotation measure of extragalactic radio sources, that is, a weighted integral of the longitudinal magnetic field along the line of sight. In addition, since these sources are given as irregularly distributed points in the sky, they adapt to the 2-D spherical situation the technique of gapped wavelets introduced previously in 1-D [172]. A final application in astrophysics, still under development, is to gravitational lensing, namely, the detection of Einstein arcs in cosmological pictures [82]. Whenever the light from a distant bright object (a quasar) is seen through a galaxy, the latter behaves as a gravitational lens, so that the point source appears as a ring, or a portion of a ring (“arclet”), if the alignment is not exact. By measuring the radius of that ring, one may infer the distance of the source. This may be done in two steps. The center of the arc is obtained with an annular-shaped wavelet, such as the Bessel filter (3.9), Frick’s wavelet (3.11), or the annular Halo wavelet (3.12), used at a rather large scale (e.g. a = 2). This determination is quite robust to noise, in particular, to spurious bright points, that mimic nearby stars. The arc itself is obtained with an isotropic Mexican hat, at a smaller scale (e.g. a = 0.5). By superposing the two transforms and applying a severe thresholding (up to 95%) for eliminating the noise, one obtains an image with three bright spots: two points of the arc, around the endpoints, and the center of the corresponding circle. From this, one can reconstruct the arc unambiguously, and thus one obtains a tool for measuring in a simple way the distance of quasars, for instance.

179

5.1 Astronomy and astrophysics

Before turning to more specific 2-D applications, we hasten to add that 1-D wavelet analysis has been used in various current problems is astrophysics. A case in point is the analysis of the solar neutrino capture rate data from the Homestake experiment [215], a crucial ingredient in the resolution of the celebrated solar neutrino problem.

5.1.3

Application of the CWT in solar astronomy Since 1996, the Extreme-ultraviolet Imaging Telescope (EIT) on board the Solar and Heliospheric Observatory (SoHO) satellite observes the Sun in four wavelengths: 171, ˚ These correspond respectively to particular emission lines of 195, 284 and 304 A. Iron (IX-X, XII, XV) and Helium (II), and thus to temperatures typical of those of the Sun corona in the first three wavelengths, and of the transition region in the fourth one [130]. The Sun corona is physically very complex and contains a huge amount of different events appearing at different locations and scales. Solar astronomers are interested in the physics which can be deduced from them in order to improve our knowledge of the global Sun. One way to achieve this is to make time statistics on features of special solar objects. In addition, because of the large number of EIT pictures (currently greater than 100 000), astronomers aim at an automatic analysis. However, many conceptual problems arise due to the difference between the human description of things and the true (logical) computer vision. These can be summarized into two main questions. r How to define a Sun corona object in simple terms, that is, in sufficiently simple concepts which can be managed by a computer program? r How to determine the relevant characteristics of such an object and how to translate them as simply as possible? After a short description of the common Sun corona objects, we will show that the continuous wavelet transform (CWT) offers tools to answer these fundamental questions (we basically follow [37]). Notice also that Bijaoui and his group have applied their vision model to the analysis of EIT images of the solar corona [316].

5.1.3.1

Special coronal objects The physical objects of the Sun corona result in general from convective motions in the solar mantle and/or of magnetic interactions with hot material. Here is a list of the principal objects, ordered by size, from the smallest to the largest (for more information, see [130,278]). ˚ images, the magnetic network constitutes a Magnetic network: In the red 304 A textured solar background resulting of the advection of small magnetic flux by the convective motion in the solar mantle. Brightenings: Brightenings are visible in all EIT images and are related to magnetic topology changes to a lower energy state. Their typical scale in an image is close

180

Applications of the 2-D CWT. II: physical applications

of the pixel size, but they brighten and fade away on a time scale ranging from several minutes to hours. Flares: A sudden and energetic local brightening in an active region (see below). Bright points: Bright points (BP) are small regions with enhanced emission. They are located above pairs of magnetic features of opposite polarity in the photosphere. We can see them in the quiet corona and in coronal holes. They present a lifetime ranging between two hours and two days. Magnetic loops (or Loops): These objects result from the filling of magnetic field lines with plasma. Because the temperature of this material varies along the loop, ˚ because they are cooler the footpoints of the loop are more precise in the 171 A, ˚ The magnetic loops may than the loop summit, which is better seen in the 195 A. be part of the same active region, connecting two regions of opposite flux, or even join different ARs. Active regions: They show up as a region of large increase in the ultraviolet flux on the image. Their typical size is about 10% of the solar radius. Physically, these active regions (AR) contain hot material in smaller and larger loops around and inside a region of enhanced magnetic flux. Because active regions are deeply related to the well-known Sun spots, they appear in two bands of latitude according to the evolution of the main solar cycle of 11 years: They live at high latitudes at the solar minimum (beginning of the cycle) and move towards the equator at the solar maximum. Coronal holes: Coronal holes (CH) are large regions where the magnetic field lines are open and are advected by the solar wind into interplanetary space. Because the energy is advected away, the CH are colder than the closed magnetic field regions and they appear effectively like dark holes in the EIT images. Their morphology evolves with time and they become very small during the Sun maximum. A visual summary of the solar objects defined above is presented in Figure 5.1. There are also features which are not related to the Sun physics, but either to defects of the SoHO satellite due to its aging, or to its interaction with some external events. The main ones are the cosmic ray hits, which, by the interactions of cosmic rays with the EIT CCD camera, plague the images with many bright pixels or bright straight lines, depending on the cosmic orientation relatively to the CCD surface. We should mention finally that all these images have a noise component, namely, a readout noise coming from the CCD camera, the solar noise and the photon-shot noise (Poisson noise). The global noise is well approximated by Gaussian statistics, because of the high counting effect (central limit theorem).

5.1.3.2

Distribution of small features To start with, let us define the specific wavelet tools needed for the present application, which aims at selecting some of the solar corona phenomena. We restrict ourselves to

181

5.1 Astronomy and astrophysics

˚ wavelength image, the Fig. 5.1. The main Sun corona objects. The top left quadrant is the 304 A ˚ rest corresponds to 171 A.

isotropic wavelets, since directions are irrelevant in the present context. Given an image a) with respect to a Mexican hat and the corresponding s( x ), we consider its CWT S(b, a)|2 . Then, as in Section 2.3.5, we define ridges R j , energy density E[s](b, a) = |S(b, and the corresponding amplitude A j (2.63) and slope S j (2.64). These two parameters are sufficient for the detection and discrimination of small features contained in the image s. A precious tool to that effect is the histogram of the amplitude as a function of the slope, or the slope–amplitude histogram. Let a0 the smallest relevant scale. Choose a sequence {b j , 0 j K − 1} of max a0 ), belonging to ridges {R j , 0 j K − 1}. Then, given the set of ima of E[s](b, all corresponding couples (S j , A j )0 j K −1 , the histogram is built by the following simple algorithm: r determine the desired size of the histogram H, say M × N , and initialize H as the zero M × N matrix; r compute S min and Smax , the minimum and the maximum of all the slopes (S j )0 j K −1 , respectively; r compute A min and Amax , the minimum and the maximum of all the amplitudes (A j )0 j K −1 , respectively; r form the discretized slope S m = Smin + m (Smax − Smin ) for 0 m (M − 1), M−1 n = Amin + n (Amax − Amin ) for 0 n (N − and the discretized amplitude A N −1 1);

182

Applications of the 2-D CWT. II: physical applications r

then, for k = 0, . . . , K − 1: – take the slope Sk and compute the index m such that Sm is the nearest discretized slope from Sk , that is, the index m such that −0.5 < (M − 1)

Sm − Sk 0.5; Smax − Smin

(5.5)

– do the same with the amplitude and determine the index n such that −0.5 < (N − 1)

n − Ak A 0.5; Amax − Amin

(5.6)

– increment the (m, n) entry of H by one, Hmn := Hmn + 1. The histogram H reflects the 2-D distribution of the slope–amplitude couples. Identifying distinct areas inside H is equivalent to detecting different classes of small objects contained in the image s. Before computing any histogram, a practical remark must be made about the difference between the continuous theoretical world and the discretized view of the programming. Indeed, actual computation requires an adequate sampling of the image s and of the wavelet ψ. Therefore, the scale a cannot effectively go to zero in (2.63) and (2.64). Indeed, the wavelet ψ must be sampled sufficiently on the grid determining the is essentially contained image. Thus, there will be a minimal scale a0 for which ψ b a0 in the frequency domain [−π, π ) × [−π, π ) (assuming the sampling period T is equal to 1).

5.1.3.3

Analysis of academic objects To test our method, we begin by analyzing two types of objects that will model small features in EIT images. Take first the smallest possible object, a singularity of height c localized on a point u, represented by a Dirac distribution s( x ) = c δ (2) ( x − u). One readily computes the CWT of s and the corresponding energy density a) = E[s](b,

+ c2 ++ +2 . u − b)) ψ(a −1 ( 4 a

(5.7)

then E[s] is maximum in b = u It is easy to see that, if ψ has a maximum in x = 0, for all scales. The equation of the associated ridge is simply ( u , a) for all a ∈ R+ . The amplitude of this ridge is given by ln Au = −4 ln a0 + ln c (thus it tends to ∞ as a0 → 0) and the corresponding slope has the value −4. The second object is a simple Gaussian localized in w, of width σ and height D, s( x ) = D exp(− 12 | x − w| 2 /σ 2 ). A detailed calculation shows that E[s] has also a vertical ridge localized in b = w with a maximum in a = σ . The amplitude of this ridge is proportional to D and the slope is now positive.

183

5.1 Astronomy and astrophysics

Fig. 5.2. Analysis of singularities and Gaussians. (a) The original academic image; (b) the

slope–amplitude histogram (the logarithm of the amplitude is plotted to reduce the range); (c) the selection of points in the singularity population (triangles), or in the Gaussian population (circles).

These two examples show that the amplitude yields a criterion for selecting small objects according to their intensity. Then the slope decides between a singularity or a larger object modeled by a Gaussian. The procedure is illustrated in Figure 5.2. We analyze an academic image s of size 256 × 256, shown in panel (a), and consisting of a collection of randomly placed singularities and Gaussians of small size, and compute the corresponding slope–amplitude

184

Applications of the 2-D CWT. II: physical applications

histogram, knowing that, in this case, the minimal scale a0 of the Mexican hat is close to 0.9. In Figure 5.2(b), we clearly see two distinct populations in the slope–amplitude histogram. The population on the left-hand side (left dashed circle) corresponds to singularities of s with a slope centered around −4. The right area (right circle) corresponds to the Gaussians. A rough selection of points according to the sign of the slope is made in Figure 5.2(c). Negative slopes are represented by triangles, and positive ones by circles; singularities and Gaussians are effectively selected separately.

5.1.3.4

Application to EIT images We can now apply the preceding technique to the selection of cosmic ray hits and of bright points in the EIT images. The former are well described by singularities, because cosmics burn only a few pixels on the CCD camera of the satellite, and the latter can be modeled by Gaussians of small size. ˚ The analyzed EIT image, shown in Figure 5.3(a), is the top-left quadrant of a 284 A wavelength image of the Sun. The slope–amplitude histogram, computed for a0 = 1, is presented in Figure 5.3(b). We notice that the distinction between populations is not as neat as in the academic example. The reason is that the white noise present in the picture recording has a main effect of spreading the well-defined areas of Figure 5.2(b). Next, we impose on the histogram of Figure 5.3(b) an additional selection criterion. For cosmics, we choose the maxima b j of E[s](., a0 ) such that ln A j > 2 and S j < 0 and, for bright points, those with S j > 0. The amplitude thresholding prevents us from taking singularities that are too faint coming from quantization and Gaussian noise. The result is shown in Figure 5.3(c). The cosmics are detected everywhere in the image (because they are not related to solar physics), while the bright points appear mainly on the solar disk (on-disk). In Figure 5.4, we make a zoom on a particular on-disk area of the Sun. The selection effect is now clearer than in the global image.

5.1.3.5

Conclusion and open questions We have presented in the previous section a simple method based on the CWT to discriminate two kinds of simple events in the Sun corona pictures, the cosmic ray hits and the bright points. However, there is ample room for improving the range and efficiency of the method. First, noise has been suppressed in the slope–amplitude histogram by a hard thresholding. However, a precise statistical study remains to be made on these selections according to the SNR of the analyzed images and their CWTs. Then, one may try to characterize more complex solar phenomena, such as the active regions, the magnetic loops or the textured magnetic network. A possibility is to exploit the full information carried by the vertical ridges of the CWT. Indeed, the method described above uses only the first relevant scale of the latter, that is, only their

185

5.1 Astronomy and astrophysics

˚ wavelength EIT image; (b) Fig. 5.3. Analysis of an EIT image. (a) The top-left quadrant of a 284 A the slope–amplitude histogram; (c) the selected cosmics (triangles) and bright points (circles).

beginning. Information about possible maxima of E[s] along these ridges is interesting too, for instance for determining the typical scale which defines each type object, as in the 1-D analysis of impact experiments [358]. Several hierarchical criteria based on the CWT may also help us to detect the inclusion of small events into larger ones, such as the brightenings within the active regions. The shape of strong response areas in the CWT at different scales could be useful in this context, as found in [344], for instance.

186

Applications of the 2-D CWT. II: physical applications

Fig. 5.4. A closer look on a small on-disk region of the Sun: (a) Bright points selection; (b) cosmics

selection.

Finally, directional wavelets may be useful, too, since many solar events present an anisotropic behavior. The magnetic loops, for instance, are locally equivalent to straight lines characterized by a particular width. At a scale proportional to this width, the anisotropic CWT coefficients should vary for different angles θ relatively to the main direction of this line. A magnetic loop signature could perhaps be found inside this variation (this is a variation on the problem of detection of oriented contours).

5.1.4

Detection of gamma-ray sources in the Universe Another topic where the CWT has been applied successfully is the analysis of the X-ray structure of various objects, such as clusters of galaxies, following a suggestion by Grebenev et al. [201]. This leads to a different class of problems. Indeed, such sources are frequently at the limit of detection, so that statistical considerations become crucial. In particular, we are here often in the photon-counting regime, the photon per pixel statistics is significantly different from Gaussian and most sources are extended. The analysis of such images by wavelet methods was further developed by Damiani et al. [115]. The work reported here uses a similar approach, but goes significantly deeper in the analysis. In particular, considerable care is devoted to the presence of Poisson noise in the photon flux. We again follow [37]. When it comes to the detection and analysis of gamma-ray sources in the Universe, the way data analysis is carried out depends on the energy range explored by the telescope. This is not only due to the nature of emitting objects, and to the difficulty of designing appropriate detectors, but also to the gradually lower photon counting rate

187

5.1 Astronomy and astrophysics

as the energy increases. For example, in average 100 ultraviolet photons from the Sun are expected to be detected in one second by each pixel of the SoHO CCD camera (Section 5.1.3), whereas about 1 gamma photon is recorded by the whole gammaray space telescope EGRET during the same period! This correspondingly decreases accuracy and significance of any statistical decision, like event detection. Equally important for the data analyst, the nature of photon-counting processes induces an intrinsic “noise,” called Poisson noise, requiring more statistical care than the usual Gaussian noise. The problem we address in this section is the detection of sources in the raw data of the above-mentioned telescope EGRET (20 MeV – 30 GeV photons). Sources are pointlike objects like pulsars or active galactic nuclei and appear in the data as a few detected photons coming from the same direction in the sky. The whole issue is to give a meaning to the coincidence of finding these photons together, hence to conclude (or not) that they were produced by chance from the diffuse background (interaction of cosmic rays with interstellar clouds). In addition to the detection significance, the position, magnitude and spectral characteristics of a source are other desirable quantities determined from the data. This may seem a very humble problem to solve, but, as outlined above, the scope of questions one can answer at 1 GeV is considerably restricted compared to the wealth of the analysis in the previous section.

Sample data and the classical solutions Every dot on Figure 5.5 is a detected photon, of energy 100 MeV or above, during the viewing period 21.0 of EGRET. A “position” on this counting map refers to a direction in the sky, and the map is modeled as approximately flat. Such counting maps are very broadly modeled as counting (Poisson) processes from two contributions, the background flux and the flux from the sources. That is, we do not directly observe light Declination

5.1.4.1

200 30

180 25 160

140 20 120 15

100 10

80 5 60

40 0 20 −5

0 0

70

50

60

100

50

150

40

200

30

250

20

10 300 Right Ascension

Fig. 5.5. Detected photons above 100 MeV during EGRET viewing period 21.0.

188

Applications of the 2-D CWT. II: physical applications

intensities, but rather photons that are randomly created from the corresponding physical objects (point-like objects or extended objects like interstellar gases). Moreover, the detector is far from being perfect: direction and energy of an incoming photon are recorded with an error that translates into the convolution of the above-mentioned fluxes by a bell-shaped function, the PSF (point-spread function), which is here more heavy-tailed than a Gaussian [268]. As most recognition tasks in data analysis, gamma-ray source detection is often carried out by a maximum likelihood (ML) method. That is, a parametrized source model is fitted to data through maximization of the probability that the data arose as a realization of the suggested model. This involves a heavy nonlinear optimization procedure and an initial guess from the user to set the parameter values (height, width and position, say) to their optimal values. But one eventually ends up with a very faithful account of the physical properties of each source. There are statistical reasons to think that it is hard to beat the quality of ML estimation, like the minimum variance property (see [Ead71]). The reference for ML source detection in EGRET data is [268]. Our concern, however, is to develop a simpler method of source detection based on the continuous wavelet transform. Roughly speaking, the idea is to group the events in a chosen interval of energies into a single 2-D counting map, as in Figure 5.5, and to take its CWT with an isotropic wavelet, typically a Mexican hat. The source candidates are the maxima of the wavelet transform. Then, based on some statistical criterion, a detection significance will be given to each maximum. The higher the significance, the more likely the candidate to be a true source. In order to give the status of source candidate to the maxima of the wavelet transform, we must make sure that relevant information (the sources) is properly decorrelated from noise (the background). For this purpose, the Mexican hat wavelet is a good choice, because: (i) its isotropic bell shape responds mostly to bell-shaped sources; (ii) its good localization in space allows to discriminate events according to their position, more efficiently than the Laplacian of the heavy-tailed PSF; and (iii) its good localization in the frequency plane permits us to discriminate events according to their relative scale. The statistical performance of a wavelet method is presumably poorer than what can be achieved by ML. However, the latter has two drawbacks, namely, high computational complexity of the implementation and supervision of the optimization process. These issues now simply disappear, since wavelets allow a real-time automatic processing of the same job. To stress again the difference between the two tools, we can think of wavelet methods as providing an initial guess for more thorough ML identification of the source parameters, or as an on-board data processing module to warn the astronomers in case of a sudden gamma-ray burst. Before giving the details of our procedure, let us mention that, as in the analysis of the distant Universe [78,79], related problems like the identification of extended objects, that is, multiscale structures, have also been attacked with the discrete WT

189

5.1 Astronomy and astrophysics

[Sta98,345]. As usual, this allows a faster implementation, but the choice of filter is severely restricted, in particular, the Mexican hat is not admissible. Thus a balance must be made between quality of the analysis and speed of implementation. We can also note at this point that the “almost flat” approximation used above, namely that a direction in the sky (a point on the sphere) can locally be represented by two planar coordinates, is not necessary. If a more global data analysis is required, we can always switch to the genuine spherical wavelet transform, mentioned in Section 5.1.2, and discussed at length in Section 9.2. This makes no conceptual difference in what follows, only the algorithms will be more CPU-time consuming.

5.1.4.2

Decision criteria and results It remains to describe the procedure itself. In what follows, we will concentrate on the wavelet aspects and skip most of the statistical arguments, which may be found in detail in [37]. The problem may be subdivided in a series of questions, each of which requires to choose a decision criterion, following more specific questions. (i) What is the detection criterion? Or in other words, how big should the values of the wavelet transform be to conclude that a peak is indeed a source? Our criterion is based on a physical model of the background interstellar gamma-ray emission, related to the distribution of hydrogen in the galaxy. The idea is to measure the discrepancy between this model and the data in the wavelet domain. Peaks will be considered sources if they significantly overshoot the model. (ii) How to estimate the total photon flux from a source? Intuitively, the bigger the value of the wavelet transform at the position of a source, the larger the flux of the source; and, since the wavelet transform is linear, this relation should be linear too. That is correct, modulo complications due to the presence of an unknown background of magnitude comparable to that of the source. We can however use our coarse a priori background model to remove part of this bias. Our estimator of the flux of a source detected to be at position x is not the value y obs (x) of the wavelet transform of the observed data, at scale a and at position x, but rather =

y obs (x) − W [µB ](x) , W [µS ](x)

where W [µB ](x) and W [µS ](x) are the wavelet transforms at x (and at a given scale, as always) of the modeled background flux and of the modeled source flux, respectively. This estimator can be shown to be asymptotically unbiased (i.e., when the exposure time or all the fluxes tend to infinity), provided the models are correct. Confidence intervals on this statistic can also be derived.

Applications of the 2-D CWT. II: physical applications

(iii) How to estimate the position of a source? Intuitively, the source candidates should be located at the maxima of the wavelet transform. As above, this is true only for flat backgrounds. We account for its nonuniformity in a correction to the quantity to maximize in order to get the position of the source candidate. The position estimator is thus not argmaxx y obs (x) but rather % $ x ∗ = argmax y obs (x) − W [µ B ](x) . x

Again, assuming the models are correct, this subtraction restores asymptotic unbiasedness. Confidence regions corrected in this way may be seen on the example presented in Figure 5.7 below. (iv) How to choose the scale of the wavelet for best performance? Since the physical source is point-like, recorded sources look like the impulse response of the detector, i.e., the PSF. Hence, the best choice for the scale parameter is that leading to a wavelet with a width comparable to that of the PSF. This width does not vary much from source to source, it only depends on the energy of the incoming photons. A source emitting proportionally more at high energies than low energies is said to be “hard,” or to have a low spectral index and has a rather peaked PSF: it is best detected at small scale. On the contrary, a “soft” source has a rather flat PSF and is best detected at a larger scale. This dependence of the optimal scale parameter on the spectral index is illustrated on Figure 5.6. Number of σ

190

7

6

5

4

3

2 0

2

4

6

8 10 Scale parameter a

Fig. 5.6. Significance for the detection of a source, expressed in number of sigmas, as a function of

the scale parameter of the wavelet. The wavelet is centered at the position of the source. Each curve refers to a different spectral index, from lower (peaked curve) to higher (flat curve). The position of the maxima of these curves changes as the spectral index changes.

5.1 Astronomy and astrophysics

Most of the 270 EGRET sources are not identified. The intrinsic resolution of high energy gamma-ray detectors is not good enough to provide strong constraints on the source position; it is therefore difficult to find counterparts at other wavelengths. Their nature is still a mystery, that the next generation telescope GLAST will help to solve. A large fraction of identified sources consists in active galaxies whose nucleus is a massive black hole (up to 109 MSun ) surrounded by an accretion disk of matter falling in the gravitational well. In addition, strong jets of ultra-relativistic matter and radiation are emitted perpendicularly to the disk. Active galactic nuclei (AGN) detected in gammarays above 100 MeV have a jet pointed towards the Earth. Their emission is very variable, so that they are often undetected when they are in a quiescent state and then, in a short time, they become very bright. To give an example, EGRET viewing period 21.0 is a high latitude observation in which an AGN is in flaring state. 3EG J0237+1635 was detected at 10 σ in [213] and is 16 σ here. Several other sources are also present above 4 σ . The procedure described above has been applied and results are shown in Figure 5.7. One can see that all but one of the sources are seen. One should also note that the bright AGN position is slightly wrong, because of the presence of a faint source in the vicinity that is not detected. To detect this kind of source, the algorithm must be applied a second time after the addition of the significant sources to the background. This attempt at developing an alternative method to the usual maximum likelihood estimation will probably prove to be relevant in years to come. Indeed, the next generation gamma-ray telescope, GLAST, is to be launched in March 2006 and its complexity Declination

191

30

25

20

15

10

5

0

−5 70

60

50

40

30

20

10

Right Ascension

Fig. 5.7. Detected sources using the wavelet analysis of the EGRET viewing period 21.0 for E>100 MeV, the contours give the significance level from 4 to 16 σ . Superimposed stars give the positions of the third EGRET catalog sources detected over 4 σ (from [213]).

192

Applications of the 2-D CWT. II: physical applications

(parameters to take into account, volume of the data stream) will make it impossible to build a source catalog following EGRET’s old-school procedure. The algorithms must be made more efficient in some way. Wavelets will not be the key to the whole problem, of course, but will hopefully help develop alternative viewpoints.

5.2

Geophysics

5.2.1

Geology: fault detection As we have seen in Section 1.7, wavelet analysis was born in geophysics, as the empirical method designed by J. Morlet for analyzing the recordings of microseisms used in oil prospection. Thus it was to be expected that wavelets would soon find applications in other geophysical problems. It was indeed the case, as can be seen from the reviews [Fou94] or [250], where mostly 1-D applications are discussed, however. Then an interesting application of 2-D directional wavelets to geology was initiated in 1995 by Ouillon [Oui95,298]. The object to be analyzed is a system of geological faults covering a large area in the Arabian peninsula, which shows a self-similar behavior over scales from a few meters to hundreds of kilometers. Standard methods for analyzing such a system are based on renormalization group techniques or on the multifractal formalism (see Section 5.4). What the authors propose here is a continuous wavelet analysis, with directional wavelets, combined with a multifractal analysis. The motivation for this choice is that the relevant information to be measured is the anisotropy of the fault field, and the variation of this anisotropy with scale. In order to understand the idea, let us consider a synthetic so-called en e´ chelons fracture [298], depicted in Figure 5.8(a). At a small scale, the dominant orientation of this object is vertical, but at large scale, one sees only an oblique line pointing NE at 45◦ . Analyzing it with an isotropic wavelet would reveal these details, without focusing on directions, whereas an anisotropic one will enhance the direction response. Thus the authors of [298] chose the latter, namely an anisotropic Mexican hat (see Section 3.3.1). The originality of the method is an optimized local filtering of the WT, as follows. The wavelet used is an anisotropic Mexican hat with anisotropy factor $. One computes the WT of the image for a number of couples (θ, $). Then, for each point in the signal, one selects the pair (θ, $) that gives the largest value of the CWT among all those computed, the Optimum Anisotropic Wavelet Coefficient (OAWC). Thus, for each point, the OAWC selects the local filter which best matches the signal at the chosen resolution (scale). Then one thresholds the OAWC map in order to keep only the most significant features. The ridges of the remaining map correspond to the dominant structures detected. In the case of a fault array, these ridges (called virtual rupture lines or VLR) correspond to the faults as seen on a map. Finally, one draws a histogram of

193

5.2 Geophysics

Fig. 5.8. NOAWC analysis of the synthetic en e´ chelons fracture. Top row: (a) the signal; (b) the

NOAWC map at the resolution of 2 pixels; (c) the same at the resolution of 4 pixels. The bottom row shows the corresponding orientation roses (from [Gai00]).

the azimuths θ of the optimal wavelets associated with the points of the VLRs. This histogram, called a rose by geologists, depicts clearly the anisotropy of the object and its variation with scale. In further papers [120,181], this OAWC method was further improved by adding an adaptive normalization, in the sense that each OAWC is divided by its theoretical maximum corresponding to a perfect match between the wavelet and the object. The so-called NOAWC so obtained is a local indicator of the quality of the match. To give an example, we present in Figure 5.8 the NOAWC analysis of the en e´ chelons fracture signal (from [Gai00]). Panels (b) and (c) show the NOAWC map at the resolution of 2 and 4 pixels, respectively, with the VLRs enhanced in white, and on the bottom row, the corresponding orientation roses. Whereas the 2 pixels rose points at 90◦ (vertical orientation), that at 4 pixels resolution points at 60◦ . The interesting information is then the critical scale corresponding to a brutal shift in the orientation of the rose (see [Gai00] and [298,299] for more details). The NOAWC method has been applied successfully to the analysis of geological fault arrays and to the so-called rock fabric analysis, where “fabric” means the “complete spatial and geometrical configuration of all those components that make up a deformed rock” [Gai00,182]. As an example, we present in Figure 5.9 the NOAWC multiscale analysis of a 150-km-wide fault field, taken from [Gai00]. As in the previous

194

Applications of the 2-D CWT. II: physical applications

Fig. 5.9. NOAWC analysis of a real map of a fault field. (a) The original map; (b)–(f) NOAWC maps at scale a = 2, 4, 8, 16, 32. On the right, we show the corresponding orientation roses (from [Gai00]).

case, we show in the successive panels the NOAWC maps at smaller and smaller resolutions, together with the corresponding orientation roses. The dominant direction of the latter clearly varies with scale. The critical scales where transitions take place are then determined by a multifractal analysis. We refer to [299] and [Gai00] for further details. As already mentioned, the authors of all these papers use for the NOAWC method an anisotropic Mexican hat, which has a rather poor directional selectivity. However, the elliptical shape of the “footprint” of the wavelet plays an essential role in the method. This suggests the use of a Morlet wavelet instead of a Mexican hat. We should expect a much better precision, but the experiment has yet to be done.

5.2.2

Seismology As it is well-known [Bur98], the wavelet saga started with Jean Morlet, a French geophysicist working in oil prospection for Elf-Aquitaine. The technique consists of the sending of an impulse into the ground (by an explosion or any other means) and analyzing the signal reflected by the various discontinuities in the underground, down

195

5.2 Geophysics

x

y

x

t

t (a)

(b)

Fig. 5.10. (a) A seismic section; (b) a seismic block (from [Bou97]).

to 8000 m. These correspond to abrupt changes in density and composition of the rock, a necessary (but by far not sufficient!) condition for the presence of oil or gas. Clearly the resulting signal will be extremely noisy. Then, on a purely empirical basis, Morlet had the idea of representing this signal by a linear superposition of contributions obtained by dilating/contracting a fixed mother function (the analyzing wavelet). The method worked reasonably well, but it took a year of work between J. Morlet and the theoretical physicist A. Grossmann to understand its exact mathematical structure – namely, the content of Chapter 1 [199,205]. From there the theory of wavelets expanded in all directions, as we have seen in the preceding pages, until the loop was completed in the Ph.D. thesis of E. Bournay Bouchereau [Bou97], where she looked again at the very problem of seismic exploration treated by Morlet. The raw data are the so-called seismic sections, that is, 2-D plots where the vertical t axis represents twice the time needed by the wave to reach the corresponding rock layer (thus depth) and the x axis the horizontal distance between two successive receptors. A typical seismic section is shown in Figure 5.10(a). For a comprehensive study, one groups together a collection of parallel sections, thus getting a 3-D seismic block, as shown in Figure 5.10(b). Now, given a section, or a portion thereof, the goal is to detect the geological faults it contains. The technique developed in [Bou97] consists in taking the 2-D continuous WT of the section with a directional wavelet, either a Morlet wavelet or a separable directional one, on the model of (3.24). (The CWT is used here, instead of a discrete WT, in order to maintain translation covariance, which is essential for pattern identification.) This CWT easily detects faults,

196

Applications of the 2-D CWT. II: physical applications

(a)

(b)

Fig. 5.11. Detection of faults in a seismic section: (a) a section; (b) faults detected with a Morlet wavelet analysis (from [Bou97]).

which can be considered as lines of singularities. In the presence of a high noise (deep layers), a preliminary adaptive filtering improves the efficiency of the detection. An example of fault detection is given in Figure 5.11. In addition to the work just described, we ought to mention other applications of directional wavelets in seismology. For instance, the NOAWC method of the previous section has been used for the study of natural seismological events (spatial distribution of hypocenters of an earthquake sequence) [Gai00]. Even 3-D wavelets (see Section 9.1), namely a 3-D Mexican hat, have been used for the description of seismicity of a large area in the western Alps [75]; and, needless to say, 1-D wavelet analysis has also been applied to seismic time series, in particular, the arrival time of the various components, S-wave, P-wave, etc. (see, for instance, [104,297,Oon00]). Finally, it is amusing to note that essentially the same technique based on 2-D directional wavelets has been used for certain problems in metallurgy [236], owing to the similarity between a metallurgical image (inner structure of a piece of metal) and a seismic image.

5.2.3

Climatology Before concluding this section, we have to say a few words on the use of wavelet analysis in climatology. Most of the applications are in 1-D, typically time series analysis, using

197

5.3 Applications in fluid dynamics

often the discrete WT. We refer to [Fou94] or [250] for a review. An exception is the work of Hudgins et al. [227] on atmospheric turbulence, which uses the CWT in an essential way (this is the paper in which they introduced the wavelet cross-correlation function or wavelet cross-spectrum). Another one in the same spirit is that of van Milligen [273] on turbulence in fusion plamas. In both cases, the wavelet method turns out to be superior to the standard Fourier techniques. As for 2-D examples, the prime domain is again that of turbulence in fluids, that we will discuss in Section 5.3.1. In addition to the latter, an interesting application was made by Kumar [249], namely, to determine the so-called scale space anisotropy of geophysical fields. By this, one means that such fields are not only highly anisotropic over a wide range of scales, for dynamical reasons, but, in addition, different scale features are oriented in different directions (exactly as for the geological fault arrays described above). A typical example is provided by hurricanes, where the scale anisotropy is obvious. In many other cases, however, the anisotropy is present in a subtle way, that cannot be properly detected by classical techniques, such as spectrum- or correlation-based techniques. As an alternative, Kumar uses a 2-D Morlet wavelet analysis to characterize scale space anisotropy in radar-depicted spatial rainfall, by studying the fraction of energy in different directions at different scales. For that purpose, he introduces the relative scale-angle spectrum (2.58), which is sufficiently sensitive to reveal the subtle presence of scale space anisotropy in random fields. In the particular example treated here, the author is able to conclude that “ . . . a rainfall field might show an anisotropic structure that might not be obvious from a typical spectral analysis and may have wider implications in modeling and sampling problems.” Another application, closely related to the previous one, is the use of 2-D wavelets for enhancing thin-line features in meteorological radar reflectivity images [212]. Thin-line features in reflectivity correspond to surface wind convergence lines that can potentially lead to the initiation of thunderstorms. Thus the detection, preferably automatic, of such features is an important ingredient in the short time forecasting of thunderstorms. It turns out that a directional wavelet is required, namely a 2-D Morlet wavelet or a separable substitute built on the model (3.24). Once again, we see the superior discrimination power of directional wavelets in physical applications! In addition to the directional aspects analyzed in the previous applications, it is a fact that many meteorological phenomena have a distinctly fractal behavior. Clouds are a good example, but several artificial examples, such as random surfaces, share the same property. Thus, it is not surprising to find several applications of wavelets to such fractal structures. We will discuss some of them in Section 5.4.1.

5.3

Applications in fluid dynamics The wavelet transform, both continuous and discrete, has been successfully applied to the analysis of 2-D developed turbulence in fluids, especially for localization of coherent

198

Applications of the 2-D CWT. II: physical applications

structures in the distribution of energy or enstrophy. This topic is briefly described in Section 5.3.1. In addition, we will describe here two other applications of 2-D wavelets in fluid dynamics, which rely on the possibility of local filtering, both in direction and in position, with directional wavelets.

5.3.1

Detecting coherent structures in turbulent fluids Turbulence in fluids is a phenomenon that has resisted analysis until now. After more than a century of research, no real theoretical understanding of the dynamics of a turbulent flow has been achieved. There only exist various statistical or phenomenological models, which are widely used in practical applications, but lack a genuine justification. Even the very definition of the terms used does not always achieve a consensus among physicists. On the other hand, there is a huge amount of experimental data. In order to understand the rˆole of wavelets in this context, we have to go back to the basics. For a general review, we refer to [165]. The starting point is the system of Navier–Stokes (NS) equations governing the evolution of an incompressible Newtonian fluid: ∂ v v + 1∇ p = ν + ( v · ∇) v + F ∂t ρ · v = 0, ∇

(5.8) (5.9)

supplemented by adequate initial and boundary conditions. In these equations, v ≡ v( x , t) is the velocity, p ≡ p( x , t) the pressure, F the external force per unit mass, ρ a constant density, and ν the constant viscosity. The NS equations (5.8)–(5.9) are × v, which measures the local often expressed in terms of vorticity, namely ω =∇ rotation rate of the fluid. In dimension 2, the NS equations are formally the same, but the velocity field reads v( x ) = (u(x, y), v(x, y)) and the vorticity reduces to the pseudoscalar ω = ∂x v − ∂ y u. Then the fundamental quantities are the total energy and the total enstrophy, defined as, respectively 1 1 2 2 E(t) = d x | v ( x , t)| , Z (t) = d 2 x |ω( x , t)|2 , 2 2 and their Fourier transforms, the energy and enstrophy spectra ( is the volume occupied by the fluid). Fully developed turbulence is the regime of very large Reynolds numbers Re ∼ 1/ν, ν → 0 (in practice, in aeronautics, meteorology or combustion, for instance, Re varies v bebetween 106 and 1012 ). In this regime, the nonlinear advection term ( v · ∇) comes dominant, by several orders of magnitude. As with the semiclassical limit → 0 in quantum mechanics, this changes the character of the equation. As a result, not only is there no analytical solution known, but the NS equations cannot be solved numerically in this regime with present day computers, unless some drastic simplifications are made. Instead, since turbulent flows are highly unpredictable, one

199

5.3 Applications in fluid dynamics

has to use statistical models, requiring some basic assumptions, such as statistical homogeneity and isotropy, or ergodicity, which allows one to replace ensemble averages by space averages. All this led, for example, to the celebrated 1941 cascade model of Kolmogorov [165]. Yet turbulent fluids often exhibit coherent structures, that is, structures in the energy or the enstrophy spectrum that persist through a large range of scales (vorticity tubes, often called filaments), but are highly unstable. Clearly, the mere existence of these invalidates the statistical assumptions. In addition, the averaging processes, while satisfactory at low Reynolds numbers, ignore the coherent structures, since they have a small extent in space and in time. These are, however, an essential aspect of fully developed turbulence, and thus statistical models are inadequate for Re ! 1. This situation led Marie Farge to introduce, back in 1988 [163], wavelet methods for detecting and analyzing the time evolution of such coherent structures. Since then, she and her collaborators have devoted a huge amount of research work in this direction. Many different techniques have been used, wavelets (CWT and DWT), wavelet packets, multifractal techniques. In retrospect, the basic idea is always to separate the coherent structures, which are analyzed with wavelets, from the background flow, treated by statistical methods. This is, of course, not the place to go into the details of this considerable, but rather specialized work, which represents one of the most spectacular applications of wavelet analysis in physics. We refer the interested reader to the extensive review papers by M. Farge et al., which contain references to the original work [164,165,335]. In addition, a 1-D application to intermittent turbulence in atmospheric data is given in [211].

5.3.2

Directional filtering We will turn instead to applications that use specifically directional wavelets. As a consequence of their good directional selectivity, the Morlet and Cauchy wavelets are quite efficient for directional filtering. In order to illustrate the point, we analyze in Figure 5.12 a pattern made of rods in many different directions (a). Applying the CWT, with a Cauchy wavelet in a fixed direction (here horizontal), selects all those rods with roughly the same direction (b), whereas the other ones, which are misaligned, yield only a faint signal corresponding to their tips, in agreement with the behavior discussed above. Since this is in fact noise, one performs a thresholding to remove it, thus getting a clear picture (c). In this way, one can count the number of objects that lie in any particular direction. Note that the same pattern was analyzed with a Morlet wavelet in [18,19], and the result is slightly less neat. Figure 5.13 presents another example of directional filtering, this one with the Gaussian conical wavelet (3.37). The picture represents bacteria, seemingly at random. However, after filtering successively at −10◦ , 45◦ , and 135◦ , one realizes this latter orientation is significantly more populated.

200

Applications of the 2-D CWT. II: physical applications

(a)

(b)

(c) ◦

Fig. 5.12. Directional filtering with a Cauchy wavelet (A R P = 20 ) oriented at θ = 0◦ : (a) the

pattern; (b) the CWT; (c) the same after thresholding at 25%.

5.3.3

Measuring a velocity field In the first example [Wis93,375], the aim is to measure the velocity field of a 2-D turbulent flow around an obstacle. Velocity vectors are materialized by small segments, by the technique of discontinuous tracers. Tiny plastic balls are seeded into the flow and illuminated by a “plane of light,” in order to get a 2-D image. Then two successive photos are taken with a fast CCD camera, with exposure times of 700 and 6000 µs, respectively. In this way one gets a “dot-bar” signature for each tracer, which materializes

201

5.3 Applications in fluid dynamics

Fig. 5.13. Another example of directional filtering, with a Gaussian conical wavelet: (a) the original image, representing bacteria; (b) filtering at −10◦ ; (c) the same at 45◦ ; (d) the same at 135◦ .

the direction and the length of the local velocity (see an example in Figure 5.14, taken from [Wis93]). In order to get sufficiently many data points, one superposes several such pictures, typically 16. First one computes the WT of the resulting image with a Morlet wavelet, which selects those vectors that are closely aligned with the wavelet. Then a second analysis is performed with a wavelet oriented in the orthogonal direction, thus completely misoriented with respect to the selected vectors. Now the WT sees only the tips of the vectors and their length may be easily measured. The same two operations are then repeated with various successive orientations of the wavelet. Using appropriate thresholdings, the complete velocity field may thus be obtained, in a

202

Applications of the 2-D CWT. II: physical applications

Fig. 5.14. The dot-bar signature of tracers in the fluid flowing from left to right (from [Wis93]).

totally automated fashion, with an efficiency sensibly better than with more traditional methods. Two examples of reconstructed velocity fields from [Wis93] are given in Figure 5.15, corresponding to a quasi-laminar flow and a turbulent flow around an obstacle (again the flow comes from the left; units are normalized to the size of the experimental area). Notice that the analysis gives in principle both the modulus and the phase of the WT. But here, contrary to the simple applications like contour detection [13], the phase cannot be exploited, the data are too noisy. Thus one loses some precision on the orientation. Nevertheless, the method is remarkably efficient. In the same vein, the 2-D CWT (again Cartesian only) has been proposed for improving the method of holographic particle velocimetry, which consists in measuring the velocity of particles in a fluid by exploiting holograms of fluid volumes [10].

5.3.4

Disentangling of a wave train A second example originates from underwater acoustics. When a point source emits a sound wave above the surface of water, the wave hitting the surface splits into several components of very different characteristics (called respectively “direct,” “lateral,” and “transient”). The resulting wave train is represented by a linear superposition of damped plane waves, and the goal is to measure the parameters of all components. This phenomenon has been analyzed successfully with the WT both in 1-D [334] and in 2-D [18], and the extension to a 3-D version is straightforward. Let us give some details of the method in the 2-D case. The signal representing the underwater wave train is taken as a linear superposition of damped plane waves: f ( x) =

N

cn ei kn ·x e−ln ·x ,

(5.10)

n=1

where, for each component, kn is the wave vector, ln is the damping vector, and cn a complex amplitude. Then, using successively the scale-angle and the position

203

5.3 Applications in fluid dynamics y 0,5 0,4 0,3 0,2 0,1 0,0 −0,1 −0,2 −0,3 −0,4 −0,5 −9,68

−8,68

−7,68

−6,68

−5,68

−4,68

x

(a) y 2,0 1,5 1,0 0,5 0,0 −0,5 −1,0 −1,5 −2,0 −2,0

x −1,5

−1,0

−0,5

0,0

0,5

1,0

1,5

2,0

2,5

3,0

3,5

(b) Fig. 5.15. Two examples of reconstructed velocity fields: (a) a quasi-laminar flow; (b) a turbulent flow around an obstacle (from [Wis93]).

representations described in Section 2.2.3, one is able to measure all the 6N parameters of this signal with remarkable ease and precision. The method proceeds in three steps and uses explicitly the phase space interpretation. First one computes the CWT of the signal (5.10) with a Morlet wavelet. By linearity, the result is the linear superposition of the contributions of the various components. Moreover, each component is the product of two factors, where the first one depends on b only and the second one on (a, θ) only: a, θ ) = F(b,

N n=1

ˇ cb,n Fn (a, θ).

(5.11)

204

Applications of the 2-D CWT. II: physical applications

Actually, the resulting function may be written explicitly in terms of the phase space variables introduced in Section 2.3.2, for instance, Fˇn (a, θ ) ≡ Fˇn ( v ). Now we go to the scale-angle representation and consider the WT (5.11) for fixed Then a straightforward calculation shows that, for each term in this superposition, b. a −1 Fˇn (a, θ ) admits a unique local maximum. Now, in the full transform (5.11), each term has its own local maximum, but these need not be well separated: one maximum may hide another one, totally or partially. This masking effect will happen, for instance, when: r one component has a much bigger amplitude, |c | ! |c |, for all m not equal to n b,n b,m (total masking); r two wave vectors are close to each other, k k , but with different amplitudes, n m |cb,n | > |cb,m | (partial masking). In both cases, the two waves can be separated, by increasing the selectivity of the wavelet (for instance, using a Morlet wavelet with a more anisotropic modulus). If the two waves have close wave vectors (kn km ) with similar amplitudes (|cb,n | |cb,m |), but different damping vectors (kn = km ), then they can still be separated, by changing the Otherwise the method will fail, the two waves interfere inextricably, observation point b. none of them dominates the other one. When the masking effect is not too important, the maxima will be sufficiently prominent that the interferences between the different components will become negligible (in the modulus) and one may write: a, θ )| |F(b,

N

ˇ |cb,n | | Fn (a, θ)|,

(5.12)

n=1

One then reverts to the position representation, choosing for (a, θ ) each maximum successively. In each case, the filtering effect of the CWT essentially eliminates all components except one, which is then easy to treat. In this way, one is able to measure all the 6N parameters of the signal easily. A striking example is given in [18], illustrating the power of the method as well as the rˆole of the anisotropy factor $ of the Morlet wavelet. The signal is the superposition of four damped plane waves, with different wave vectors kn , except that the directions of k1 and k4 differ by 20◦ only. As a result, wave # 1 partially masks wave # 4: when the analysis is performed with a Morlet wavelet with $ = 1, the corresponding maxima in the scale-angle representation are not well separated [Figure 5.16(a)]. When one uses instead a wavelet with $ = 5, the two maxima are clearly identifiable and and can be localized precisely (b). Notice that the “footprint” of the wavelet is not an ellipse, because the radial coordinate used is a −1 , not a. Now the procedure allows to reconstruct each of the four components almost perfectly [18]. Only wave # 4 keeps some trace of interference with wave # 1, the others are indeed pure waves. In order to remove this effect, one should first subtract wave # 1 from the signal and redo the analysis.

205

5.4 Fractals and the thermodynamical formalism

2 3

2

3

1

1

4

4

(a)

(b)

Fig. 5.16. Disentangling of a four component wave train with a Morlet wavelet: the four maxima (a) with $ = 1; (b) with $ = 5 (from [18]).

5.4

Fractals and the thermodynamical formalism

5.4.1

Analysis of 2-D fractals: the WTMM method Many physical phenomena require a wide range of scales for a complete description of their properties. The paradigm, of course, are fractals, which are complex mathematical objects that have no natural length scale. More precisely, a fractal, be it in 1-D or in 2-D, is by definition self-similar under dilation, either globally (genuine fractal) or locally (multifractal). Physical examples abound. For instance, all kinds of random walks used to mimic various noisy dynamical behaviors, financial time series, geologic shapes (such as the fault systems descibed in Section 5.2.1), interfaces that develop in growth processes far from equilibrium, fractal growth processes (such as the so-called diffusion-limited aggregates), electrodeposition clusters, etc. A fractal is in general a very irregular object (for instance, its support may be a Cantor-like set), hence it should be represented by a singular measure, rather than a function. In order to cope with such situations, the so-called multifractal formalism has been developed. Now, central concepts of the theory, such as generalized fractal dimensions or spectrum of singularities of the measure, are closely related to ideas from statistical mechanics (as a matter of fact, the standard “box counting” method is already of a statistical nature). Thus one speaks also of a thermodynamical formalism of fractal analysis. A review of this formalism may be found in [47] or [293]. Since scaling is the most significant operation in the context of fractals, the CWT is a natural tool for analyzing them. Clearly the continuous version of the WT is essential here, since the characteristic scaling ratio is unknown a priori. The first step is to extend

206

Applications of the 2-D CWT. II: physical applications

the CWT to singular measures. This was achieved in 1-D by Holschneider [221] and Arn´eodo et al. [45], then extended to 2-D by Arn´eodo and his group in Bordeaux, including one of us (R.M.) [43]. We briefly describe the key steps of the formalism. We take the most general case, namely, an object described by a fractal measure µ on R2 . The standard thermodynamical formalism (spectrum of generalized fractal dimensions) yields only statistical information about the object as a whole. To get precise local information requires a wavelet transform. The CWT of the measure µ with respect to the wavelet ψ is defined as T [µ](b, a, θ) = dµ( x ) ψ(a −1r−θ ( x − b)). (5.13) In the isotropic case, we write simply x − b)). T [µ](b, a) = dµ( x ) ψ(a −1 (

(5.14)

Assume now that the measure has the following scaling behavior (self-similarity) around the point xo : x , $)), µ(B( x , λ$)) ∼ λα(xo ) µ(B(

λ > 0,

(5.15)

xo ) is the local scaling exponent. where B( x , $) is a ball of radius $ around xo and α( Using the covariance property of the CWT (Proposition 2.2.3), it is easily shown that the WT of the measure µ scales in the same way: λa, θ ) ∼ λα(xo ) T [µ]( a, θ), T [µ]( xo + λb, xo + b,

λ → 0+ .

(5.16)

This relation is the key to the wavelet analysis of fractals. For instance, the local exponent versus log a, for a small enough. α( xo ) may be obtained by plotting log |T [µ](a, θ, b)| This would suffice for an exact (global) fractal, such as a numerical snowflake, for which α is constant over the whole object. For a genuine multifractal, α( xo ) varies from point to point, and then (5.16) allows one to compute the generalized fractal dimensions and the singularity spectrum of the object. This wavelet approach to the thermodynamical formalism of fractal analysis has been developed systematically by Arn´eodo and his collaborators in Bordeaux. It is instructive to give some hints about the basic ideas of the formalism. We first go back to the standard multifractal formalism. Given a singular measure µ, it can be described in terms of its q-moments as follows. Suppose we cover the support of µ with N ($) boxes Bi ($) of size $. Then the scaling behavior of µ will be deduced from the partition function Z(q, $) =

N ($) i=1

q

µi ($),

where µi ≡ µ(Bi ($)), q ∈ R.

(5.17)

207

5.4 Fractals and the thermodynamical formalism

In the limit $ → 0+ , Z(q, $) behaves as a power law: Z(q, $) ∼ $ τ (q) , $ → 0+ . Finally, the spectrum of generalized fractal dimensions is obtained from the exponents τ (q) by the relation Dq = τ (q)/(q − 1), and these in turn completely characterize the singular behavior of the original measure µ, in particular, the local H¨older exponents that describe it (see [47,292]). Now the rationale for reinterpreting this formalism in terms of wavelets is that the WT tends to “forget” the regular part of the signal (because of the vanishing moments) and focus on the singular part. A naive way of achieving this would be to take as partition function, instead of (5.17), Z(q, a) = d x |T [µ]( x , a)|q , q ∈ R (5.18) (in 1-D, one uses simply |T (x, a)| as WT modulus, see (2.66)). This is a bad choice, however, since Z(q, a) might diverge for q < 0. Instead one replaces the integral over x by a discrete sum over the local maxima of T [µ]( x , a), for fixed a, that is, precisely the WTMM introduced in Section 2.3.5. The justification of this choice is that [262] (i) the maxima lines (ridges) have the same scaling behavior as the WT itself, and (ii) each maxima line l = (bl (a), a) points, as a → 0, to a point bl (0) which corresponds to a singularity of µ and, in addition, the WT modulus scales along the line as |T [µ](bl (a), a)| ∼ a α(bl (0)) .

Thus the wavelet plays the role of a generalized “oscillating box” and the scale a defines its size. Actually, the definition of the partition function can be further refined by using explicitly the WTMMM, as follows. Let L(a) denote the sets of ridges that exist at scale a and contain a maximum at a scale a a. Then one defines finally the partition function q

Z(q, a) = |T [µ]( x , a )| . (5.19) sup l∈L(a)

( x ,a )∈l,a a

In fact, introducing the “sup” amounts to adapting the size of the wavelet along the ridge so as to avoid divergences. Here again, the exponents τ (q) are defined from the power-law behavior of Z(q, a) as a → 0: Z(q, a) ∼ a τ (q) , a → 0+ . (In the analogy with thermodynamics, q and τ (q) play the rˆole of inverse temperature and free energy, respectively.) The WTMM technique has been applied successfully to a wide variety of examples [43,44,47], that cover both artificial fractals (numerical snowflakes, diffusion limited

208

Applications of the 2-D CWT. II: physical applications

aggregates, recursive fractal functions) and natural ones (electrodeposition clusters, various arborescent phenomena, fully developed turbulence data, clouds). The method permits the measurement of the fractal dimensions and the unraveling of universal laws (mean angle between branches, azimuthal Cantor structures, etc.). In 2-D, it should be remarked that the analysis uses exclusively an isotropic wavelet (usually a 2-D Mexican hat), and thus there is no θ dependence in (5.16). However, this may not be the end of the story. Indeed we shall exhibit in Section 4.5.2 below a fractal (“twisted snowflake”) whose structure requires a directional wavelet for its complete determination. In more recent work, the attention has focused on 2-D applications, around the theme of rough surfaces: fractional Brownian surfaces, anisotropic self-affine rough surfaces [Dec00,53], synthetic multifractal rough surfaces [129], cloud structure [52,54]. This last item opens a whole world of applications to physical processes, in particular to meteorology, since fractal objects abound there. Another application, closely related to the previous analysis, concerns the analysis of real rough surfaces, that is, metal surfaces obtained after various kinds of machining processes. Here too, the 2-D wavelet transform yields a useful tool [235]. Actually the object to analyze is essentially the texture of the surface, which brings us to another field of application, namely, texture analysis, that we will discuss in Section 5.5. A last point to notice in this context is that, considering the heavy computational cost of the 2-D CWT, the Bordeaux group has designed an ingenious hardware version, called the Optical WT [Arn95,46]. The technique, based on Fraunhofer diffraction, a familiar tool in optics, amounts to obtaining the WT with a binary approximation to the isotropic Mexican hat, that is, using as isotropic wavelet the Bessel filter described in (3.9)– (3.10). With this tool, they obtained beautiful pictures of CWT analyses of diffusionlimited aggregates [Arn95,43,46]. The technique could not, however, be pushed to full implementation for lack of a sufficiently fast CCD camera – or, equivalently, a sufficiently large budget! Actually, several optical implementations of the WT have been proposed in the literature. However, they are often limited to 1-D data or, in the 2-D case, to Cartesian coordinates (translations and separate scaling factors in the x and y directions). We refer the interested reader, for instance, to a feature issue of Applied Optics dedicated to this topic [255]. An alternative optical approach, based on a special type of grating and able to reproduce many types of wavelets, has been presented in [269]. The isotropic Mexican hat has also found applications in optics proper, for instance, in the determination of the wave aberration coefficients of a rotationally invariant optical system, from the measured data of wave front deformations [340]. The result is that the wavelet method is more efficient and more robust to noise than standard least squares methods. Another application of the 2-D CWT in optics is to moir´e interferometry, a well-established optical technique for measuring displacement and strain in materials, based on phase analysis of interference fringe patterns [238]. Thus a complex wavelet is needed, and the authors resort to a modified Morlet wavelet, obtained by replacing

209

5.4 Fractals and the thermodynamical formalism

the Gaussian by a cubic spline, as advocated by Unser [357], and once again, the CWT method proves superior to the standard Fourier techniques.

5.4.2

Shape recognition and classification of patterns The characterization of a 2-D shape from its outlines is an important problem in several applications of image analysis, such as character recognition, machine parts inspection for industrial applications, characterization of biological shapes such as chromosomes and neural cells, and so on. Furthermore, in the field of human vision and perception, 2-D shape analysis also plays a central role in psychophysics and neurophysiology. There are two general approaches to shape characterization: region based, which deals with the region in the image corresponding to the analyzed object; and boundary based, where the shape is characterized in terms of its silhouette [Pav77]. The former is intrinsically 2-D, dealing directly with planar primitives and concepts, and thus 2-D wavelets may be used directly. The latter, however, mimics 2-D operations through 1-D representations, and is referred to henceforth as contour characterization. An alternative to standard techniques consists of representing the shape by the complex curve that describes its boundary, and applying the 1-D CWT to this complex signal [Ces97,21], as outlined in Figure 5.17. This leads to the so-called W-representation,

SEGMENTATION

CONTOUR EXTRACTION

300 t

250 x(t)

200 CONTOUR TRACKING

150 100 50 0 0

y(t)

200

400

600

800

1000 t

Fig. 5.17. Basic framework for 2-D shape characterization from its outline (from [21]).

210

Applications of the 2-D CWT. II: physical applications

which allows an easy way of performing a number of standard tasks (for instance, in machine vision), such as detection of dominant points, shape partitioning, natural scales analysis. As compared to other techniques, the W-representation has the following useful characteristics, which are desirable for purposes of shape analysis [275] and follow directly from the basic properties of the CWT: r uniqueness, because of the invertibility of the CWT; r invariance under translation, scaling and rotation; r robustness to local modifications of the shape; r efficiency and ease of implementation. Notice that an essential ingredient of the technique is the wavelet-based fractal analysis discussed above. In particular, one resorts systematically to the different types of local maxima lines defined in Chapter 2, Section 2.3.5. More precisely: r The algorithm for the detection of dominant points, e.g. corners, is based on the vertical maxima lines, namely, their position and some relevance measure, for instance, their length. r The detection of periodic patterns and the so-called natural scales is based on the horizontal maxima lines, since this amounts essentially to determining the instantaneous frequencies in the signal. r Finally, the fractal behavior of certain contours is analyzed with the standard technique sketched in Section 5.4.1. This analysis has numerous applications. An original one is the classification of neurons according to their complexity, that is, their fractal dimension [Ces97]. Of course, shape analysis is a whole different world. For an up-to-date review of it, we refer to the recent monograph of L. da F. Costa and R. M. Cesar, Jr. [Cos01].

5.5

Texture analysis The determination and classification of textures in images is an old and difficult problem, with many potential applications, notably in computer vision. Numerous methods have been designed to that effect. Most of them are of a statistical nature, such as Markov random fields, but one has also used Gabor analysis [234] and various kinds of wavelet transforms. We shall concentrate on the latter, of course. Some proposals have been made with the standard (Cartesian) discrete WT (see [166] for instance), but, since most textures possess directional features, it is more natural to use oriented wavelets for attacking the problem. An example is the steerable wavelet pyramid developed by Simoncelli and his collaborators (see Section 2.7). This technique was used in [317] as a basis for a parametric texture model based on joint statistics of wavelet coefficients. It has been considerably developed by Do and Vetterli [Do01,140,141], using appropriate generalizations of the ridgelets and curvelets (these

211

5.5 Texture analysis

will be described in Chapter 9 and Section 11.1), and combining them with waveletdomain hidden Markov models. Alternatively, one can use the plain 2-D CWT with an oriented wavelet. Actually a similar solution was proposed long ago by Rao and Schunck [323], using the first derivative of a 2-D Gaussian, (3.15), already considered by Canny [98] (see Section 3.3.1.1). This is indeed an oriented wavelet, but it is not directional in the technical sense. (In addition, one may find in [323] a survey of the early literature on texture determination.) Texture analysis with a genuine directional wavelet (a Morlet wavelet) was done first by Gonnet [195], using the characterization of the instantaneous frequency of the signal as a vector field [196] (Section 2.3.5). Further results were obtained by Murenzi et al. [290], using the same wavelet (truncated Morlet, also called Gabor filter) and the scale-angle representation. A more efficient technique yet is that of the directional dyadic wavelet packets [360] (Section 2.6.4). Substantial advances in the classification of textures have been obtained recently along this line by Menegaz et al. [270]. As usual in the framework of computer vision, a pattern is represented by a feature vector, as explained in Chapter 4, and the classification is made in defining similar images as those whose feature vectors are close to each other, in the sense of some notion of distance (often Euclidean or L 2 ). In order to use directional wavelet packets for this problem, one simply includes the directional details at each scale among the components of the feature vectors. The latter become longer, but the efficiency of the method for texture discrimination increases significantly. A related topic is the so-called shape from texture problem, which can be formulated as follows. We are given a 2-D photograph of a 3-D surface, which displays a pattern or a texture, more or less regular. The image gives a distorted view of this texture, which depends of the geometry of the surface. The goal is to reconstruct the original surface from the distorted image. The problem is usually split into two steps. First one measures the local distortion of the image, then one recovers the original surface. For the first step, some assumptions are usually made about the type of surface and texture. It turns out that the CWT is an efficient tool for the estimation of distortions, as first proposed by Super and Bovik [348] and Hwang et al. [229]. The planar surface is assumed to have a homogeneous texture, which is modeled as a linear superposition of plane waves. A further paper [231] considers a plane containing several textures with different orientations, a situation which requires first a segmentation step. In both cases, the authors rely on the properties of the ridges of the 2-D wavelet transform (Section 2.3.5). Further work along the same line was made in [179], exploiting the well-known covariance properties of the CWT under translation, rotation, and scaling. Here too, the authors consider a single homogeneous (sinusoidal) texture, and study its deformations. An alternative to this deterministic model is to describe the texture by a stationary random process. This approach has been developed by Clerc and Mallat [Cle99,106,107] for the general case of a curved surface. In addition, they consider general distortions,

212

Applications of the 2-D CWT. II: physical applications

for instance, anisotropic scalings. Since the conventional CWT is no longer covariant under such transformations, they introduce instead the so-called warplet transform, replacing the familiar global scaling by a distortion matrix which is not a multiple of the identity (the name refers to the fact that the image presents a warped view of the surface). This technique, while conceptually interesting, leads to high computational costs, in particular, the formalism of [179] cannot cope with it.

5.6

Applications of the DWT For the sake of completeness, we conclude this chapter with some remarks on the applications of the DWT. As we said already, the latter is used in the majority of applications, but this is not the main subject of the present book. Thus we will give only a few indications. As with other methods, wavelet bases may be applied to all the standard problems of image processing. The main problem of course is data compression, and for achieving useful rates one has to determine which information is really essential and which one may be discarded with acceptable loss of image quality. Significant results have been obtained in the following directions. r Representation of images in terms of wavelet maxima [264], as a substitute for the familiar zero-crossing schemes [Mar82]. r In particular, application of this maxima representation to the detection of edges, and more generally detection of local singularities [262]. r Image compression and coding using vector quantization combined with the WT [41]. r Image compression, combining the previous wavelet maxima method for contours and biorthogonal wavelet bases for texture description [176]. r Image and signal denoising, by clever thresholding methods [144]. Some applications are less conventional. For instance, a technique based on the biorthogonal wavelet bases [108] has been adopted by the FBI for the identification of fingerprints. The advantages over more conventional tools are the ease of pattern identification and the superior compression rates, which allows one to store and transmit a much bigger amount of information in real time. The full story may be found in [87]. Another striking application is the deconvolution of noisy images from the Hubble Space Telescope, by a technique combining the DWT with a statistical analysis of the data [Bou93,85,328]. The results compare favorably in quality with those obtained by conventional methods, but the new method is much faster. One should also quote a large amount of work under development in the field of High Definition Television, where wavelet techniques are being actively exploited; here again the huge compression rates make them specially interesting. As for applications of the multidimensional DWT more specifically oriented to physics, we like to mention two. The first one is in quantum field theory (although

213

5.6 Applications of the DWT

it was done before the wavelet techniques were born): various perturbation expansions (the so-called “cluster expansion”) used in the analysis of Euclidean field theory models are in fact discrete wavelet expansions [65]. Actually the summation over scales, indexed by j, was originally motivated by renormalization group arguments. In the same domain, we may note that wavelet bases have been used also ( [66] and references therein) for estimating the time evolution of solutions of some wave equations (Klein-Gordon, Dirac, Maxwell or the wave equation), or even to expand solutions of the equations in terms of dedicated “wavelets” (although the functions introduced in the last case seem rather far away from genuine wavelets [239]). The other application resorts to solid state physics, namely the Quantum Hall Effect (quantization of the electric conductivity) that occurs when a 2-D electron gas is submitted to a strong transverse magnetic field. Here orthonormal wavelet bases may be used for generating localized orthonormal bases for the lowest Landau level, a necessary step towards the analysis of the Hall effect [14,36,57,58].

6

Matrix geometry of wavelet analysis. I

6.1

Group theory and matrix geometry of wavelets In Chapters 1 and 2, we have studied systematically the continuous wavelet transform in one and two dimensions, respectively. As already emphasized there, the properties of the transforms in the two cases are remarkably similar. In 2-D we have formalized them in the three propositions 2.2.1, 2.2.2 and 2.2.3, and essentially the same statements may be made in 1-D. A moment’s reflection shows that one could write out, without difficulty, an entirely parallel mathematical description in any dimension n 1. Clearly there must be some unifying principle underlying the picture. The question is, of course, what is this principle? As so often in such situations, the answer is to be found in group representation theory, i.e., by looking at the underlying geometry of the space of signals. The various transformations (translation, rotation, zoom, etc.) that a signal may undergo, determine a set of mathematical symmetries, which, interestingly enough, can be expressed in simple matrix terms and, as will be made clear in the following, the signal space itself – as a mathematical object – emerges as a consequence of this geometry. But we have been using group theory all along! Indeed, to draw on a literary analogy, like Moli`ere’s Monsieur Jourdain speaking in prose without knowing so, we have been using group-theoretical language throughout our analysis! It is the aim of the present chapter to demonstrate this fact. By so doing, we intend to achieve three goals. First, group theory will provide us with a unifying picture to view all the mathematical properties of signals that we have derived by hand, so to speak. Second, we hope to convince the reader that the group-theoretical approach is not only aesthetic, it is also simpler, in that it allows us to understand the deeper mathematical structures involved in a simple language. Finally, rewriting our results in this language allows us to easily extend the concept of a CWT to more general manifolds (i.e., more complex spaces of signal parameters), which we shall explore in later chapters: the n-dimensional space Rn , space–time R1+1 or R1+3 , the two-sphere S 2 or the n-sphere S n , etc. Indeed, as recalled in the Prologue, this was the decisive factor in extending the CWT from one to two dimensions in R. Murenzi’s thesis [Mur90], the key step being to identify the

214

215

6.1 Group theory and matrix geometry of wavelets

relevant group and its realization in the space of signals. To set the stage, let us analyze this mathematical structure in some detail. We shall do this first in 1-D, in this chapter, and then take up the 2-D problem in the next chapter. We emphasize that the next three chapters, while interesting and intellectually satisfying, are not prerequisites for the last three, Chapters 9 to 11, which will treat various extensions of the wavelet transform studied so far. In addition, we have collected in the Appendix some pertinent definitions and results from group theory, that we hope would help the group theoretically uninitiated reader to understand the material and to appreciate better the breadth of its scope.

6.1.1

The 1-D CWT revisited We start with the basic 1-D transformation (1.4): x −b −1/2 ψ(x) → ψb,a (x) = |a| ψ , b ∈ R, a = 0, a and rewrite it in the form ψb,a (x) = |a|−1/2 ψ (b, a)−1 x ,

(6.1)

(6.2)

where we have introduced the affine transformation of the line, consisting of a dilation (or scaling) by a = 0 and a (rigid) translation by b ∈ R: x = (b, a)y = ay + b,

(6.3)

and its inverse y = (b, a)−1 x =

x −b . a

(6.4)

Writing φ = ψb,a and making a second transformation on φ we get φ(x) → φb ,a (x) = | a | − 2 φ((b , a )−1 x) 1

= | aa | − 2 ψ((b , a )−1 (b, a)−1 x) x − (b + ab )

− 12 = | aa | ψ . aa 1

(6.5)

This shows us that the effect of two successive transformations is captured in the composition rule (b, a)(b , a ) = (b + ab , aa ), which, if we represent these transformations by 2 × 2 matrices of the type a b (b, a) ≡ , a = 0, b ∈ R, 0 1

(6.6)

(6.7)

216

Matrix geometry of wavelet analysis. I

is reproduced by ordinary matrix multiplication. The point to be noted about these matrices is that the product of two of them is again a matrix of the same type and so also is the inverse, −1 a b a −1 −a −1 b −1 (b, a) = = 0 1 0 1 of such a matrix. Furthermore, the 2 × 2 identity matrix is also in this class. In other words, the set of matrices (6.7) constitute a group, called the (full) affine group and denoted G aff . Note also, that if we consider only those matrices in (6.7) for which a > 0, then this set is also stable under multiplication and inverse taking and hence, it constitutes a subgroup of G aff , denoted G + aff . Coming back to the relation (6.2), we + observe that G aff or G aff consists precisely of the transformations we apply to a signal: translation (time-shift) by an amount b and zooming in or out by the factor a. Hence, the group G aff relates to the geometry of the signals. Next let us study the effect of the transformation given by the group element (b, a) on the signal itself. Writing, ψ → U (b, a)ψ ≡ ψb,a , we may interpret U (b, a) as a linear operator on the space L 2 (R, d x) of finite energy signals, with the explicit action, x −b . (6.8) (U (b, a)ψ)(x) = |a|−1/2 ψ a Additionally, for each (b, a), the operator U (b, a) is unitary, i.e., it preserves the Hilbert space norm of the signal: ∞ d x | ψ(x) | 2 . ψb,a 2 = ψ2 = −∞

More interestingly, the association, (b, a) → U (b, a) is a group homomorphism, preserving all the group properties. Indeed, the following relations are easily verified: U (b, a)U (b , a ) = U (b + ab , aa ) U ((b, a)−1 ) = U (b, a)−1 = U (b, a)† U (e) = I, with e = (0, 1), the unit element.

(6.9)

We say that the association (b, a) → U (b, a) provides us with a unitary representation of G aff . Note that we may also write, U (b, a) = Tb Da , where Tb , Da are the well-known shift and dilation operators, familiar from standard time-frequency analysis (see also (2.7)–(2.12)): (Tb s)(x) = s(x − b),

(Da s)(x) = |a|− 2 s(a −1 x). 1

217

6.1 Group theory and matrix geometry of wavelets

We shall see later that the representation U (b, a) is in a sense minimal or irreducible, in that the entire Hilbert space of finite energy signals L 2 (R, d x) is needed to realize it completely. However, let us first attend to another question which is pertinent here, namely, why is it that G aff is made to act as a transformation group on R (see (6.3) and (6.4)), even without manifestly identifying R with any set of signal parameters? The answer to the above question lies in realizing that this space is intrinsic to the group itself. Indeed, let us factor an element (b, a) ∈ G aff in the manner a b 1 b a 0 (b, a) = = , (6.10) 0 1 0 1 0 1 and note that the first matrix on the right-hand side of this equation basically represents a point in R. We also note that the set of matrices of the type appearing in the second term of the above product is a subgroup of G aff . Dividing out by this matrix, we get (b, a)(0, a)−1 = (b, 0), which enables us to identify the point b ∈ R with an element of the quotient space G aff /H , (where H is the subgroup of matrices (0, a), a = 0). Next we see that, since a b 1 x a ax + b 1 ax + b a 0 = = , 0 1 0 1 0 1 0 1 0 1 the action of the group G aff on its quotient space G aff /H is exactly the same as its action on R as given in (6.3). Thus, the parameter space R on which the signals ψ(x) are defined is a quotient space of the group and hence intrinsic to the set of signal symmetries. We shall see below that the parameter space on which the wavelet transform of ψ is defined can also be identified with a quotient space of the group. In fact this space will turn out to be a phase space, in a sense to be specified later. Before moving on, let us reemphasize that the philosophy which seems to be emerging here is that the group (of signal symmetries) is the determinative quantity and all aspects of the signal and its various transforms emanate from it. We come back now to the point made earlier, that the representation U (b, a) was irreducible. We shall see that it also enjoys a second crucial property, that of being square integrable. The group G aff has a natural action on itself (by matrix multiplication from the left), according to which, for a given (b0 , a0 ) ∈ G aff , a general point (b, a) ∈ G aff is mapped to (b , a ) = (b0 , a0 )(b, a) = (b0 + a0 b, a0 a). It is not hard to see that the measure dµ(b, a) =

db da , a2

(6.11)

is invariant under this action: db da db da = . a2 a 2

(6.12)

218

Matrix geometry of wavelet analysis. I

We call the measure dµ the left Haar measure of G aff . In a similar manner we could obtain a right Haar measure dµr (invariant under multiplication from the right), which would turn out to be dµr (b, a) = a −1 db da. It is important to realize, that while these two measures are (measure theoretically) equivalent, they are not the same measure. The function *(b, a) = a −1 , for which dµ(b, a) = *(b, a) dµr (b, a), is called the modular function of the group. The square-integrability of the representation U (b, a) now means that there exist signals ψ ∈ L 2 (R, d x) for which the matrix element U (b, a)ψ | ψ is square integrable as a function of the variables b, a, with respect to the left Haar measure dµ, i.e., dµ(b, a) |U (b, a)ψ|ψ|2 < ∞, (6.13) G aff

and a straightforward computation would then establish that the function is also square integrable with respect to the right Haar measure. Furthermore, it is a fact that the existence of one such (nonzero) vector implies the existence of an entire dense set of them. Indeed, the condition for a signal to be of this type is precisely the condition of admissibility required of mother wavelets. To derive the admissibility condition, and also to verify our claim of irreducibility of the representation U (a, b), it will be convenient to go over to the Fourier domain. It is not hard to see that, on the Fourier-transformed space, the unitary operator U (b, a) (b, a), with explicit action, transforms to U (b, a)ψ (ξ ) = |a|1/2 ψ(aξ )e−ibξ (b ∈ R, a = 0). U (6.14) , dξ ) the image of The Fourier transform is a linear isometry, and we denote by L 2 (R 2 L (R, d x) under this map. It follows that the operators U (b, a) are also unitary and , dξ ) ∈ L 2 (R that they again constitute a unitary representation of the group G aff . Let ψ be a fixed nonzero vector in the Fourier domain. We will now show that the set of all , dξ ) and this is what will (b, a)ψ as (b, a) runs through G aff is dense in L 2 (R vectors U . Indeed, let constitute the mathematically precise statement of the irreducibility of U , dξ ) be a vector which is orthogonal to all the vectors U (b, a)ψ: χ ∈ L 2 (R (b, a)ψ = 0, ∀b ∈ R, a = 0. χ |U Using (6.14) we get, (b, a)ψ = |a|1/2 χ |U

∞

−∞

(aξ ) e−ibξ = 0. dξ χ (ξ ) ψ

(aξ ) = 0, almost every (ξ ) ψ By the unitarity of the Fourier transform, this yields χ where, for all a = 0. Since ψ ≡ 0, this in turn implies χ (ξ ) = 0, almost everywhere. , dξ ) which are stable under the action of all the Thus, the only subspaces of L 2 (R , dξ ) itself and the trivial subspace containing just the zero (b, a) are L 2 (R operators U , dξ ) is sort of a minimal space for the representation. vector. In other words, L 2 (R

219

6.1 Group theory and matrix geometry of wavelets

The unitarity of the Fourier transform also tells us that the representations U (b, a) and (b, a) are equivalent and since U (b, a) is irreducible, so also is U (b, a). (Note, this U is also clear from the fact that the linear isometry property of the Fourier transform implies that (b, a)ψ, χ | U (b, a)ψ = χ |U respectively.) χ , ψ denoting the inverse Fourier transforms of χ , ψ, Now we address the question of square integrability. We require that da db ψ| 2= |U (b, a)ψ| a2 G aff da

) ψ(aξ )ψ(ξ ) ) ψ(ξ db eib(ξ −ξ ) ψ(aξ = dξ dξ |a| da )|2 )|2 |ψ(ξ = 2π dξ |ψ(aξ |a| ∞ dξ 2 |ψ(ξ )| < ∞ = 2π ψ2 −∞ |ξ | (the integral over b yields a delta distribution, which can be used to perform the ξ integration and the interchange of integrals can be justified using standard distribution theoretic arguments). This means that the vector ψ is admissible in the sense of (6.13) if and only if ∞ dξ 2 cψ ≡ 2π (6.15) |ψ(ξ )| < ∞. −∞ |ξ | From this discussion we draw two immediate conclusions. First, there is a dense set of which satisfy the admissibility condition (6.15). Second, the admissibility vectors ψ condition (1.10) or (6.15), cψ < ∞, simply expresses the square integrability of the representation U . (Note that a vector ψ ∈ L 2 (R, d x) is admissible if and only if its , dξ ), on L 2 (R Fourier transform satisfies (6.15).) Defining an operator C !1 2π 2 (C ψ)(ξ ) = ψ(ξ ), (6.16) |ξ | and denoting by C its inverse Fourier transform, we see that the vector ψ is admissible if and only if cψ = Cψ2 < ∞ .

(6.17)

This operator, known as the Duflo–Moore operator, is positive, self-adjoint and unbounded. It also has an inverse. A straightforward computation, using (6.14) and (6.16) now shows that if a vector ψ satisfies (6.17), then so also does the vector U (b, a)ψ, for any (b, a) ∈ G aff . A word now about the form of the representation U (b, a). How does one arrive at it? In fact, given the way the group acts on R, x → ax + b, the representation U (b, a) is

220

Matrix geometry of wavelet analysis. I

recognized as being the most natural, nontrivial way to realize a group homomorphism onto a set of unitary operators on the signal space L 2 (R, d x). (Unitarity is required in order to ensure that the signal ψ and the transformed signal U (b, a)ψ both have the same total energy.) Indeed, given any differentiable mapping T : Rn → Rn , the operator U (T ), on the Hilbert space L 2 (Rn , d n x), defined as x )), (U (T ) f )( x ) = |det[J (T )]|− 2 f (T −1 ( 1

where J (T ) is the Jacobian of the map T , is easily seen to be unitary. (Recall that d(T ( x )) = |det[J (T )]| d x .) This provides the rationale for defining the representation U (b, a) by (6.8). Of course, the interesting point here is that this representation turns out to be both irreducible and square integrable. But then, why is square integrability of the representation a desirable criterion for wavelet analysis? In order to answer this question, let us take a vector ψ satisfying the admissibility condition (6.17) and use it to construct the wavelet transform of the signal s: S(b, a) = ψb,a | s. As we already know, the total energy of the transformed signal is given by the integral dµ(b, a) | S(b, a) | 2 , (6.18) E(S) = G aff

and we would like this to be finite, like that of the signal itself. An easy computation now shows that E(S) = Cψ2 s2 = cψ s2 ,

(6.19)

which means that the total energy of the wavelet transform will be finite if and only if the mother wavelet can be chosen from the domain of the operator C, i.e., if and only if it satisfies the square integrability condition (6.13). However, this is not the whole story, for let us rewrite the above equation in the expanded form E(S) = dµ(b, a) s | ψb,a ψb,a | s G aff ! = s | dµ(b, a) | ψb,a ψb,a | s G aff

= cψ s | I s, I being the identity operator on L 2 (R, d x) and where, for any nonzero vector φ ∈ L 2 (R, d x), the quantity | φφ | /φ2 denotes the one-dimensional projection operator along this vector. (We have also interchanged two integrations with the taking of a scalar product, without justifying it, but the manipulation can easily be justified using standard

221

6.1 Group theory and matrix geometry of wavelets

absolute convergence techniques.) Using the well-known polarization identity for scalar products, we infer from this the operator relation 1 dµ(b, a) | ψb,a ψb,a | = I, (6.20) cψ G aff also known as the resolution of the identity equation. It is immediately clear that (6.20) is completely equivalent to the square integrability of the representation U (b, a) as expressed in (6.13). The resolution of the identity also incorporates within it the possibility of reconstructing the signal s(x) from its wavelet transform S(b, a). To see this, let us act on the vector s ∈ L 2 (R, d x) with both sides of (6.20). We get 1 dµ(b, a) ψb,a ψb,a | s = I s = s, cψ G aff implying s(x) =

1 cψ

dµ(b, a) S(b, a)ψb,a (x),

almost everywhere,

(6.21)

G aff

which is the celebrated reconstruction formula we encountered before. Summarizing, we conclude that square integrability (which is a group property) is precisely the condition which ensures, in this case, the very desirable consequences of (i) the finiteness of the energy of the wavelet transform, and (ii) the validity of the reconstruction formula (two properties also shared by the Fourier transform of a signal). The resolution of the identity condition (6.20) has independent mathematical interest. First of all, it implies that any vector in L 2 (R, d x) which is orthogonal to all the wavelets ψb,a is necessarily the zero vector, i.e., the linear span of the wavelets is dense in the Hilbert space of signals. This fact, which could also have been inferred from the irreducibility of the representation U (b, a), is what enables us to use the wavelets as a basis set for expressing arbitrary signals. In fact we have here what is also known as an overcomplete basis. Second, this overcomplete basis is a continuously parametrized set, meaning that this is an example of a continuous basis and a continuous frame. There is a host of other useful mathematical properties of the wavelet transform and spaces of transforms, which emanate from square integrability. We proceed to examine a few.

6.1.2

The space of all wavelet transforms A finite energy wavelet transform S(b, a) is an element of the Hilbert space L 2 (H, dµ). Here we have written H = R × R∗ (R∗ = real line with the origin removed). Although, H and G aff are homeomorphic as topological spaces, we prefer to denote them by different symbols, for presently we shall identify H with a phase space of G aff , arising from its matrix geometry. Using (6.18) and (6.19) to compare the L 2 -norm of the

222

Matrix geometry of wavelet analysis. I

wavelet transform S (as an element in L 2 (H, dµ)) to the L 2 -norm of the signal s (as an element in L 2 (R, d x)), we get S2 = cψ s2 ,

(6.22)

which just means that, up to a constant, the wavelet transform preserves norms (i.e., energies). We define a map Wψ : L 2 (R, d x) → L 2 (H, dµ), by the relation $ %− 1 $ %− 1 (Wψ s)(a, b) = cψ 2 ψb,a | s L 2 (R,d x) = cψ 2 S(b, a).

(6.23)

This map is linear and, in view of (6.22), an isometry, so that its range, which is the set of all wavelet transforms corresponding to the mother wavelet ψ, is a closed subspace of L 2 (H, dµ). We denote the range by Hψ : $ % Hψ = Wψ L 2 (R, d x) ⊂ L 2 (H, dµ). (6.24) From the defining equation (6.23) we infer that Hψ consists of continuous functions over H and hence is a proper subspace of L 2 (H, dµ). It is worthwhile reiterating here the fact that the condition of Wψ being an isometry implies, not only that the wavelet transform (with respect to the mother wavelet ψ) of any signal s ∈ L 2 (R, d x) is an element in Hψ , but also that every element in Hψ is the wavelet transform of some signal s ∈ L 2 (R, d x).

6.1.2.1

An intrinsic characterization of the space of wavelet transforms Is there some convenient, intrinsic way to characterize the subspace Hψ ? To answer this question we appeal to the resolution of the identity and a bit of group theory. The final characterization will be spelled out in Theorem 6.1.1. Multiplying each side of equation (6.20) by itself we find, 1 dµ(b , a ) dµ(b, a) | ψb,a K ψ (b, a ; b , a ) ψb ,a | = I, (6.25) cψ H×H where we have written 1 ψb,a | ψb ,a . K ψ (b, a ; b , a ) = cψ

(6.26)

Acting on the signal vector s with both sides of (6.25) and using (6.23), we obtain ! 1

dµ(b, a) dµ(b , a ) K ψ (b, a ; b , a )S(b , a ) ψb,a (x) = s(x), cψ H H almost everywhere (the change in the order of integrations being easily justified by Fubini’s theorem). Comparing the above equation with the reconstruction formula (6.21) we obtain the interesting identity dµ(b , a ) K ψ (b, a ; b , a )S(b , a ) = S(b, a), (6.27) H

223

6.1 Group theory and matrix geometry of wavelets

for almost all (b, a) in H (with respect to the measure dµ). This then is the condition which characterizes wavelet transforms coming from the mother wavelet ψ. It is also known as the reproducing property of the integral kernel K ψ . As we know, the kernel K ψ : H × H → C is called a reproducing kernel. It has the easily verifiable properties: K ψ (b, a ; b, a) > 0, for all (b, a) ∈ H,

(6.28)

K ψ (b, a ; b , a ) = K ψ (b , a ; b, a), dµ(b

, a

) K ψ (b, a ; b

, a

) K ψ (b

, a

; b , a ) = K ψ (b, a ; b , a ),

(6.29) (6.30)

H×H

the last relation being again the reproducing property (6.27) in a different guise (and in fact following directly from it). The last two equations hold pointwise, for all (b, a), (b , a ) ∈ H. Next let us compute the wavelet transforms of the wavelets ψb,a themselves. Denoting the transforms by Sb,a we find Sb,a (b , a ) = ψb ,a | ψb,a = cψ K ψ (b , a ; b, a).

(6.31)

Since the vectors ψb,a , (b, a) ∈ H, are overcomplete in L 2 (R, d x), the wavelet transforms Sb,a must also be overcomplete in Hψ . One also has the easily verifiable resolution of the identity, $ %−2 cψ dµ(b, a) | Sb,a Sb,a | = Iψ , (6.32) H

where we have written Iψ for the identity operator on Hψ . Thus, any vector F ∈ L 2 (H, dµ) which lies in the orthogonal complement of Hψ must satisfy dµ(b , a ) K ψ (b, a ; b , a )F(b , a ) = 0. H

All this goes to say that the reproducing kernel K ψ defines the projection operator Pψ from L 2 (H, dµ) to Hψ : dµ(b , a ) K ψ (b, a ; b , a )F(b , a ), F ∈ L 2 (H, dµ), (6.33) (Pψ F)(b, a) = H

equations (6.29) and (6.30) mirroring the conditions Pψ = P∗ψ = P2ψ (star denotes the adjoint). Stated differently, an arbitrary vector F ∈ L 2 (H, dµ) can be uniquely written as the sum F = Fψ + Fψ⊥ , of a part Fψ ∈ Hψ and a part Fψ⊥ orthogonal to it. The operator Pψ , acting on F, projects out the part Fψ (which is a wavelet transform). It is natural to ask at this point if Fψ⊥ could also be written as the wavelet transform with respect to some other mother

224

Matrix geometry of wavelet analysis. I

wavelet. As will be seen below, generally Fψ⊥ can be written as an infinite sum of orthogonal wavelet transforms, corresponding to different mother wavelets. To proceed further, we go back to the affine group, G aff , and note that there is a natural unitary representation of it on the Hilbert space L 2 (H, dµ), given by its natural action on H. This representation U (b, a), called the left regular representation, acts in the manner (U (b, a)F)(b , a ) = F((b, a)−1 (b , a )) b − b a , , =F a a

F ∈ L 2 (H, dµ).

(6.34)

The unitarity of this representation, U (b, a)F2L 2 (H,dµ) = F2L 2 (H,dµ) , is guaranteed by the invariance of the measure dµ (see (6.12)). However, the left regular representation is by no means irreducible, since as we shall see below, the subspace Hψ carries a subrepresentation of it. The isometry Wψ (see (6.23)) maps the unitary operators U (b, a) onto unitary operators Uψ (b, a) = Wψ U (b, a)Wψ−1 on L 2 (H, dµ). Computing the action of these operators, using (6.8) and (6.23), we find b − b a

(Uψ (b, a)F)(b , a ) = F (6.35) , , F ∈ Hψ. a a This is the same action as that of the operators U (b, a) of the left regular representation, except that now it is expressed exclusively in terms of vectors in Hψ . This means, first of all, that the subspace Hψ is stable under the action of the operators U (b, a) and, secondly, that restricted to this subspace, it gives an irreducible unitary representation of G aff .

6.1.2.2

Decomposition of the space of all finite energy wavelet transforms Let ψ and ψ be two different mother wavelets. This means that they are both vectors in the domain of the operator C (with their Fourier transforms satisfying (6.15)). How are wavelet transforms, taken with respect to these mother wavelets, related? In particular, denoting by Sψ the wavelet transform of the signal s, taken with respect to the mother wavelet ψ, and by Sψ the wavelet transform of the signal s , taken with respect to the mother wavelet ψ , we would like to evaluate the overlap

dµ(b, a) Sψ (b, a) Sψ (b, a) I (ψ, ψ ; s , s) = H

dµ(b, a) s | ψb,a ψb,a | s. (6.36) = H

225

6.1 Group theory and matrix geometry of wavelets

Let us begin by assuming that s, s are taken from a class of smooth functions (e.g., the Schwartz class, S(R)), which is dense in L 2 (R, d x). Then db da b,a | s | ψ b,a ψ s I (ψ, ψ ; s , s) = ∗ a2 R R ! db da 1

(ξ ) ψ (aξ ) 2 e ibξ = dξ | a | s a2 R R∗ R ! 1

−ibξ

2 s(ξ ) ψ(aξ ) . dξ | a | e × R

We exploit the smoothness of the functions s, s to use the identity 1

db eib(ξ −ξ ) = δ(ξ − ξ ), 2π R which holds in the sense of distributions, and then perform the ξ -integration to obtain da (aξ ) )ψ s(ξ ) I (ψ, ψ ; s , s) = 2π dξ ψ(aξ s (ξ ). R R a Changing variables, we get " # ! (y) ψ ψ(y)

dy dξ s(ξ ) s (ξ ) · I (ψ, ψ ; s , s) = 2π |y| R R ψ ψ |C s | s. = C Thus,

dµ(b, a) Sψ (b, a)Sψ (b, a) = H

H

dµ(b, a) s | ψb,a ψb,a | s

= Cψ | Cψ s | s.

(6.37)

Using the continuity of the scalar product s | s in s and s , we may now extend the above expression to all signals s, s ∈ L 2 (R, d x). Equation (6.37) is a general orthogonality relation for wavelet transforms. In particular, if Cψ and Cψ are orthogonal vectors, then the corresponding wavelet transforms are also orthogonal in L 2 (H, dµ). We may also write this equation in the form of an operator identity on L 2 (R, d x):

dµ(b, a) | ψb,a ψb,a | = Cψ | Cψ I, (6.38) H

which clearly is a generalization of the resolution of the identity (6.20). At the risk of being pedantic, we would still like to emphasize that the above orthogonality relation implies: r If s and s are signals which are orthogonal vectors in L 2 (R, d x), then their wavelet transforms S and S , whether with respect to the same or different mother wavelets, are orthogonal as vectors in L 2 (H, dµ).

226

Matrix geometry of wavelet analysis. I r

Spaces of wavelet transforms, Hψ , Hψ , corresponding to mother wavelets ψ, ψ which satisfy the orthogonality condition Cψ ⊥ Cψ , are orthogonal subspaces of L 2 (H, dµ). Equation (6.38) is a remarkable result. Acting on a signal s ∈ L 2 (R, d x) with both sides of this equation, and assuming that Cψ | Cψ = 0, we get 1

dµ(b, a) Sψ (b, a)ψb,a , ψb,a = U (b, a)ψ , s= Cψ | Cψ H where Sψ (b, a) = ψb,a | s is the wavelet transform of s computed with respect to the mother wavelet ψ. Thus, although the wavelet transform is computed with respect to the mother wavelet ψ, it can be reconstructed using the wavelets of any other mother wavelet ψ , so long as Cψ | Cψ = 0. Moreover, up to a multiplicative constant, the reconstruction formula is exactly the same as that in which the same wavelet ψ is used both for analyzing and reconstructing (see (6.21)). This indicates, that in some sense, analysis and reconstruction are independent of the mother wavelet chosen. Let us choose a set of mother wavelets {ψn }∞ n=1 such that the vectors φn = Cψn form an orthonormal basis of L 2 (R, d x), φn | φm = Cψn | Cψm = δnm ,

n, m = 0, 1, 2, . . . , ∞.

(6.39)

Such a basis is easy to find and we shall construct one below. If Hψn , n = 0, 1, 2, . . . , are the corresponding spaces of wavelet transforms and K ψn the associated reproducing kernels, then Hψn ⊥ Hψm , for n = m, and dµ(b

, a

) K ψn (b, a ; b

, a

) K ψm (b

, a

; b , a ) = δnm K ψm (b, a ; b , a ). H

(6.40)

More interestingly, it is possible to show that the complete decomposition, L 2 (H, dµ)

∞

Hψn ,

(6.41)

n=1

of the space of all finite energy signals (on the parameter space H) into an orthogonal direct sum of spaces of wavelet transforms, holds. Thus, in an L 2 -sense, any element S ∈ L 2 (H, dµ) has the orthogonal decomposition, S(b, a) =

∞

Sn (b, a),

almost everywhere,

n=1

into orthogonal wavelet transforms Sn , with respect to a basis of mother wavelets. This result, which can be proved by direct computation in the present case, is actually a particular example of a much more general result on the decomposition of the left regular representation of a group into irreducibles. For a more detailed mathematical discussion of this point, we refer the reader to [Ali00]. The components Sn (b, a) have the form

227

6.1 Group theory and matrix geometry of wavelets

Sn (b, a) = U (b, a)ψn | sn ,

n = 0, 1, 2, . . .

(6.42)

for some signal vectors sn ∈ L 2 (R, d x), which, in general, are different for different n. We also have the relations dµ(b , a ) K ψn (b, a ; b , a )Sm (b , a ) = δnm Sm (b, a). (6.43) H

Finally, we construct an explicit example of a basis set of mother wavelets satisfying (6.39). Let Hn (ξ ), n = 0, 1, 2, . . . , ∞, be the Hermite polynomials, normalized in the manner 0, if m = n, −ξ 2 (6.44) dξ e Hm (ξ ) Hn (ξ ) = √ 2n n! π , if m = n. R The first few are: H0 (ξ ) = 1,

H1 (ξ ) = 2ξ,

H2 (ξ ) = 4ξ − 2,

H3 (ξ ) = 8ξ 3 − 12ξ,

2

etc.

In the Fourier domain, define n (ξ ) = ψ

1 3 4

π 2

n+1 2

ξ2

√ | ξ | 2 e− 2 Hn (ξ ). n! 1

(6.45)

Then, it is easily verified that n 2 2 ψ < ∞, L (R,dξ ) and ψ ψ n = 2π m | C C

R

dξ n (ξ ) = δmn . ψm (ξ ) ψ |ξ |

Thus, in the inverse Fourier domain, the vectors ψn are in the domain of the operator C, satisfying the condition for being mother wavelets, while from the well-known properties of Hermite polynomials, the vectors φn constitute an orthonormal basis of L 2 (R, d x). More generally, since the range of C is dense in L 2 (R, d x), we can take any 2 orthonormal basis, {φn }∞ n=1 of L (R, d x), chosen from vectors in this range and then ∞ {ψn = C −1 φn }n=1 will be the desired wavelet basis. We collect the above results into a theorem: Theorem 6.1.1 The wavelet transform of the space of signals L 2 (R, d x), with respect to the mother wavelet ψ, is a closed subspace of L 2 (H, dµ). This subspace has a reproducing kernel K ψ , which is the integral kernel of the projection operator, Pψ : L 2 (H, dµ) → Hψ . The Hilbert space L 2 (H, dµ) can be completely decomposed into an orthogonal direct sum of an infinite number of subspaces Hψn , each a space of wavelet transforms with respect to a mother wavelet ψn . The vectors ψn are constructed by

228

Matrix geometry of wavelet analysis. I 2 taking an orthonormal basis {φn }∞ n=1 of L (R, d x), chosen from the range of the Duflo– −1 Moore operator C, and writing ψn = C φn .

6.1.2.3

Decomposition into orthogonal channels The above theorem can be used to analyze a given wavelet transform into orthogonal channels along the lines of [105]. Referring back to (6.39), since the mother wavelets {ψn }∞ n=1 form a complete, linearly independent set, any mother wavelet ψ can be written as a linear combination, ψ=

∞

an ψn ,

an = φn | Cψ = ψn | C 2 ψ.

(6.46)

n=1

Hence if s ∈ L 2 (R, d x) is any signal vector and Sψ (b, a) its wavelet transform with respect to ψ, then clearly Sψ (b, a) =

∞

Sn (b, a) ,

where

Sn (b, a) = an U (b, a)ψn | s.

(6.47)

n=1

In this way, the wavelet transform of the signal Sψ (b, a) has been decomposed into a set of mutually orthogonal wavelet transforms Sn (b, a) (of this same signal). We call this a decomposition into orthogonal channels. Note that although the wavelet transforms are orthogonal in L 2 (H, dµ), the mother wavelets ψn are not orthogonal in L 2 (R, d x) (see (6.39)). We shall see in the next chapter (at the end of Section 7.2.3), that in the case of a 2-D wavelet transform, it will actually be possible to obtain a decomposition into orthogonal (angular) channels using a family of mother wavelets which are themselves mutually orthogonal.

6.1.3

Localization operators Let ψ be a mother wavelet and Hψ the corresponding space of wavelet transforms. Let

⊂ H be a measurable set (with respect to the measure dµ). We associate to this set an integral kernel aψ : H × H → C:

dµ(b

, a

) K ψ (b, a ; b

, a

)K ψ (b

, a

; b , a ), (6.48) aψ (b, a; b , a ) =

and an operator aψ ( ) on Hψ acting via this kernel: dµ(b , a ) aψ (b, a ; b , a )S(b , a ), (aψ ( )S)(b, a) = H

S ∈ Hψ .

(6.49)

If we compute the matrix element of this operator for the wavelet transform S, using the properties of the reproducing kernel (see (6.29)–(6.30)), we easily obtain, dµ(b, a) | S(b, a) | 2 . (6.50) S | aψ ( )S =

229

6.1 Group theory and matrix geometry of wavelets

which shows that this operator is bounded, positive and self-adjoint. We call aψ ( ) a localization operator, since in view of the above relation, the quantity p S ( ) =

S | aψ ( )S S2

(6.51)

is the fraction of the wavelet transform which is localized in the region of phase space. If S(b, a) = ψb,a | s, then we shall also write ps ( ) for the above probability, for indeed, it measures the concentration of the phase space content (e.g., time–frequency content) of the signal s in the region . As a set function, p S ( ) has the properties of a probability measure: p S (∅) = 0, ∅ = empty set, p S (H) = 1, 7 8 p S ( i ) = p S ( i ), if i

j = ∅, whenever i = j, i∈J

(6.52)

i∈J

J being some index set. Since this holds for all S ∈ Hψ , we say that the operators aψ ( ) themselves constitute a positive operator-valued measure, or POV-measure, satisfying the properties: aψ (H) = Iψ , aψ (∅) = 0, 7 8 aψ ( i ), if i

j = ∅, whenever i = j, aψ ( i ) = i∈J

(6.53)

i∈J

where the sum on the right-hand side of (6.53) has to be understood in the weak sense, i.e., in the sense of (6.52). It is instructive to see how p S ( ) changes if the set gets transformed under the action of the group G aff . Since, for any (b0 , a0 ) ∈ G aff , U (b0 , a0 )ψb,a | ψb ,a L 2 (R,d x) = ψb,a | U (b0 , a0 )∗ ψb ,a L 2 (R,d x) ,

(6.54)

using the group properties (6.9) and the definition of the reproducing kernel in (6.26), we find that it satisfies the following covariance property: b − b0 a K ψ (b0 + a0 b, a0 a ; b , a ) = K ψ (b, a ; , ), (6.55) a0 a0 i.e., K ψ ((b0 , a0 )(b, a) ; b , a ) = K ψ (b, a ; (b0 , a0 )−1 (b , a )).

(6.56)

Let (b0 , a0 ) denote the translate of the set by (b0 , a0 ): (b0 , a0 ) = {(b0 + a0 b, a0 a) ∈ H | (b, a) ∈ }. Then, taking note of the action (6.35) of the left regular representation of G aff on wavelet transforms and exploiting the invariance of the measure dµ, we easily find that Uψ (b, a)S | aψ ( )Uψ (b, a)S = S | aψ ((b, a)−1 )S , i.e., we have the operator identity

S, S ∈ Hψ ,

230

Matrix geometry of wavelet analysis. I

Uψ (b, a)aψ ( )Uψ (b, a)∗ = aψ ((b, a) ).

(6.57)

This is a group covariance condition satisfied by the localization operators aψ ( ), and is generally known as an imprimitivity relation (see, e.g., [Ali00], for a more detailed discussion). For the probability measure p S ( ), this condition implies the transformation property p S ((b, a) ) = pUψ (b,a)−1 S ( ),

or,

ps ((b, a) ) = pU (b,a)−1 s ( ).

(6.58)

Physically, this relation means that the fraction of the signal s, localized in the transformed set (b, a) , is the same as the fraction of the transformed signal U (b, a)−1 s localized in the original set .

6.2

Phase space analysis We turn our attention to a different way of understanding the wavelet transform, namely, as a function on a phase space (in a sense to be made clear in the sequel). First let us (b, a) of the full affine group (see (6.14)) to the connected restrict the representation U + affine group G aff characterized by a > 0. We immediately see that this representation is no longer irreducible for this smaller group. Indeed, consider the two subspaces + (R ) = { , dξ ) | H f ∈ L 2 (R f (ξ ) = 0, ∀ ξ < 0}, 2 H − (R ) = { f ∈ L (R, dξ ) | f (ξ ) = 0, ∀ ξ > 0},

(6.59)

, dξ ) of the representation U (b, a). From (6.14) it is evident that of the carrier space L 2 (R vectors in any one of these subspaces are mapped to vectors in the same subspace under (b, a), when we only consider elements (b, a) ∈ G + the action of the operators U aff . This means that each one of these subspaces carries a unitary representation of this smaller group and, as before, we can show that both these representations, which we denote by + (b, a) and U − (b, a), respectively, are irreducible but unitarily inequivalent. In fact, U these are the only two nontrivial, unitary irreducible representations of G + aff . Moreover, , dξ ) is the orthogonal direct sum of these two subspaces: the Hilbert space L 2 (R + (R − (R , dξ ) = H ) ⊕ H ). L 2 (R

(6.60)

(b, a) In other words, we have here a complete decomposition of the representation U + of the connected affine group G aff and we write − (b, a). (b, a) = U + (b, a) ⊕ U U

(6.61)

In the inverse Fourier domain, the representation U (b, a) similarly breaks up (because of the unitarity of the Fourier transform) into two irreducible representations, U+ (b, a) and U− (b, a), on the two subspaces

231

6.2 Phase space analysis

H+ (R) = { f ∈ L 2 (R, d x) | f (ξ ) = 0, ∀ ξ < 0}, 2 H− (R) = { f ∈ L (R, d x) | f (ξ ) = 0, ∀ ξ > 0},

(6.62)

of L 2 (R, d x), respectively. These spaces are known as Hardy spaces [55,207,208]. Elements of H+ (R) (respectively, H− (R)) extend to functions analytic in the upper (respectively, lower) complex half-plane, and accordingly they are called upper (respectively, lower) analytic signals [Lyn82,Pap77].

6.2.1

Holomorphic wavelet transforms It is an interesting fact that, for appropriate choices of a mother wavelets, the wavelet transforms of signals in the spaces H± (R) can (up to a factor) become holomorphic functions. We study this property in some detail in this section. For any ν 0, consider + (R ), ∈H the mother wavelet ψ 12 ν+1 2ν ξ 2 e−ξ , for 0 < ξ < ∞, π (ν+1) )= ψ(ξ 0, otherwise, ψ 2 = 1, C

(6.63)

(ν + 1) being the usual Gamma function. The wavelets for this vector have the form b,a (ξ ) = ψ

2ν π (ν + 1)

! 12

ν

a 1+ 2 ξ

ν+1 2

e−iξ z ,

where z = b + ia,

(6.64)

and the reproducing kernel is ν

b,a | ψ b ,a = K ψ (b, a ; b , a ) = ψ

2ν (ν + 1) (aa )1+ 2 . i 2+ν π (z − z)2+ν

(6.65)

Computing the wavelet transform of a signal s ∈ H+ (R), b,a | s= S(b, a) = ψ

2ν π (ν + 1)

! 12 a

1+ ν2

∞

dξ ξ

1+ν 2

s(ξ ). eiξ z

(6.66)

0

ν

This, apart from the factor a 1+ 2 , is a holomorphic function of z = b + ia. Indeed, writing this function as !− 12 2ν ν F(z) = a −(1+ 2 ) S(b, a) π (ν + 1) ∞ 1+ν s(ξ ), z = b + ia, dξ eiξ z ξ 2 = 0

we have,

(6.67)

232

Matrix geometry of wavelet analysis. I

+ + | F(z) | = ++ 2

∞

1+ν 2

+2 + + + s(ξ ) ++ = ++

dξ ξ e ∞ s2 dξ ξ 1+ν e−2ξ a , iξ z

0

∞

dξ ξ

0

1+ν 2

e

−ξ a iξ b

e

+2 + s(ξ ) ++

by the Cauchy–Schwarz inequality

0

=

s2 (ν + 2) . (2a)2+ν

Thus, the convergence of the integral representing F(z) is uniform in any bounded open set containing z and differentiation with respect to it, under the integral sign, is permissible, implying that F is holomorphic in z, on the complex, upper half plane which we identify with H+ = R × R∗+ (where, R∗+ = (0, ∞)). We shall call F(z) a holomorphic wavelet transform. Furthermore, since db da 2 | S(b, a) | = dµν (z, z) | F(z) | 2 , a2 H+ H+ where we have introduced the measure, dµν (z, z) =

(2a)ν db da, π (ν + 1)

(6.68)

the set of all holomorphic wavelet transforms constitutes a closed subspace of the Hilbert space L 2 (H+ dµν ) of functions supported on the upper half plane. We denote this subspace by Hνhol and note that it is also a reproducing kernel Hilbert space, with reproducing kernel ν K hol (z, z ) =

(ν + 2) 1 . 2+ν i (z − z )2+ν

(6.69)

One has indeed ν dµν (z, z) K hol (z, z ) F(z ) H+

2ν (ν + 1) = i 2+ν = F(z).

H+

a ν da db

F(z ) (z − z )2+ν

The vectors ηz , z ∈ H+ , with ν ηz (z ) = K hol (z , z),

(6.70)

which are the holomorphic wavelets, are again overcomplete in Hνhol and satisfy the resolution of the identity: ν dµν (z, z) | ηz ηz | = Ihol ( = identity operator of Hνhol ). (6.71) H+

233

6.2 Phase space analysis ν There is also the holomorphic representation of G + aff on Hhol , unitarily equivalent to ν U+ (b, a). Denoting this by Uhol (b, a), we easily compute its action: z−b ν ν (b, a)F)(z) = a −(1+ 2 ) F (Uhol . (6.72) a

The appearance of the holomorphic Hilbert spaces of wavelet transforms is remarkable in many ways. First of all, their existence is related to a geometrical property of the half plane H+ , which is a differential manifold with a complex K¨ahler structure. This means, from a physical point of view, that it has all the properties of being a phase space of a classical mechanical system and, additionally, that this phase space can be given a complex structure (consistent with its geometry). In particular, it has a metric and a preferred differential two-form, which gives rise to the invariant measure dµ and using which classical mechanical quantities, such as Poisson brackets, may be defined. We will not go into the details of this here, but only point out the existence of a potential function in this context. Consider the function, (z, z ) = − log[−(z − z )2 ].

(6.73)

This function is called a K¨ahler potential for the space H+ and it generates all the interesting quantities characterizing its geometry, such as the invariant two-form and the invariant measure. Indeed, we immediately verify that

ν

e(1+ 2 ) (z,z ) =

i 2−ν K ν (z, z ). (ν + 2) hol

(6.74)

Next we define (z, z) =

1 ∂ 2 (z, z) db ∧ da , dz ∧ dz = i ∂z ∂z a2

(6.75)

which is the invariant two-form (under the action of G + aff ). This gives the invariant measure dµ of the group and furthermore, ν

e−(1+ 2 ) (z,z) = 4(−1)ν (2a)ν db ∧ da,

(6.76)

from which follows the measure with respect to which the holomorphic functions F(z) are square integrable and form a Hilbert space. (Recall that if u and v are two vectors in a vector space V , then u ∧ v is the antisymmetric tensor product, u ∧ v = u ⊗ v − v ⊗ u. The differentials, dz, da, etc., are considered as being elements in the dual of the tangent space of the manifold – in this case G + aff – at each point.) It ought to be emphasized here that Hνhol contains all holomorphic functions in L 2 (H+ , dµν ). Note also that, in view of (6.67), any such function can be obtained by computing the Fourier transform of 1+ν a function f (ξ ) = ξ 2 s(ξ ) and then analytically continuing it to the upper half plane, where s is a signal in the Fourier domain, with support in (0, ∞). Additionally, it ought to be noted that, for each ν > 0, we get a family of holomorphic wavelet transforms,

234

Matrix geometry of wavelet analysis. I

so that depending on the value of ν, the same signal s can be represented by different holomorphic functions on phase space.

6.2.2

Matrix analysis of phase space We have said earlier that the variables (b, a) parametrizing the space H+ , and in terms of which the wavelet transform is written, should be identified as phase space variables. In this section, we proceed to elaborate on this. To begin, let us determine the Lie algebra of the group G + aff . This group has two subgroups, formed by matrices of the type t e 0 1 t t (e , 0) = and (1, t) = , t ∈ R, 0 1 0 1 and a general element of the group can be obtained by multiplying two such matrices. Consider now the following two matrices X 1 , X 2 , which generate the Lie algebra gaff of this group: 1 0 0 1 + d t ++ d , X2 = , (6.77) (e , 0) t=0 = (1, t) + t=0 = X1 = dt dt 0 0 0 0 and satisfy the commutation relation, [X 1 , X 2 ] = X 1 X 2 − X 2 X 1 = X 2 .

(6.78)

Exponentiating these matrices we get, a 0 1 b (log a)X 1 bX 2 = , e = . e 0 1 0 1

(6.79)

The Lie algebra of the group is the two-dimensional vector space spanned by X 1 and X 2 and equipped with the commutation relation (6.78). A general element in the Lie algebra can be written as, 1 x2 x 1 2 X = x X1 + x X2 = (6.80) , x 1 , x 2 ∈ R. 0 0 Any group element can be obtained by exponentiating a suitable element of this Lie algebra. This is made clear if we write x1 x2 x1 a b (e − 1) e x1 X = . (6.81) (b, a) = e = 0 1 0 1 and note the inverse map from the group to the algebra: X = log(b, a) = x 1 X 1 + x 2 X 2 ,

x 1 = log a,

x2 =

b log a . a−1

(6.82)

235

6.2 Phase space analysis

Since every X ∈ gaff is mapped to an element (b, a) ∈ G + aff by the exponential map (6.81), we identify the domain of this map with the full real plane and use x = (x 1 , x 2 ) ∈ R2 as the coordinates for the elements of the Lie algebra. A group has a natural action on its Lie algebra, called the adjoint action. For (b, a) ∈ + G aff this action, which we denote by Ad(b,a) , is defined by 1 1 2 x −bx + ax . (6.83) Ad(b,a) X = (b, a)X (b, a)−1 = 0 0 The matrix of thistransformation, computed in the basis {X 1 , X 2 }, and acting on the x1 ∈ R2 is vectors x = x2 M(b, a) =

1 −b

0 . a

(6.84)

As a vector space, the Lie algebra has a dual space, which we denote by g∗aff . On it the * adjoint action of G + aff induces, by duality, the coadjoint action, denoted Ad(b,a) . To compute this action we take the dual basis {X ∗ 1 , X ∗ 2 } in g∗aff and write ageneral element γ1 in it as X ∗ = γ1 X ∗ 1 + γ2 X ∗ 2 . We identify X ∗ with the vector γ = ∈ R2 . On γ2 such vectors, the coadjoint action of the group is given (by definition) by the transposed inverse of the matrix M(b, a). We write −1 % $ 1 ba T M * (b, a) = M(b, a)−1 = . (6.85) 0 a −1 The determinants of these transformations are related as *

det[Ad(b,a) ] = a = det[Ad(b,a) ]−1 . Explicitly, a point γ ∈ R2 transforms under the coadjoint action as, −1 γ + ba γ 1 2 γ → γ = M * (b, a)γ = , a −1 γ2

(6.86)

(6.87)

and for fixed γ0 ∈ R2 , its orbit under G + aff is the set: 2 Oγ∗0 = {γ = M * (b, a)γ0 | (b, a) ∈ G + aff } ⊂ R .

(6.88)

Such an orbit is called a coadjoint orbit of the group G + aff . Orbits of different points either coincide entirely or are disjoint. In this way, the entire dual space g∗aff becomes the union over disjoint coadjoint orbits. Indeed, we easily establish the following orbit structure:

236

Matrix geometry of wavelet analysis. I

(i)

(ii)

(iii) (iv)

0 The orbit of the vector , 1 γ1 ∗ (6.89) ∈ R2 | γ2 > 0} = R × R∗+ . O+ = {γ = γ2 0 The orbit of the vector with the same matrices, −1 γ1 ∗ (6.90) ∈ R2 | γ2 < 0} = R × R∗− . O− = {γ = γ2 α α , The orbits of vectors , one for each α ∈ R. Such orbits are singletons, 0 0 and we denote them by Oα∗ . It is obvious that ∗ ∗ ∪ O− ∪α∈R Oα∗ = R2 . O+

(6.91)

We note also that each of the two orbits (i) and (ii) is homeomorphic to the group ∗ = H+ = R × R∗+ . As a manifold, this space itself. Consider the first one of these, O+ has a symplectic structure, i.e., there exists a preferred nondegenerate, antisymmetric, closed differential two-form on it, which is invariant under the coadjoint action. Indeed, it is trivially verified that the two-form (γ ) =

dγ1 ∧ dγ2 γ2

(6.92)

satisfies this condition. This also gives the invariant measure on this phase space, in ∗ these coordinates. We look upon O+ as an abstract differential manifold, with (γ1 , γ2 ) representing a particular choice of coordinates. In this context, we study two other possible choices of coordinates and see how the two-form (6.92) appears in these new coordinates. As the first of these coordinate transformations, we write q γ2−1 γ1 γ → η = = , (6.93) p γ2 which maps R × R∗+ onto itself. Under the coadjoint action of the group element (b, a), these coordinates transform as (q, p) → (q , p ) with, q = aq + b,

p = a −1 p.

The invariant two-form is simply ( η) = dq ∧ dp,

(6.94)

237

6.2 Phase space analysis

giving also the invariant Liouville measure in these coordinates. We call these coordi∗ nates the canonical or Darboux coordinates of the phase space O+ , in view of the form, familiar from classical Hamiltonian mechanics, assumed by the two-form . This also allows us to identify the variables q and p as position and momentum, respectively. We also immediately recognize the coadjoint action as inducing canonical transformations on phase space. As the second coordinate transformation, we choose −1 ξ γ γ 1 1 2 γ → ξ = = , (6.95) ξ2 γ2−1 which again maps R × R∗+ onto itself. Under the coadjoint action, these coordinates change in the manner: ξ1 = aξ1 + b,

ξ2 = aξ2 ,

(6.96)

with the invariant two-form now being (ξ ) =

dξ1 ∧ dξ2 . ξ22

(6.97)

The invariant measure arising from this should be compared to the left invariant measure, dµ, of the group G + aff (see (6.11)). Indeed, if we identify ξ1 , ξ2 with group parameters, then the transformation (6.96) is just the left multiplication in the group: (ξ1 , ξ2 ) = ∗ (b, a)(ξ1 , ξ2 ). It is because of the possibility of coordinatizing the orbit O+ in this particular way, that we may legitimately look upon the group elements themselves as phase space variables and the wavelet transform as a transform on phase space. ∗ Similar considerations apply to the other nontrivial orbit O− . We shall see in the next chapter that the two-dimensional wavelet transform can also be analyzed in a completely analogous manner. Before concluding this section, we should point out one crucial fact about the orbit structure of G + aff . The composition rule (6.6) equips this group with the structure of a semidirect product, R R∗+ , in which the subgroup R∗+ has an action on the subgroup R, in this case by simple multiplication. This action and α ∈ R∗ ), on the induces an action (again by multiplication, x → αx, for x ∈ R + dual space R (which we naturally identify with R). Under this action the dual space splits up into three orbits, the open half-spaces R∗+ , R∗− and the singleton {0}. The first two orbits are open and free, in the sense that they are both homeomorphic to the subgroup R∗+ itself and the map identifying the subgroup with the orbit is continuous . The fact and open. Furthermore, the union of the two open free orbits is dense in R ∗ that the orbits O± are open and free, is what leads to the representations U± (b, a) being square integrable (see, for example, [Ali00] for a detailed discussion of this point). ∗ Geometrically, the coadjoint orbits O± (of the whole group G + aff ) are the cotangent ∗ bundles of the orbits R± (of the subgroup R∗+ ).

238

Matrix geometry of wavelet analysis. I

6.3

The case of Gabor wavelets For the sake of completeness and comparison, we briefly look at the case of Gabor wavelets, or short-time Fourier transforms, for they too can be understood in group theoretical and phase space terms which are similar to those of the standard wavelets considered above. The one difference here is that the phase space of signals is not the entire group in question (the Weyl–Heisenberg group) and square integrability of the representation, giving rise to the Gabor wavelets, has also to be understood in the light of this fact. Another remarkable difference here is that for building Gabor type wavelets, any vector in the Hilbert space of signals can be used as a mother wavelet – the admissibility condition is trivial. The Gabor transform was introduced in (1.3). It was obtained by taking a window function ψ ∈ L 2 (R, d x), translating by b (in time) and modulating it in frequency 1/a, to obtain the Gabor wavelets or gaborettes. It had the form: ψb,a (x) = ei(x−b)/a ψ(x − b). The Gabor transform of a signal s is then defined in the same way as the wavelet transform: S(b, a) = ψb,a | s = d x e−i(x−b)/a ψ(x − b) φ(x). (6.98) R

It is a remarkable fact that the Gabor transform, and the 1-D continuous wavelet transform are built on exactly the same pattern (and this similarity persists in higher dimensions). Both are based on transformation groups of signals in L 2 (R, d x), the action of the group being implemented by unitary operators, s → U (b, a)s: r for the Gabor transform, the group is the Weyl–Heisenberg group G , acting as in WH (1.3); r for the continuous wavelet transform it is the ax + b or affine group, acting as in (1.4), and in fact the translation parts are identical in the two cases.

6.3.1

Group theoretical analysis In order to understand the Gabor transform group theoretically, let us first rewrite the gaborettes by adopting a slightly different notation and also introducing an additional phase variable. We write q for b and p for 1/a and denote the phase variable by θ : q

ψθ,q, p (x) = eiθ ei p(x− 2 ) ψ(x − q),

θ ∈ R.

The gaborettes would then correspond to setting θ = − pq/2. Writing

(6.99)

239

6.3 The case of Gabor wavelets

ψθ,q, p = U (θ, q, p)ψ,

(6.100)

an easy computation shows that U (θ, q, p) defines a unitary operator on the space of signals. Computing the effect of two such operators in succession, on a vector ψ yields the composition rule, U (θ1 , q1 , p1 ) U (θ2 , q2 , p2 ) = U (θ1 + θ2 + ξ ((q1 , p1 ); (q2 , p2 )), q1 + q2 , p1 + p2 ), (6.101) where, 1 (6.102) ( p1 q2 − p2 q1 ), 2 which in fact defines the multiplication rule between elements of the Weyl–Heisenberg group G WH . The function ξ is called a multiplier. The Weyl–Heisenberg group is a three-parameter group, homeomorphic to R3 . An arbitrary element g of G WH is of the form

ξ ((q1 , p1 ); (q2 , p2 )) =

g = (θ, q, p),

θ ∈ R,

(q, p) ∈ R2 ,

and the multiplication law in the group is g1 g2 = (θ1 + θ2 + ξ ((q1 , p1 ); (q2 , p2 )), q1 + q2 , p1 + p2 ).

(6.103)

The multiplier ξ equips the group with the structure of a central extension, of the group of translations of R2 , corresponding to phase space translations (i.e., translations q and modulations p). This just means that the subgroup consisting of the elements + = {g = (θ, 0, 0) | θ ∈ R}, is the center of the group G WH , i.e., these elements commute with every element g in the group, g(θ, 0, 0) = (θ, 0, 0)g, and it is the introduction of ξ which extends the commutative group R2 , with elements (q, p), into the noncommutative Weyl–Heisenberg group. The Weyl–Heisenberg group is unimodular, the measure dµ = dθ dq d p being invariant under both the left action, (θ, q, p) → (θ0 , q0 , p0 )(θ, q, p), and the right action, (θ, q, p) → (θ, q, p)(θ0 , q0 , p0 ). As with the affine group, the Weyl–Heisenberg group also has a matrix realization. It is given by the 4 × 4 matrices 1 12 ζ T ω θ 0 −1 (θ, q, p) = ω= , (6.104) ζ I2 0 , 1 0 0 0 T 1 with θ ∈ R,

ζ =

q p

∈R , 2

0 =

0 0

,

I2 =

1

0

0

1

.

Also, as for the affine group, we can compute the generators of the Lie algebra of G WH , by considering the one-parameter subgroups of elements of the type, g1 (t) = (t, 0, 0), g2 (t) = (0, t, 0) and g3 (t) = (0, 0, t), with t ranging through R. Computing

240

Matrix geometry of wavelet analysis. I

+ d gi (t) +t=0 , dt we obtain the three elements which generate the Lie algebra, gWH : T 1 T 1 T 0 − 2 e3 0 1 e 0 2 2 X0 = , X2 = , , X1 = O e1 O e2 O 0 Xi ≡

where 0 0 = 0 , 0

1 e 1 = 0 , 0

0 e 2 = 1 , 0

(6.105)

0 e3 = 0 , 1

and O is the 3 × 3 zero matrix. The matrices X i satisfy the commutation relations [X 0 , X 1 ] = [X 0 , X 2 ] = 0,

[X 1 , X 2 ] = X 0 .

(6.106)

A general element of gWH can be written as x 1 , x 2 , x 3 ∈ R;

X = x 0 X 0 + x 1 X 1 + x 2 X 2,

the commutation relations (6.106) then define a Lie bracket, [X, Y ] = X Y − Y X , between any two elements X, Y ∈ gWH . Next, noting that for any X ∈ gWH , (X )2 is the null matrix (which, in group theoretical terms, is stated by saying that the group G WH is nilpotent), we see that e X = (x 0 , x 1 , x 2 ) = I4 + X ∈ G WH .

(6.107)

Thus, the group and the Lie algebra can be given the same parametrization. In order to make the connection with gaborettes, we need to find unitary irreducible representations of G WH on the Hilbert space of signals, L 2 (R, d x). That is, we need to find a set of unitary operators, U (θ, q, p), for all (θ, q, p) ∈ G WH , which realize a group homomorphism, are stable under inverse taking and map the identity element (0, 0, 0) of the group to the identity operator on L 2 (R, d x). But this is already done by the operators defined in (6.100), so that U (θ, q, p) realize a unitary representation of the group. The fact that it is also irreducible can be proved in much the same way in which we proved irreducibility for the representation of the affine group in Section 6.1.1. More generally, it can be shown that any (nondegenerate) unitary irreducible representation of G WH is characterized by a real number λ = 0 and may be realized on the Hilbert space L 2 (R, d x) by the operators U λ (θ, q, p): q

(U λ (θ, q, p)s)(x) = eiλθ eiλp(x− 2 ) s(x − q),

φ ∈ L 2 (R, d x).

(6.108)

Two representations U λ (θ, q, p) and U λ (θ, q, p) are unitarily inequivalent if λ = λ . For the construction of gaborettes we shall mostly work with the representation for which λ = 1 and denote it simply by U instead of U 1 . (The other representations will be used in our discussion of holomorphic Gabor transforms below.)

241

6.3 The case of Gabor wavelets

A general element (θ, q, p) in G WH can be factorized as (θ, q, p) = (0, q, p)(θ, 0, 0). Thus, the quotient space G WH /+ (where the phase is factored out) is identifiable with R2 . It is this space which plays the rˆole of a phase space for the Weyl–Heisenberg group. We parametrize an element in G WH /+ by (q, p) ∈ R2 and since for a fixed element (θ0 , q0 , p0 ) ∈ G WH we again have the factorization 1 (θ0 , q0 , p0 )(0, q, p) = (00 , q + q0 , p + p0 )(θ0 + ( p0 q − pq0 ), 0, 0), 2 the action of the group on the quotient space is simply given by (q, p) → (q + q0 , p + p0 ). The invariant measure under this action is just the Lebesgue measure dq d p. For any ψ ∈ L 2 (R, d x), let us define the vectors qp ψq, p = U (− , q, p)ψ, (6.109) (q, p) ∈ G WH /+. 2 These are just gaborettes, expressed in the parameters q, p. The corresponding Gabor transform of a signal s is ∞ d x e−i p(x−q) ψ(x − q) s(x). (6.110) S(q, p) = ψq, p | s = −∞

Once again, computing the total energy of the transform, dq d p |S(q, p)|2 = 2π ψ2 s2 , E(S) =

(6.111)

R2

which should be compared to (6.19). From this there follows the resolution of the identity (exactly as in the derivation of (6.20)) 1 dq d p |ψq, p ψq, p | = I, (6.112) 2πψ2 R2 and the reconstruction formula (see (6.21)), 1 dq d p S(q, p)ψq, p (x), s(x) = 2πψ2 R2

for almost all x.

(6.113)

In the physical literature, gaborettes, and indeed also wavelets, are referred to as coherent states, of their respective groups [Kla85, Ali00, 6]. Note, however, that unlike the wavelets ψb,a , which were defined for all elements (b, a) of the affine group, the gaborettes ψq, p are only defined for points in the quotient space G WH /+. Moreover, the two (equivalent) conditions (6.111) and (6.112) imply that the representation U (θ, q, p) is square integrable with respect to this space: qp dq d p |U (− , q, p)ψ|ψ|2 < ∞, ∀ψ ∈ H. (6.114) 2 G WH /+ Again, one ought to emphasize here that the above admissibility condition is satisfied by all vectors ψ in the signal space, but that it is defined with respect to a quotient

242

Matrix geometry of wavelet analysis. I

space. Indeed, it also follows from the above that, for any vector ψ in the Hilbert space of signals, dq d p dθ |U (θ, q, p)ψ|ψ|2 = ∞, G WH

so that there is no vector which is admissible with respect to the entire group. We shall return to the subject of building wavelet transforms on general quotient spaces in the next chapter (see Section 7.1.5). To complete this cycle of properties, we could again verify that the space Hψ of all Gabor transforms, corresponding to a particular window function ψ, is a closed subspace of L 2 (G WH , dq d p), which is also a Hilbert space with a reproducing kernel: K ψ (q, p; q , p ) =

1 ψq, p | ψq , p . 2π ψ2

(6.115)

This relation is the exact analog of (6.26). The complete decomposition of the space L 2 (G WH /+, dq d p), of all Gabor transforms, could be worked out in the manner of (6.41), with very little change in the derivation.

6.3.1.1

Holomorphic Gabor wavelets It is possible to construct spaces of holomorphic Gabor transforms and associated holomorphic gaborettes, in much the same way as we constructed holomorphic wavelet transforms in Section 6.2.1. We show below the existence of one such space and later we shall indicate how others may be obtained. However, in order to do so, it is first necessary to modify somewhat the definition of the Gabor transform. We proceed by first choosing the window function x2

ψ(x) = (π )− 4 e− 2 , 1

ψ2 = 1.

(6.116)

Next, we define the modified gaborettes, q

ψ(0,q, p) (x) = (U (0, q, p)ψ)(x) = (π )− 4 ei(x− 2 ) p e− 1

(x−q)2 2

qp

= ei 2 ψq, p

(6.117)

and use them to define the modified Gabor transform of a signal s: S(q, p) = ψ(0,q, p) | s = e−i

qp 2

S(q, p).

(6.118)

Then S obviously satisfies the finiteness of energy condition and moreover, the modified gaborettes also satisfy the resolution of the identity. Thus, signal reconstruction is possible using these modified gaborettes as well. Introducing the complex variable z = q − i p, it is easy to verify that |z|2

e− 4 S(q, p) = 1 π4 Writing

∞

−∞

z2

d x e− 2 (x−z) e 4 s(x) ≡ S(z). 1

2

(6.119)

243

6.3 The case of Gabor wavelets |z|2

e− 4 F(z), S(z) = √ 2π

(6.120)

we see that F(z) is an entire analytic function of z. These functions constitute the holomorphic Hilbert space Hhol = L 2 (C, dν(z, z)), where dν is the measure |z|2 dq d p dν(z, z) = e− 2 . 2π This Hilbert space is also a reproducing kernel Hilbert space, with the kernel given by zz

K hol (z, z ) = e 2 .

(6.121)

We call the ψ(0,q, p) holomorphic gaborettes and the functions F(z) holomorphic Gabor transforms. Other spaces of holomorphic Gabor transforms can now be similarly constructed by replacing the window function (6.116) by a Gaussian with standard deviation λ−1 , λ > 0 and using the representation (6.108) of the Weyl–Heisenberg group. Consider the window function !1 λ2 4 − λ2 x 2 ψ(x) = e 2 , ψ2 = 1, λ > 0, (6.122) π and define the generalized gaborettes λ2 ψ(0,q, p) (x) = (U (0, q, p)ψ)(x) = π λ

! 14

q

eiλ(x− 2 ) p e−

λ2 (x−q)2 2

,

(6.123)

where U λ is the unitary representation of the Weyl–Heisenberg group defined in (6.108). It can then be verified that the resolution of the identity λ dq d p |ψ(0,q, p) ψ(0,q, p) | = I, (6.124) 2π R2 holds. The generalized Gabor transform is now S(q, p) = ψ(0,q, p) | s,

s ∈ L 2 (G WH /+, dq d p),

(6.125)

which, by virtue of (6.124), enjoys the finiteness of energy condition 2π s2 . S2 = λ Introducing the complex variable √ p z = λq −i√ , λ (6.123) may be rewritten as: λ2 ψ(0,q, p) (x) = ψ(0,z) (x) = π

! 14

e−

λ|z|2 4

e

λz 2 4

√ − λ2 ( λx−z)2

The generalized Gabor transform now becomes

.

(6.126)

244

Matrix geometry of wavelet analysis. I

λ2 S(q, p) ≡ S(z) = π

! 14 e

− λ|z| 4

2

e

λz 2 4

∞

λ

d x e− 2 (

√ λx−z)2

s(x),

(6.127)

−∞

so that 2π F(z) = λ

! 12 e

λ|z|2 4

2

λz S(z) = e 4

∞

λ

√

d x e− 2 (

λx−z)2

s(x)

−∞

is the analytic Gabor transform and it then follows that 2 dq d p | S(q, p)| = dνλ (z, z) |F(z)|2 , G WH /+

C

where we have introduced the measure λ − λ|z|2 e 2 dq d p. dνλ (z, z) = 2π Thus, F is a vector in a Hilbert space of entire analytic functions, which we denote by Hλhol and note that this space of holomorphic Gabor transforms is a subspace of L 2 (C, dνλ ). The Hilbert space Hλhol has the reproducing kernel λ (z, z ) = e K hol

λzz 2

,

(6.128)

so that, for any F ∈ Hλhol , λ dνλ (z, z) K hol (z, z )F(z ) = F(z). C

Finally, the complex plane C is also a K¨ahler manifold, with potential function zz , 2 and invariant two-form (z, z) =

1 ∂ 2 (z, z) dz ∧ dz = dq ∧ d p, i ∂z ∂z so that, once again, we verify the relations (see (6.74)–(6.76)) (z, z) =

λ (z, z ) = eλ (z,z ) , K hol

and e−λ (z,z) = e−

λ|z|2 2

dq ∧ dp,

from which follows the measure dνλ . Thus, group theoretically, the analysis of Gabor wavelets using the Weyl–Heisenberg group runs entirely parallel to the analysis of the 1-D CWT using the affine group, except for the following two differences. First, here every vector in L 2 (R, d x) is admissible (the group theoretical reason being that G WH is a unimodular group). Second, the square integrability of the representation U (θ, q, p) is not on the entire Weyl–Heisenberg

245

6.3 The case of Gabor wavelets

group G WH itself, but on the quotient G WH /+ of the group by its center +. However here, as with the affine group, it is the square integrability of the representation in question which leads to the finiteness of the energy of the Gabor transform and enables us to reconstruct the signal from its transform.

6.3.2

Phase space considerations It is instructive to carry out a phase space analysis of the Weyl–Heisenberg group, for it sheds light on both the differences and the similarities between it and the affine group. In particular, it will clearly emerge why the phase space over which the Gabor transform is built is two-dimensional, although the group itself is a three-dimensional manifold. Let (θ, q, p) ∈ G WH and X ∈ gWH . We compute the adjoint action of the group on the Lie algebra, using (6.107) to get

x 0 = x 0 + px 1 − q x 2 ,

x 1 = x 1,

x 2 = x 2, where

X = Ad(θ,q, p) X = x 0 X 0 + x 1 X 1 + x 2 X 2 = (θ, q, p) [x 0 X 0 + x 1 X 1 + x 2 X 2 ] (θ, q, p)−1 . Representing X by the vector in R3 with components x 0 , x 1 , x 2 , the adjoint action is given by the matrix

1

p

M(θ, q, p) = 0

1

0

0

−q

0 . 1

(6.129)

Note that there is no dependence of M on the phase θ . The adjoint action on the dual vectors γ is effected by the matrices M * (θ, q, p), which are the transposed inverses of the matrices M(θ, q, p). Writing γ = M * (θ, q, p)γ , we get the dual transformation equations γ 0 = γ0 , γ1 = − pγ0 + γ1 , γ2 = qγ0 + γ2 . Thus, the orbits of vectors under the coadjoint action fall into two categories:

246

Matrix geometry of wavelet analysis. I

λ 0 3 (i) The orbits of vectors of ∈ R , 0 = , one for each λ = 0: 0 0 Oλ∗

λ = {γ = | x ∈ R2 }, x

(6.130)

which are planes orthogonal to the γ0 -axis. 0 ∈ R2 . These orbits are singletons consisting of the (ii) The orbits of vectors , λ λ 0 vector itself. We denote them by Oλ∗ and note that together they form a set λ of Lebesgue measure zero in R3 . Clearly, the union of all the orbits is the entire dual space, g∗WH , of the Lie algebra, now identified with R3 . The nontrivial orbits Oλ∗ , λ = 0, can each be identified with the quotient space G WH /+, since under the coadjoint action the phase subgroup + (consisting of elements of the type (θ, 0, 0)) is stable. These are the (two-dimensional) phase spaces of the problem and from the general theory of group representations it is known that each unitary irreducible representation of G WH is associated to one such orbit. As stated earlier, these representations U λ (θ, q, p) can be realized as in (6.108). The invariant two form under the coadjoint action on the orbits Oλ∗ is just dγ1 ∧ dγ2 and for λ = 1, this action is simple: γ1 → γ1 − p , γ2 → γ2 + q. The preceding discussion makes clear the mathematical sense in which the space of parameters of the Gabor transform S(q, p) is to be thought of as a phase space. The space of parameters is a coadjoint orbit of the group, which has the structure of a classical mechanical phase space.

7

Matrix geometry of wavelet analysis. II

The last chapter has already familiarized us with the use of group theoretical methods for the construction and analysis of wavelets and gaborettes. We aim in this chapter to first indicate the general applicability of these techniques and then to look at the case of the two-dimensional continuous transform, using the SIM(2) group. Later, we look at general matrix groups of the type that can be used for constructing other types of wavelet transforms in two dimensions. We shall be led, in this manner, to studying a class of semidirect product type groups, certain coadjoint orbits of which are isomorphic to the group itself. In all these cases, the common features of such a matrix-group analysis will be: (a) the group will refer to a set of possible symmetry transformations which the signal may undergo; (b) the space over which the signals are defined (as L 2 -functions) is intrinsic to the group; (c) the parameters in terms of which the wavelet transform is expressed are the parameters of the group itself, i.e., symmetry parameters of the signal, and (d) these parameter spaces, which arise as coadjoint orbits of the group, are also identifiable with phase spaces of signals. Referring back to the 2-D wavelet transform introduced in Chapter 2, we shall see that this transform is again related to a square integrable representation of a matrix group. The coadjoint orbit of this group (there is only one nontrivial orbit in this case) will again allow us to carry out a phase space analysis. As in the 1-D case, square integrability will enable us to obtain a resolution of the identity, lead to finiteness of the total energy of the wavelet transform and yield a reconstruction formula for the signal. In order to put the discussion in the context of a more general framework, we begin with a word about the choice of an appropriate group for building wavelet transforms and the general rationale for appealing to group theory in the first place.

7.1

A group-adapted wavelet analysis As mentioned earlier, a group-theoretical approach enables one to exploit mathematically the symmetries underlying the particular geometry which the signal space may have. If symmetries exist, it is natural to try to build these into the wavelet transform itself. This generally implies finding a continuous wavelet transform by exploiting a

247

248

Matrix geometry of wavelet analysis. II

representation of the group on some Hilbert space, the key ingredient required for such a construction being square integrability of the representation. Assume that the class of finite energy signals under consideration can be realized as functions on a manifold Y , i.e., s ∈ L 2 (Y, dµ) ≡ H. This manifold could in fact be the space of some parameters (e.g., frequency, time, position, etc.) of the signal. As examples, Y could be space Rn , the 2-sphere S 2 , space–time R × R or R × R2 , and so on. Assume that we measure the signals with the help of a probe η : s → η[s]. Usually such probes are taken to be linear functionals, representing the action of a measuring apparatus, a reference frame, etc. Mathematically, the measurement of the signal is given usually by some sort of an overlap integral that is, in the present case, an inner product s → ψ | s, with ψ representing the probe. Note that, if we were to restrict the signals to smooth functions on Y , measurements could also be represented by distributions of some type.

7.1.1

Some generalities Suppose there is a group G of symmetries of the signal, which acts as a set of transformations of the manifold Y . This means that any element g ∈ G, representing a specific symmetry operation on the parameter space (e.g., rotation of the signal through some angle, translation by some amount, etc.), induces a transformation of Y : y → g[y]. Successive transformations of this kind can be composed: g[g [y]] = gg [y] and the composite transformation gg is again an element of G. To each transformation, y = g[y], there exists an inverse transformation, y = g −1 [y ], g −1 ∈ G and in addition the identity transformation e for which e[y] = y, for all y ∈ Y is also a member of the group G. We assume further that this action is transitive, i.e., for any pair y, y ∈ Y , there is at least one g ∈ G such that g[y] = y . It should be noticed, however, that the transformation group G acting on Y is in general not unique, its choice may depend on the problem at hand. In the case of the 1-D CWT, for instance, we could choose between the full affine group G aff (a = 0) and its connected subgroup G + aff (a > 0). However, when talking about wavelets, the groups that we consider will always contain a dilation as a symmetry of the signal. Furthermore, in all the cases that will concern us here, the group action will be realized through matrices acting on a vector space into which Y can be embedded. Recall, this means that the set of matrices constituting this group is closed under multiplication, inverse taking and that the identity matrix is also a member of this set. The action of the group on Y induces an action on the signal space and there are two possible ways in which this could happen. (1) Action on the signals themselves. This is the active point of view: s → sg , in which one measures the transformed signal, which we denote by sg , with the fixed probe η. (2) Action on the probes. This is the passive point of view: η → ηg , in which one measures the fixed signal s with the transformed probe ηg . Now, if the manifold Y is globally G-invariant, consistency requires that the two points of view be equivalent,

249

7.1 A group-adapted wavelet analysis

ηg [sg ] = η[s], ∀ g ∈ G.

(7.1)

If now signals are taken to be vectors s in the Hilbert space H ≡ L 2 (Y, dν), where dν is some convenient measure, one naturally identifies probes with linear functionals on H, that is, η[s] ≡ η | s, for some fixed vector η ∈ H. Next, if one imposes that the action of G be linear, one ends up with a linear representation of G in H, sg = U (g)s, ηg = U (g)η. This means that to each g ∈ G one associates an operator U (g) on H, and the mapping g → U (g) is a group homomorphism (see (6.9)). The consistency condition (7.1) now requires that the U (g) be unitary operators, i.e., one has a unitary representation of G on the space H of signals: U (g)s | U (g)s = s | s , ∀ g ∈ G, s, s ∈ H

⇒

U (g)−1 = U (g)∗ .

(7.2)

Being a representation of G, the operators U (g) must also satisfy U (g1 )U (g2 ) = U (g1 g2 ),

U (g −1 ) = U (g)−1 ,

U (e) = I (= identity operator of L 2 (Y, dν)).

(7.3)

for all g1 , g2 in G and where e denotes the unit element of G (i.e., ge = eg = g, ∀ g). As an additional requirement, the representation U (g) will be assumed to be irreducible. Technically this means that for any nonzero vector ψ ∈ H, the set of vectors U (g)ψ, g ∈ G, span the Hilbert space. (At times a weaker condition, e.g., the existence of just one such nonzero vector ψ may be sufficient.) Irreducibility also means that H is a minimal space for realizing the symmetries unitarily. As a measure space, the group G has generally two invariant measures defined on it, a left Haar measure, dµ, invariant under g → g0 g, for fixed g0 ∈ G and a right Haar measure, dµr , invariant under g → gg0 . dµ(g0 g) = dµ(g) ,

dµr (gg0 ) = dµr (g),

g ∈ G.

The existence of these measures is a general property of all topological groups; for the affine group these were explicitly written down as dµ = db db/a 2 and dµr = db da/a (see (6.11) and the discussion following). The two invariant measures on G are equivalent, but generally not equal. In particular, for all wavelet related groups, which include dilations, the two Haar measures are different. We shall usually work with the left Haar measure, although everything we do could just as well be done using the right Haar measure. The space L 2 (G, dµ) is usually taken to be the Hilbert space of all wavelet transforms and it will turn out that finding wavelet transforms of signals implies mapping the signal space L 2 (Y, dν) isometrically into a (closed) subspace of L 2 (G, dµ). Given this setting, one may derive a wavelet analysis on Y , adapted to the symmetry group G, in which the wavelet transform of a signal s would be an associated function S, defined on the group (the group being identified with a phase space for the signal) or some other phase space related to the group. In case the group itself can be identified with a phase space (as was the situation with the affine group or as it will be for the 2-D

250

Matrix geometry of wavelet analysis. II

wavelet transform discussed below), the obvious choice of the function S is to write it as S(g) = U (g)η | s. Two questions immediately arise: first, does the transform S(g) have finite energy and second, is it possible to reconstruct the signal s uniquely from its transform S?

7.1.2

Square integrability of representations The first question posed above can be reformulated as follows. If the total energy of the transform S is identified with the square of its L 2 -norm: 2 E(S) = S = dµ(g) |S(g)|2 , G

then of course, we want S < ∞. In fact we require that the mapping s → S be (up to a constant) an isometry. Thus, the original question about the finiteness of energy becomes: is it possible to find ψ ∈ H such that S is finite for all s or that the map s → S be a multiple of the isometry? It will turn out that a positive answer to this question will also guarantee the possibility of reconstructing the signal from its transform, i.e., a positive answer to the second question. Interestingly enough, from a mathematical point of view, in order to obtain finiteness of energy for all transforms S, it is enough to require this of the transform of the probe ψ only. Thus, we require the existence of a nonzero vector ψ ∈ H such that I (ψ) = dµ(g) |U (g)ψ | ψ|2 < ∞ . (7.4) G

In case such a vector exists, we call it an admissible vector or, in signal analytic language, a generalized mother wavelet. The representation U (g) is then said to be square integrable. Square integrability is a property of both the representation U (g) and of the group G itself. Not all groups have square integrable representations and the same group may have representations which are square integrable as well as other ones which are not. From the general theory of square integrable group representations (see, e.g., [Ali00] for a detailed account) one knows the following: r The existence of one admissible vector guarantees the existence of a dense set of such vectors. In particular, if ψ is admissible, then so also are all the vectors U (g)ψ, g ∈ G. Let us denote the set of all admissible vectors by A. Then there exists an operator C, in general unbounded, on the Hilbert space H such that it is self-adjoint, has positive spectrum and such that its domain is precisely the set of all admissible vectors: cψ ≡ Cψ2 < ∞

⇔

ψ ∈ A.

(7.5)

Moreover, C −1 also exists as a densely defined positive (spectrum) operator. As in the 1-D wavelet case (6.17), we call C the Duflo–Moore operator of the representation. r If the group is nonunimodular (i.e., the left and right Haar measures are different) then C is an unbounded operator and A is a proper subset of H. If G is unimodular,

251

7.1 A group-adapted wavelet analysis

then C = λI, λ > 0, and A coincides with the entire Hilbert space H (every vector is an admissible vector). r For any two admissible vectors ψ, ψ and arbitrary vectors s, s ∈ H, the following orthogonality relation holds: dµ(g) U (g)ψ | s U (g)ψ | s = Cψ | Cψ s | s. (7.6) G

Note that (6.37) is just a special case of this relation, when G is the affine group. Proceeding in the same way by which we arrived at (6.38) from (6.37), we derive from (7.6) the resolution of the identity, 1 dµ(g) |U (g)ψ U (g)ψ| = I, (7.7) Cψ | Cψ G provided Cψ | Cψ = 0. Taking ψ = ψ in the above gives, 1 dµ(g) |ψg ψg | = I. cψ G

7.1.3

(7.8)

Construction of generalized wavelet transforms Given the existence of a square integrable representation U (g) of the group G, on the signal space H = L 2 (Y, dν), (generalized) wavelets and wavelet transforms can be constructed by exploiting the orthogonality relation (7.6), in much the same way as was done for the affine group in the previous chapter. We start by taking a mother wavelet ψ ∈ H and defining generalized wavelets as the vectors ψg = U (g)ψ ∈ H, g ∈ G. In the physical literature, the vectors ψg are called coherent states of the representation U (g). Our convention will be to call these vectors generalized wavelets only when the group contains some sort of a dilation transformation on the space Y and the group space itself can be identified with a phase space. Otherwise we shall use the term coherent state. The question of when the group also has the structure of a phase space will be analyzed later. Using the generalized wavelets ψg , the (generalized) wavelet transform of the signal s ∈ H is defined to be the function S(g) = ψg | s on G. Since the group G is derived by analyzing symmetry transformations of signals, its elements g are defined in terms of these very symmetry parameters (e.g., rotation angle, translation distance, zoom factor, etc.) and hence the wavelet transform also becomes a function of these parameters. Computing now the expectation value of both sides of (7.8) with respect to s ∈ H yields 2 S L 2 (G,dµ) = dµ |S(g)|2 = cψ s2H, (7.9) G

implying that the map Wψ : H → L 2 (G, dµ), given by −1/2

(Wψ s)(g) = cψ

ψg | s,

(7.10)

252

Matrix geometry of wavelet analysis. II

is a linear isometry. Similarly, acting on a signal s with the operators appearing on both sides of (7.7), we derive the general reconstruction formula 1 s= dµ(g) S(g)ψg . (7.11) Cψ | Cψ G Again, it ought to be emphasized here that the wavelet transform S(g), appearing in this expression, is computed using the wavelets ψg , while the reconstruction is done using the different set ψg . Of course, the formula is valid also if ψ = ψ . The above discussion again illustrates how in the general situation, as in the case of the affine group discussed in the last chapter, it is the square integrability of the representation which leads to the finiteness of the energy of the transform, on the one hand, and to the reconstruction formula, on the other. The orthogonality relation (7.6) acquires a more transparent physical meaning if we express it in terms of wavelet transforms. If Sψ (g) is the wavelet transform of the signal s, with respect to the mother wavelet ψ and Sψ (g) the transform of s with respect to ψ , then (7.6) can be rewritten as, dµ(g) Sψ (g) Sψ (g) = Cψ | Cψ s | s. (7.12) G

In other words, wavelet transforms of orthogonal signals are always orthogonal, independent of the mother wavelets chosen to represent them, while wavelet transforms of arbitrary signals, when computed with respect to a mother wavelet ψ, are all orthogonal to their transforms computed with respect to ψ , if Cψ is orthogonal to Cψ .

7.1.4

Reproducing kernels, partial isometries and localization operators Let ψ be an admissible vector and let Wψ [L 2 (Y, dν)] ≡ Hψ be the range of the isometry Wψ . This means that Hψ is a closed Hilbert subspace of L 2 (G, dµ), consisting of all wavelet transforms S(g) associated to the mother wavelet ψ. Let Pψ be the projection operator onto Hψ , i.e., Pψ L 2 (G, dµ) = Hψ . If ψ and ψ are two admissible vectors, chosen so that Cψ | Cψ = 0, then from (7.6) we infer that the corresponding spaces of wavelet transforms, Hψ and Hψ , are orthogonal. However, unlike in the case of the affine group (see Theorem 6.1.1), this fact cannot in general be used to obtain a complete decomposition of L 2 (G, dµ) into orthogonal subspaces of wavelet transforms – generally the space L 2 (G, dµ) contains more than just wavelet transforms. An interesting feature, connecting the spaces of wavelet transforms Hψ corresponding to different mother wavelets ψ, now emerges. All these spaces of transforms sit inside L 2 (G, dµ) and the same signal s can be mapped into different transform spaces, by choosing different mother wavelets. It is natural to ask whether it is possible to interpolate between these spaces of transforms, for that would allow us to take the transform of a signal with respect to one mother wavelet and re-express it as a transform with respect to a different mother wavelet. The following discussion, in particular Theorem 7.1.1, will show exactly how this can be done.

253

7.1 A group-adapted wavelet analysis

7.1.4.1

Partial isometries For two arbitrary mother wavelets ψ, ψ , define the function K ψ,ψ : G × G → C,

K ψ,ψ (g, g ) = (cψ cψ )−1/2 ψg | ψg .

(7.13)

This function has the easily verifiable properties, K ψ,ψ (g, g ) = K ψ ,ψ (g , g) , dµ(g

) K ψ,ψ

(g, g

) K ψ

,ψ (g

, g ) = K ψ,ψ (g, g ),

(7.14)

G

the first following from the definition of K ψ,ψ in (7.13) and the second from the resolution of the identity in (7.8). Next let us define an integral operator Vψ,ψ , with K ψ,ψ (g, g ) as its kernel: dµ(g ) K ψ,ψ (g, g ) F(g ), F ∈ L 2 (G, dµ). (7.15) (Vψ,ψ F)(g) = G

We show below that K ψ ≡ K ψ,ψ is the integral kernel of the projection operator Pψ and hence defines a reproducing kernel of the type seen in the last chapter (see (6.28)– (6.30)), while the operator Vψ,ψ , for different ψ, ψ , is a partial isometry on L 2 (G, dµ). This means that the range of Vψ,ψ is the space Hψ , of wavelet transforms corresponding to the mother wavelet ψ , while its kernel is the orthogonal complement, H⊥ ψ , of the space of transforms Hψ . Between Hψ and Hψ the operator interpolates as a linear isometry. Theorem 7.1.1 . The operator Vψ,ψ is a partial isometry on L 2 (G, dµ), which maps the subspace Hψ isometrically onto Hψ , and has the properties Vψ,ψ Vψ ,ψ

= Vψ,ψ

,

∗ Vψ,ψ

= Vψ ,ψ ,

∗ Vψ,ψ Vψ,ψ

Vψ,ψ = Pψ ,

= Pψ ,

∗ Vψ,ψ

Vψ,ψ

(7.16) = Pψ .

(7.17)

Proof . The two relations in (7.16) follow from (7.14). For any F ∈ L 2 (G, dµ), it can be shown that dµ(g ) K ψ (g, g ) F(g ), (7.18) (Pψ F)(g) = G

in exactly the same way as (6.33) was proved for the affine group. This establishes the first of the relations in (7.17). Also,

−1/2 dµ(g ) K ψ,ψ (g, g ) F(g ) = (cψ cψ ) ψg | dµ(g ) ψg F(g ) G

where s = (cψ cψ )−1/2

= ψg | s, G

dµ(g) ψg F(g) ∈ L 2 (Y, dν).

G

254

Matrix geometry of wavelet analysis. II

Thus, Vψ,ψ F is the wavelet transform of s with respect to the mother wavelet ψ, implying that the range of Vψ,ψ is contained in the subspace Hψ ⊂ L 2 (G, dµ). To see that it actually coincides with Hψ , let F be an arbitrary element of Hψ . Then there exists a signal s ∈ L 2 (Y, dν) for which F(g) = ψg | s. But, by (7.8), 1 dµ(g ) ψg | ψg ψg | s ψg | s = cψ G = dµ(g ) K ψ,ψ (g, g )F (g ) = (Vψ,ψ F )(g), G

where we have written 1/2 cψ

ψg | s . F (g) = cψ Thus every vector in Hψ is also in the range of Vψ,ψ . On the other hand, using (7.14) and (7.18), dµ(g ) K ψ,ψ (g, g ) F(g ) G dµ(g ) dµ(g

) K ψ,ψ (g, g

) K ψ (g

, g ) F(g ) = G G dµ(g

) K ψ,ψ (g, g

) (Pψ F)(g

), = G

implying that, for any vector F ∈ H⊥ ψ , Vψ,ψ F = 0.

Thus H⊥ ψ is contained in the kernel of Vψ,ψ . Finally, if F ∈ Hψ , so that F (g) = ψg | s, for some signal vector s ∈ L 2 (Y, dν), and F 2L 2 (G,dµ) = cψ s2L 2 (Y,dν) , then

−1/2 (Vψ,ψ F )(g) = (cψ cψ ) dµ(g ) ψg | ψg ψg | s

=

c cψ

ψ

1/2

G

ψg | s,

by (7.8).

But the function F(g) = ψg | s is the wavelet transform of s with respect to the mother wavelet ψ. Hence, cψ s2L 2 (Y,dν) = cψ s2L 2 (Y,dν) = F 2L 2 (G,dµ) . Vψ,ψ F 2L 2 (G,dµ) = cψ cψ Thus, Vψ,ψ maps Hψ isometrically onto Hψ and its kernel coincides with H⊥ ψ .

This theorem is useful for it shows, first of all, that the various transforms that can be obtained for the same signal but using different mother wavelets, are in a sense equivalent. Second, all these transforms have the same energy content as well as the same information content. From a theoretical point of view, therefore, it does not matter

255

7.1 A group-adapted wavelet analysis

which mother wavelet is used. However, for practical or computational reasons, it may be preferable to use mother wavelets having specific properties.

7.1.4.2

Left regular representation and localization operators We had started out by choosing an irreducible representation U (g) of the group G, on the space of signals L 2 (Y, dν). Assuming U to be square integrable and taking different admissible vectors (or mother wavelets) ψ, we defined generalized wavelets ψg , using which we were able to map the signal space isometrically into subspaces Hψ of signal transforms. The isometric mapping Wψ was defined (see (7.10)) by (Wψ s)(g) = −1/2 cψ ψg | s. As a consequence of this isometry, the unitary operators U (g) are mapped to certain unitary operators on Hψ and it is interesting to obtain these image operators. Since, for arbitrary s ∈ L 2 (Y, dν) and fixed g0 ∈ G, −1/2

ψg | U (g0 )s = cψ

−1/2

U (g0−1 g)ψ | s = cψ

(Wψ U (g0 )s)(g) = cψ = cψ

−1/2

U (g0 )∗ U (g)ψ | s

−1/2

ψg0−1 g | s,

we obtain, Wψ U (g0 )Wψ−1 = Uψ (g0 ) (the inverse of Wψ being computed on its range), where, for any S ∈ Hψ , (Uψ (g0 )S)(g) = S(g0−1 g),

g ∈ G,

(7.19)

which ought to be compared with (6.35). Being the isometric image of a unitary irreducible representation, the representation given by these operators, on Hψ , is also unitary and irreducible. In other words, U (g) and Uψ (g) are equivalent representations of the group G, however, expressed on different Hilbert spaces. What is important to note here is that the right-hand side of (7.19) is independent of ψ. In other words, one obtains the same form for the transformed operators Uψ (g) on all the subspaces Hψ ⊂ L 2 (G, dµ), irrespective of the mother wavelet chosen. Using this fact let us define the operators U (g), g ∈ G, on the whole of L 2 (G, dµ), adopting this same form: (U (g)F)(g ) = F(g −1 g ), Since

U (g)F2 =

g, g ∈ G,

F ∈ L 2 (G, dµ).

(7.20)

dµ(g )|(U (g)F)(g )|2 G

dµ(g ) |F(g −1 g )|2 = F2 ,

= G

the last step following from the invariance of the measure dµ, the operators U (g) are unitary on L 2 (G, dµ). Also, clearly, the map g → U (g) is a group homomorphism. Thus, we have obtained a unitary representation of G in terms of these operators, called the left regular representation. The situation is exactly what we also had in the case of the affine group (see (6.34) and the discussion leading up to and following it). The left regular representation is not in general irreducible, and indeed, (7.19) shows that, restricted to each one of the subspaces Hψ ⊂ L 2 (G, dµ), it admits the irreducible subrepresentation Uψ (g). Another way to express this fact is to write

256

Matrix geometry of wavelet analysis. II

Wψ U (g) = U (g)Wψ = Uψ (g).

(7.21)

Summarizing, the rˆole of the representation U (g) on the signal space is played by the left regular representation U (g) on the spaces of wavelet transforms. The construction of localization operators on the group space G follows the same pattern as laid out for the affine group in Section 6.1.3. Let ψ ∈ L 2 (Y, dν) be an admissible vector, K ψ (g, g ) the corresponding reproducing kernel and ⊂ G a measurable set (with respect to the Haar measure dµ). Associated to , we define the integral kernel, aψ (g, g ) = dµ(g

) K ψ (g, g

) K ψ (g

, g ), (7.22)

and the resulting operator aψ ( ) on L 2 (G, dµ), (aψ ( )F)(g) = dµ(g ) aψ (g, g ) F(g ), F ∈ L 2 (G, dµ).

(7.23)

From (7.18) and the definition of aψ ( ), it follows that F | aψ ( )F = dµ(g) |(Pψ F)(g)|2 ,

(7.24)

G

which shows that the operator is bounded, self-adjoint and has positive spectrum. In particular, for S ∈ Hψ , S | aψ ( )S = dµ(g) |S(g)|2 .

Thus, the quantity p S ( ) =

S | aψ ( )S , S2

(7.25)

represents the fraction of the wavelet transform of s (where S(g) = ψg | s) which is localized in the region . This also motivates the term localization operator for aψ ( ). Additionally, these operators have the measure theoretical properties as those obtained for the localization operators of the affine group (see (6.53)): aψ (G) = Pψ , aψ (∅) = 0, 8 7

j = ∅, whenever i = j, aψ ( i ), if i aψ ( i ) = i∈J

(7.26)

i∈J

where, again, the sum in (7.26) is to be understood in the sense of scalar products. These relations also imply that p S ( ) is a probability measure. Finally, the localization operators satisfy the imprimitivity or covariance condition, U (g) aψ ( ) U (g)∗ = aψ (g ), where g is the shifted set,

(7.27)

257

7.1 A group-adapted wavelet analysis

g = {gg ∈ G | g ∈ }. The above covariance relation can be derived in the same way as (6.57) and again leads to the corresponding relation, p S (g ) = pU (g)−1 S ( ),

(7.28)

for the probability measure p S ( ). We interpret this relation in the same way as in the case of the affine group namely, that the probability of localization in the transformed set is the same as the probability of localization of the transformed signal in the original set.

7.1.5

Wavelet transforms on general quotient spaces The group theoretical analysis outlined above was intended to underscore the fact that, using purely symmetry arguments, one can arrive at a general wavelet transform, which then displays all the basic properties of the standard 1-D wavelet transform. However, the power of this general group theoretical analysis lies in its applicability to a vast number of other symmetry groups, thus opening up the possibility of constructing extremely general classes of wavelet transforms. Many of these turn out to be of enormous practical value, as well. It is in this light that we will undertake in Section 7.2 a general analysis of the 2-D wavelet transform. Building the generalized wavelet transform, defined in (7.10), depended on the assumption that the underlying group representation U (g) was square integrable – in the sense that there existed a vector ψ satisfying the admissibility condition (7.4). Furthermore, since we generally wish to identify the space of signal variables with the structure of a physical phase space, the group G itself would have to possess such a structure, if wavelet transforms are to be defined as functions over it. However, already in the case of the Gabor transform, we saw that these conditions were not fulfilled in the strict sense. Indeed in that case admissibility was only defined with respect to a quotient space of the group (see (6.114) and the discussion following it); it was this quotient space which had the structure of a phase space and on which the Gabor transform was defined. Let us briefly indicate here how this sort of a construction can be put against a more general setting. As a first case, consider the situation where the analyzing wavelet ψ has a nontrivial isotropy subgroup Hψ ⊂ G, up to a phase. This means that ψ satisfies the condition U (h)ψ = eiα(h) ψ,

h ∈ Hψ ,

(7.29)

where α(h) is a (real) phase factor, generally depending on h. (In the physical literature, this is the setting for the construction of Gilmore–Perelomov type of coherent states [191,192,Per86,312].) Clearly now, the integrand in (7.4) does not depend on g, but only on the coset g Hψ ∈ G/Hψ , so that the finiteness of the integral would force the subgroup Hψ to be compact. Failing that, we assume that the quotient space X = G/Hψ

258

Matrix geometry of wavelet analysis. II

carries an invariant measure ν. (Recall that elements of the quotient space X are the cosets g Hψ , g ∈ G, and they transform under the action of an element g0 of the group in the manner, g Hψ → g0 g Hψ .) We impose the weaker admissibility condition [191,192,312], dν(x) |U (g)ψ|φ|2 < ∞, ∀ φ ∈ H (x ≡ g Hψ ), (7.30) X

on ψ. The integrand in (7.30) manifestly does not depend on individual elements g ∈ G, only on their cosets modulo Hψ , x ≡ g Hψ ∈ G/Hψ , and the condition (7.30) makes sense. This condition means that the representation U is square integrable on the coset space X = G/Hψ or, as it is called, square integrable modulo the subgroup Hψ . Notice that the latter need no longer be compact. In order to define wavelets, it is necessary to go back to the group. We do this using the notion of a section. This is a map σ : X → G, chosen so that if σ (x) = g then x = g Hψ . In these terms, the admissibility condition (7.30) may be rewritten in the slightly different, but completely equivalent form: c X (ψ, φ) = dν(x) |U (σ (x))ψ|φ|2 < ∞, ∀ φ ∈ H, (7.31) X

where σ is an arbitrary section σ : X → G. Indeed, since two different sections σ and σ are related as σ (x) = σ (x)h(x), where h(x) ∈ Hψ , it is obvious that the integrand does not depend on the choice of the section. Correspondingly, the wavelet vectors are written as ψσ (x) = U (σ (x))ψ, x ∈ X , which emphasizes that the proper index set is X = G/Hψ and not G. Under the condition (7.30) or (7.31), the whole construction may be performed exactly as before [Ali00,6]. In particular, the map Wψ : H → L 2 (X, dν) given by −1/2 (Wψ s)(x) ≡ c X ψσ (x) |s is an isometry, where c X ≡ c X (ψ, ψ); in other words, one has a resolution of the identity c−1 dν(x) |ψσ (x) ψσ (x) | = I. (7.32) X X

From this follows, as before, that the range of Wψ is a closed subspace Hψ of L 2 (X, dν), the corresponding projection Pψ = Wψ Wψ∗ is an integral operator with (reproducing) kernel K (x , x) = c−1 X ψσ (x ) |ψσ (x) , the familiar reconstruction formula holds, etc. Coming back to the subject of this book, it is true that the continuous wavelet transform, both in one and two dimensions, are examples of wavelet transforms living directly on the associated group, G (+) aff and SIM(2), respectively (see below). However, we have also seen in Section 6.3 that the Gabor transform is an example of a construction modulo a subgroup, in this case the phase subgroup + of the Weyl–Heisenberg group G WH . In the same way, in dimensions higher than 2, the CWT with respect to an axisymmetric wavelet leads to wavelet transforms defined on a quotient of the above type (see Section 9.1). Physically, this means that while the total set of signal symmetries may be large,

259

7.2 The 2-D continuous wavelet transform

because of the needs of the problem at hand, the wavelet transform is defined over a smaller set of parameters. Moreover, it is this quotient space which turns out to be the relevant phase space of the problem. Actually, one can go a step further, and extend the whole construction to the case of an arbitrary coset space X = G/H , where H is not the stability subspace of any vector ψ in the sense of (7.29). The main difference is that (i) the validity of the admissibility condition (7.31) may depend on the choice of the section σ ; and (ii) when the condition holds, it reads 0< dν(x) |U (σ (x))ψ|φ|2 = φ | Aσ φ , ∀ φ ∈ H, (7.33) X

where Aσ is a bounded positive invertible operator, sometimes called the resolution operator. Equivalently, the resolution of the identity (7.32) becomes dν(x) |ψσ (x) ψσ (x) | = Aσ . (7.34) c−1 X X

Note that A−1 σ may be unbounded in general. In the case where it is bounded, the system of wavelets {ψσ (x) , x ∈ X } is called a (continuous) frame. (The unbounded case yields a far reaching generalization of the notion of a frame discussed in Section 2.4.1.) However, this extension of the theory of wavelets will not concern us in this book, with the sole exception of wavelets on the 2-sphere, discussed at length in Section 9.2.

7.2

The 2-D continuous wavelet transform Group theoretically, we expect the 2-D continuous wavelet transform to arise from a group which should be an appropriate generalization of the affine group. This indeed is the case, and we are led to it by a relatively straightforward analysis of the symmetries which might be attributed to a two-dimensional signal. It turns out that the group in question, the 2-D similitude group, is a generalization of the affine group and in fact contains it as a subgroup.

7.2.1

The similitude group and 2-D wavelets We begin with a model of a two-dimensional image. For our purposes, a 2-D image will be a finite energy signal s ∈ L 2 (R2 , d 2 x), as discussed in Section 2.1.1. The operations we want to apply to s are translations in the image plane (b ∈ R2 ), global dilations (zooming in and out by a > 0) and rotations around the origin (θ ∈ [0, 2π )). Together these transformations constitute a four-parameter group, called the similitude group of the plane and denoted by SIM(2). The action on the plane is a, θ )y = arθ y + b, x = (b, where rθ is the 2 × 2 rotation matrix,

(7.35)

260

Matrix geometry of wavelet analysis. II

rθ =

cos θ sin θ

− sin θ cos θ

.

(7.36)

Our convention is to take a quantity such as x to be a column vector, the corresponding a, θ ) row vector being x T . A convenient representation of the joint transformation (b, is in the form of 3 × 3 matrices b ar θ a, θ ) ≡ , 0 T = (0, 0). (7.37) (b, 0 T 1 Matrix multiplication then replicates the composition of successive transformations and thus we derive the group law, a, θ )(b , a , θ ) = (b + arθ b , aa , θ + θ ) (b, 1, 0), (unit element) e = (0, a, θ)−1 = (−a −1r−θ b, a −1 , −θ). (b, From this, we deduce the following. r The set of rotations (0, 1, θ), θ ∈ [0, 2π), is a subgroup of SIM(2) and so also is the set of dilations (0, a, 0), a > 0. Moreover, these two subgroups commute, i.e., (0, 1, θ)(0, a, 0) = (0, a, 0)(0, 1, θ ) = (0, a, θ ). r The set of all translations (b, 1, 0), b ∈ R2 , is also a subgroup. Moreover, it has the a, θ) ∈ SIM(2) is structure of an invariant subgroup in the following sense: if (b, arbitrary and (b0 , 1, 0) any element of the translation subgroup, then a, θ)−1 = (arθ b0 , 1, 0), a, θ )(b0 , 1, 0)(b, (b, which again is an element of the same subgroup. Thus, the similitude group SIM(2) has the structure of a semidirect product: SIM(2) = R2 (R+ ∗ × SO(2)) where R2 is the subgroup of translations, R+ ∗ that of dilations, and SO(2) of rotations. Topologically, we can identify R2 with C, the complex plane and R+ ∗ × SO(2) with ∗ C , the complex plane with the origin removed. Thus we may write SIM(2) = C C∗ , and denoting a group element by (z, w), where z ∈ C and w ∈ C∗ , the group composition law is, very simply, (z 1 , w1 )(z 2 , w2 ) = (z 1 + w1 z 2 , w1 w2 ). In particular, if we only consider elements (z, w) for which z = b + ic, with c = 0, and w = aeiθ , with θ = 0, then these elements clearly constitute a subgroup, which is just the affine group of the line. Thus, G aff ⊂ SIM(2), meaning that the similitude group

261

7.2 The 2-D continuous wavelet transform

is a generalization of the affine group, as indicated earlier. In fact, we may consider SIM(2) as being a complexification of G aff . Let us compute next the left (invariant) Haar measure on SIM(2). If (b0 , a0 , θ0 ) is a a, θ ) arbitrary, then writing fixed element of the group and (b, a0 a, θ0 + θ ), a, θ) = (b0 + a0rθ0 b, (b , a , θ ) = (b0 , a0 , θ0 )(b, and noting that det [rθ0 ] = 1, we get d 2 b = a02 d 2 b,

da = a0 da,

dθ = dθ.

Thus, the measure 1 2 (7.38) d b da dθ, a3 is invariant under left transformations. Similarly, the right Haar measure can be computed to be

a, θ) = dµ(b,

a, θ ) = 1 d 2 b da dθ, dµr (b, a and thus like the affine group, the SIM(2) group is also nonunimodular. a, θ represent parameters in terms of which we want to analyze the signals Since b, 2 s ∈ L (R2 , d 2 x), we shall identify the Hilbert space L 2 (SIM(2), dµ) with the space of 1 all finite energy 2-D wavelet transforms. It will turn out that SIM(2) R2 × R+ ∗ ×S again has the structure of a phase space (S 1 being the unit circle). We shall later analyze the orbits of SIM(2) under the coadjoint action and we shall see that there is only one nontrivial orbit, which topologically is isomorphic to the group itself. Correspondingly, there is only one nontrivial unitary irreducible representation of SIM(2). This representation, which is a straightforward realization of the action (7.35) on the space of signals a, θ ) (see (2.13)): L 2 (R2 , d 2 x), is given by the operators U (b, a, θ)s ( x − b)), b ∈ R2 , a > 0, 0 θ < 2π. (7.39) U (b, x ) = a −1 s(a −1 r−θ ( The fact that these operators define a unitary representation is straightforward to verify. 2 , d 2 k) of Fourier-transformed signals In the space L 2 (R s, this representation acquires the form (see (2.14)) (b, a, θ] = a e−i b· k U s (k) s(ar−θ k). (7.40) This representation is also square integrable, with an admissibility condition on mother wavelets, see (2.16), which is analogous to (6.15). Indeed we have the result: a, θ) defines a unitary irreducible repTheorem 7.2.1 . The family of operators U (b, resentation of SIM(2) in the Hilbert space L 2 (R2 , d 2 x), which is unique up to unitary equivalence. This representation is square integrable, and a vector ψ ∈ L 2 (R2 , d 2 x) is admissible if, and only if, it verifies the condition:

262

Matrix geometry of wavelet analysis. II

cψ ≡ (2π)

2 R2

d 2 k 2 |ψ(k)| < ∞. 2 |k|

(7.41)

Proof . That U is a representation of SIM(2) results from explicit computation; its unitarity is obvious and its irreducibility follows from Proposition 2.1.2. As for the square integrability, it is proved by a direct calculation, using the Fourier-transformed (see above), realization U + +2 + | ψ ++ d 2 b da dθ I (ψ) ≡ + U (b, a, θ ) ψ a3 SIM(2) ∞ da 2π 2 k) ψ( r−θ (k)) d b dθ d 2 k ei b·k ψ(a = a 0 0 R2 R2 × d 2 k e−i bk ψ(ar −θ (k )) ψ(k ). R2

Integrating first over b (the permutation of integrals is allowed by Fubini’s theorem) yields a factor (2π )2 δ(k − k ) and, therefore, ∞ + +2 + +2 da 2π + k) ++ ++ ψ( ++ . I (ψ) = (2π)2 r−θ k) dθ d 2 k + ψ(a a 0 0 From this, we get, exactly as in the proof of Proposition 2.2.1, I (ψ) = cψ ψ 2 , with cψ given by (7.41), which proves the statement.

in the Fourier transformed space: Introducing the Duflo–Moore operator C, ψ 2, cψ = C we obtain, ψ)( k) = (C

2π ψ(k), |k|

(7.42)

which should be compared to (6.16). Thus every function ψ ∈ L 2 (R2 , d 2 x), such that (i.e., satisfies (7.41)), is an admissible its Fourier transform lies in the domain of C vector and can be used to build wavelets. Following our established practice, we shall call such vectors mother wavelets. Choosing a mother wavelet ψ, we define the 2-D wavelets as: 1 1 ψb,a,θ r−θ ( ( x ) = U (b, a, θ)ψ ( x) = ψ x − b) , (7.43) a a

263

7.2 The 2-D continuous wavelet transform

and the 2-D continuous wavelet transform as the inner product of the signal s with the wavelet ψb,a,θ : a, θ) = ψb,a,θ | s, S(b,

(7.44)

which is a function on SIM(2) (see (2.18)–(2.20)). All the general properties of wavelets, as outlined in Section 7.1 and in particular the relations (7.6)–(7.11) and (7.21)–(7.27), follow in a straightforward manner. Some of these were worked out in detail in Section 2.2. Specific examples of 2-D wavelets with special symmetry properties have also been worked out in Chapter 2. For the sake of illustration, we display here the general resolution of the identity and reconstruction formula for signals. Following (7.7) we may write, ∞ ∞ ∞ 2π 1 da

db1 db2 3 dθ |ψb,a,θ ψb,a,θ | = I, (7.45)

Cψ | Cψ −∞ −∞ 0 0 a provided, (b1 , b2 being the components of the vector b), Cψ | Cψ = (2π )2

R2

d 2 k

k) (k) ψ ψ( = 0. 2 |k|

From this we obtain the reconstruction formula for a signal, ∞ ∞ ∞ 2π 1 da a, θ)ψ , db1 db2 3 dθ Sψ (b, s= b,a,θ Cψ | Cψ −∞ −∞ 0 0 a

(7.46)

a, θ) = ψb,a,θ where Sψ (b, | s is the wavelet transform of s in terms of the mother wavelet ψ. One ought to comment here on the freedom that one has in designing the 2-D wavelet transform. On the one hand, one may ignore the rotation variable θ, for instance, if directions are irrelevant. This is achieved by choosing an isotropic or rotation invariant wavelet, ψ(rθ ( x )) = ψ( x ). Equivalently, one may consider as transformations of the plane only translations and dilations, with the corresponding group R2 R+ ∗ . In this case, however, the representation structure is much more complicated, since every a, θ ), subspace of the form L 2 (C, d 2 k) is invariant under the action of the operators U (b, where C is a cone with apex at the origin in the k-plane. More interesting is the opposite move. If, besides the similitude operations, one considers also certain types of deformations, one gets a larger group, namely, the group obtained by replacing in (7.37) the matrix arθ ∈ R+ ∗ × SO(2) by an arbitrary nonsingular 2 × 2 real matrix. This group is much more complicated and so also is its representation structure. In fact, it is unlikely that any of its representations would be of much use for our purposes. Thus, the similitude group SIM(2) seems to occupy a privileged position in the construction, although later, in Section 7.4, we shall study a second group in which the spatial rotations rθ are replaced by hyperbolic rotations.

264

Matrix geometry of wavelet analysis. II

7.2.2

The group as the primary object We started out by defining a two-dimensional image as a function on R2 and then obtained the group SIM(2) by considering a set of transformations on R2 (see (7.35)), which would lead to the physically desirable transformations (2.7)–(2.9) on the signal space. It is possible to reverse the argument, i.e., to start with the group SIM(2) as the primary object and then to obtain the space R2 , over which the images are to be defined, as intrinsic to the group and on which it has the natural action given by (7.35). To see this, note first that the set of matrices in SIM(2), arθ 0 , a > 0, θ ∈ [0, 2π ), 0 T 1 constitute the subgroup H of rotations and dilations and since, I2 b arθ b arθ 0 , = 0 T 1 0 T 1 0 T 1

(7.47)

the set of matrices I2 y , y ∈ R2 , 0 T 1 is identifiable with the quotient space SIM(2)/H . Acting on such a matrix from the left, by an element of SIM(2), we see that, I2 y arθ b arθ arθ y + b = 1 0 T 1 0 T 1 0 T =

I2 0 T

arθ y + b 1

arθ

0

0 T

1

,

implying the transformation y → arθ y + b on R2 . Thus, the action of SIM(2) on the quotient space SIM(2)/H is the same as its action (7.35) on R2 . The situation here is the same as that encountered in the case of the affine group (see (6.10) and the discussion following). One can just as well adopt the point of view that the group is the basic geometrical object, from which signals, their transformation properties and their representations in various spaces of functions, all follow as mathematical consequences.

7.2.3

Decomposition theory of 2-D wavelet transforms As in the case of the affine group, we would like to identify the space L 2 (SIM(2), dµ) with the set of all finite energy 2-D wavelet transforms. In order to do this, we have to be able to decompose any vector S ∈ L 2 (SIM(2), dµ) into a sum (possibly infinite) of

265

7.2 The 2-D continuous wavelet transform

wavelet transforms of appropriate signals with respect to appropriate mother wavelets. For the affine group this was achieved in (6.41) and we would like to do the same in the 2-D case. Since the Duflo–Moore operator C has an inverse, we can choose an orthonormal 2 2 2 basis, {φn }∞ ), such that each φn is in the domain of n=1 , in the signal space L (R , d x −1 −1 C . Thus, the vectors ψn = C φn are admissible. We recall next that the orthogonality condition (7.12) implies that all wavelet transforms are elements of L 2 (SIM(2), dµ). a, θ) denotes the wavelet transform of a signal s with respect to Specifically, if Sn (b, a, θ) that of a signal s with respect to the mother the mother wavelet ψn and Sm (b, wavelet ψm , then a, θ ) Sn (b, a, θ) Sm (b, a, θ) = δnm s | s . dµ(b, (7.48) SIM(2) For each n = 1, 2, . . . , ∞, denoting by Hψn the space of all wavelet transforms with respect to the mother wavelet ψn , we infer from the above equation that these spaces are 2 mutually orthogonal. Furthermore, ⊕∞ n=1 Hψn ⊆ L (SIM(2), dµ), and since the group has only one irreducible unitary representation, one can in fact show that ∞

Hψn L 2 (SIM(2), dµ),

(7.49)

n=1

where denotes (unitary) equivalence. (A general discussion of such decompositions, and related results, may be found in [Ali00].) Thus, we have justified the expansion S=

∞

Sn ,

S ∈ L 2 (SIM(2), dµ),

(7.50)

n=1

of an arbitrary element of L 2 (SIM(2), dµ) in terms of wavelet transforms. Note that, a, θ) = U (b, a, θ)ψn | s, for some signal vector s ∈ L 2 (R2 , d 2 x), in the above, Sn (b, where in general, s is different for different n. Note also that the sum in (7.50) holds a, θ ) are continuous in the sense of the L 2 -norm, so that, although the functions Sn (b,

∞ in all variables, we only have S(b, a, θ) = i=1 Sn (b, a, θ ) almost everywhere (with respect to dµ). We know, from the general theory outlined in Section 7.1.4, that each one of the subspaces Hψn is a reproducing kernel Hilbert space. Let a, θ; b , a , θ ) = U (b, a, θ)ψn | U (b , a , θ )ψn K ψn (b, be the reproducing kernel for Hψn . Then, given S ∈ L 2 (SIM(2), dµ), the component Sn appearing in (7.50) can be computed using (7.18): a, θ; b , a , θ ) S(b , a , θ ), dµ(b , a , θ ) K ψn (b, (7.51) Sn (b, a, θ) = SIM(2)

a, θ) = U (b, a, θ )ψn | s, the signal vector s may be computed using and, writing Sn (b, (7.46):

266

Matrix geometry of wavelet analysis. II

s= SIM(2)

a, θ ) Sn (b, a, θ) ψb,a,θ dµ(b, .

(7.52)

If we introduce the basic wavelet transforms, a, θ ) = U (b, a, θ )ψn | φm , Snm (b,

φm = Cψm , n, m = 1, 2, . . . , ∞,

(7.53)

then, by (7.48) and the orthonormality of the φn , these functions are seen to satisfy a, θ ) Snm (b, a, θ) Sk (b, a, θ) = δnk δm . dµ(b, (7.54) SIM(2)

Hence, any S ∈ L 2 (SIM(2), dµ) has the orthogonal decomposition ∞

a, θ ) = S(b,

a, θ), cnm Snm (b,

(7.55)

n,m=1

with

a, θ) Snm (b, a, θ) S(b, a, θ), dµ(b,

cnm =

(7.56)

SIM(2)

and

a, θ) |S(b, a, θ)|2 = dµ(b,

S2 = SIM(2)

7.2.3.1

∞ ∞

|cnm |2 .

n=0 m=−∞

A concrete example Finally, we give a concrete example of the decomposition (7.50), in terms of mother wavelets built out of the well-known trigonometric functions and Laguerre polynomials. 2 , d 2 k), we choose the basis vectors In the Fourier-transformed signal space L 2 (R , 1 ,−1/2 e− 2 L n (,) eimϑ , (2π )1/2 n = 0, 1, 2, . . . , ∞ , m = 0, ±1, ±2, . . . , ±∞ ,

nm (k) = φ

(7.57)

where ,, ϑ are the polar coordinates of k and the L n (,) are the Laguerre polynomials: n n (−,)k L n (,) = . k k! k=1 These satisfy the orthogonality relations, ∞ L m (,) L n (,) e−, d, = δmn , 0

nm are orthonormal, implying that the φ k = δnk δm . nm | φ φ

(7.58)

267

7.2 The 2-D continuous wavelet transform

2 , d 2 k), follows from well-known properties of The fact that they form a basis of L 2 (R Laguerre polynomials and trigonometric functions. Moreover, it is clear that the vectors −1 φ nm = C nm , where ψ nm (k) nm (k) = , φ ψ 2π 1 1/2 − ,2 = e L n (,) eimϑ , 3 , (2π ) 2 n = 0, 1, 2, . . . , ∞, m = 0, ±1, ±2, . . . , ±∞,

(7.59)

2 , d 2 k) and hence legitimate mother are also elements of the Hilbert space L 2 (R wavelets. nm , we can construct the spaces Hψnm of wavelet Using the mother wavelets ψ 2 , d 2 k). (b, nm | The total space of a, θ) = U a, θ)ψ transforms Snm (b, s , s ∈ L 2 (R all transforms would then decompose as: L 2 (SIM(2), dµ)

∞ ∞

Hψnm .

n=0 m=−∞

7.2.3.2

Decomposition into orthogonal angular channels We saw, at the end of Section 6.1.2, how a 1-D wavelet transform can be analyzed into wavelet transforms in orthogonal channels. Here we carry out a similar decomposition of a 2-D wavelet transform into orthogonal angular channels, again following a suggestion nm defined above. Let ψ be an arbitrary mother in [105]. We use the mother wavelets ψ wavelet in the Fourier domain. We may then write, k) = ψ(

∞ ∞

mn (k) = amn ψ

n=1 m=1

∞

am (,) eimϑ ,

(7.60)

m=1

where am (,) =

∞

1 (2π )

3 2

,

anm ,1/2 e− 2 L n (,),

mn | C ψ. amn = φ

(7.61)

n=1

The sum in (7.60) may be looked upon as a decomposition of the mother wavelet into angular channels. Next, writing m (k) = am (,) eimϑ , ψ

(7.62)

m is a vector which is in the domain of the Duflo–Moore operator C we easily see that ψ a, θ ) is the 2-D wavelet transform and hence it can be used as a mother wavelet. If Sψ (b, 2 2 2 of a signal s ∈ L (R , d k), then it is straightforward to verify that a, θ) = Sψ (b,

∞ m=1

a, θ), Sm (b,

(b, m | a, θ ) = U a, θ )ψ Sm (b, s.

(7.63)

268

Matrix geometry of wavelet analysis. II

It is clear that, for m = n, the wavelet transforms Sm and Sn are orthogonal functions m and ψ n are themselves orthogonal in L 2 (SIM(2), dµ), while the mother wavelets ψ 2 2 2 vectors in L (R , d k). Thus we call (7.63) a decomposition of the wavelet transform a, θ ) into orthogonal angular channels, the transform Sm (b, a, θ ) being the comSψ (b, ponent along the m-th channel.

7.3

2-D wavelets on phase space In Section 2.3.2 we pointed out how the 2-D wavelet transform could also be looked upon as a function on a physical phase space. Here we take up this point again and give a more exhaustive mathematical treatment of it. The SIM(2) group has only one nontrivial coadjoint orbit and hence only one phase space. Moreover, this phase space is topologically homeomorphic to the group itself, meaning that wavelet transforms may also be viewed upon as transforms built on this phase space. In order to analyze these features, it will first be necessary to study the matrix structure of the generators of the various transformations constituting the group.

7.3.1

Lie algebra and orbits The four basic sets of operations of dilation, rotation and the two translations, each constitute one-parameter subgroups of SIM(2). More precisely, these subgroups are generated by group elements of the type et , 0), a, θ ) = (0, (b, t ∈ R, or (b, a, θ ) = (0, 1, t), t ∈ [0, 2π ), a, θ) = (ei t, 1, 0), t ∈ R, i = 1, 2, or (b, where 0 0 = , 0

1 , e1 = 0

0 e2 = , 1

constitute one-parameter subgroups. A general element of SIM(2) can be written as a product of elements of these subgroups: 1, θ ) (0, a, 0) , where b = b1 . a, θ ) = (e1 b1 , 1, 0) (e2 b2 , 1, 0) (0, (b, b2 Generically, writing elements in any one of these subgroups as g(t) and computing the derivative at the identity: + + d , g(t)++ dt t=0

as was done for the affine group in (6.77), we obtain the four 3 × 3 matrices

269

7.3 2-D wavelets on phase space

1 0 0 D = 0 1 0, 0 0 0 0 0 1 P1 = 0 0 0 , 0 0 0

0 −1 0 J = 1 0 0, 0 0 0 0 0 0 P2 = 0 0 1 . 0 0 0

(7.64)

et , 0), J that of rotations, et J = (0, 1, t), Here D is the generator of dilations, et D = (0, t Pi and P1 , P2 those of translations, e = (tei , 1, 0), i = 1, 2. The four generators satisfy the commutation relations [D, J ] = 0,

[D, Pi ] = Pi , i = 1, 2,

[J, P1 ] = P2 ,

[J, P2 ] = −P1 ,

[P1 , P2 ] = 0, and together they generate the Lie algebra of SIM(2) which, in this case, is a fourdimensional real vector space. We denote this Lie algebra by sim(2) and its dual space by sim(2)∗ . A general element of the Lie algebra sim(2) has the form λ −θ β1 = λD + θ J + β1 P1 + β2 P2 = (7.65) X ( α , β) θ λ β2 , 0 0 0 with θ, λ, β1 , β2 ∈ R and λ β1 . α = , β = θ β2

(7.66)

v ), Corresponding to a 2-vector v, consider again, as in (2.52), the 2 × 2 matrix s( v1 −v2 v1 s(v ) = , v = . (7.67) v2 v1 v2 Then, for any two vectors v and w,

s(v )s(w) = s(w) s( v ),

s(v )w = s(w) v.

Using s, the general Lie algebra element (7.65) may be rewritten as s ( α ) β = X ( α , β) . 0 0 yields the group element Exponentiating the matrix X ( α , β)

(7.68)

(7.69)

270

Matrix geometry of wavelet analysis. II

g=e

X ( α ,β)

=

eλ rθ

F(s( α ))β

0

1

,

(7.70)

α )) is defined as the sum of an infinite where now 0 θ < 2π and the 2 × 2 matrix F(s( series: α )) = I2 + F(s(

[s( s( α ) [s( α )]2 α )]3 + + + ... . 2! 3! 4!

(7.71)

We shall see below that every group element can be so obtained, by exponentiating an appropriate Lie algebra element. It will be useful for the sequel to express the operator F(s( α )) in a somewhat different form. Let us define a function, sinch, of a real variable u as sinch u =

sinh u , u

sinch 0 = 1.

(7.72)

This is a positive, infinitely differentiable function and so also is the related function u u F(u) = e 2 sinch ( ). 2

(7.73)

This latter function has the Taylor expansion u u u u2 u3 + + + ... . F(u) = e 2 sinch ( ) = 1 + 2 2! 3! 4!

(7.74)

For u = 0, we may also write u u F(u) = e 2 sinch ( ) = u −1 (eu − 1). 2

Using the function F we now define the 2 × 2 matrix valued function A A A2 A3 A = F(A) = I2 + + + + . . . , F(O2 ) = I2 , e 2 sinch 2 2! 3! 4!

(7.75)

(7.76)

for any 2 × 2 real matrix A and where O2 and I2 are, respectively, the 2 × 2 null and identity matrices. If detA = 0, then A A F(A) = e 2 sinch (7.77) = A−1 [e A − I2 ], 2 and if, det[e A − I2 ] = 0, we shall also write A

F(A)

−1

e− 2 = [e A − I2 ]−1 A. = sinch A2

Hence, for | α | = 0,

(7.78)

271

7.3 2-D wavelets on phase space

−eλ sin θ λ θ eλ cos θ − 1 eλ sin θ −θ λ eλ cos θ − 1 s( α) sinch , 2

1 F(s( α )) = 2 λ + θ2 =e

s( α)

2

and α ))] [F(s(

−1

1 = 2(cosh λ − cos θ ) =

e− sinch

cos θ − e−λ − sin θ

sin θ cos θ − e−λ

(7.79) λ θ

−θ λ

s( α)

2

s( α)

.

(7.80)

2

Going back to (7.70), we rewrite it as s( α) α) λ 2 sinch s( r e β e θ 2 , g = e X (α,β) = T 0 1

(7.81)

θ, a) in the form given in (7.37), we find the relations and writing (b, λ = log a,

α ))]−1 b , β = [F(s(

(7.82)

between the group parameters and those of the Lie algebra. This also shows that any θ, a) can be written as the exponential of some element X ( in group element (b, α , β) the Lie algebra. The group SIM(2) acts on its Lie algebra sim(2) via the adjoint action: a, θ) X ( a, θ)−1 = X ( = (b, (b, X ( α , β) α , β) α , β) Ad(b,a,θ) α + a rθ β s( α ) −s(b) , = 0 T 0

(7.83)

(in computing the above, we have made use of the fact that rθ s( α )r−θ = s( α ).) Writing

X ( α , β) = X ( α , β ), we get the transformation rules for the components: α = α α + a rθ β , β = −s(b)

(7.84)

θ, a) of the adjoint action in the {D, J, P1 , P2 } basis becomes so that the matrix M(b, I2 O2

α α a, θ ) = , (7.85) M(b, = M(b, a, θ) β . β −s(b) a rθ The adjoint action induces an action on the dual sim(2)∗ of the Lie algebra, which we now proceed to determine. Let {D ∗ , J ∗ , P1∗ , P2∗ } denote the elements of the basis in sim(2)∗ which is dual to the basis {D, J, P1 , P2 } of sim(2). We write a general element X ∗ ∈ sim(2)∗ as

272

Matrix geometry of wavelet analysis. II

X ∗ = α1∗ D ∗ + α2∗ J ∗ + β1∗ P1∗ + β2∗ P2∗ , We also set α1∗ α ∗ = , α2∗

β∗ =

β1∗ β2∗

α1∗ , α2∗ , β1∗ , β2∗ ∈ R.

and

γ =

α ∗ β∗

(7.86)

,

(7.87)

so that that the dual pairing between X and X ∗ is given by X ∗ ; X = α ∗ · α + β∗ · β.

(7.88)

a, θ) on sim(2)∗ , in the The matrix of the coadjoint action of a group element (b, a, θ), is the transposed inverse of the matrix (7.85) of the above basis, denoted M * (b, adjoint action: −1 T I a s ( b) r 2 θ * −1 T a, θ) )] = . (7.89) M (b, a, θ ) = [M((b, O2 a −1 rθ Writing γ =

α ∗ ∗

β

a, θ)γ = M * (b, a, θ) = M * (b,

α ∗ β∗

,

(7.90)

we obtain the transformation rules for the components of the dual vectors X ∗ : T β∗ , α ∗ = α ∗ + a −1 rθ s(b) β∗ = a −1 rθ β∗ , |β∗ | = a −1 |β∗ | .

(7.91)

Since orbits under the coadjoint action are the sets a, θ)γ0 | (b, a, θ ) ∈ SIM(2)}, O∗ = {M * (b,

(7.92)

for fixed vectors γ0 ∈ R4 , it is easy to see that there are exactly two types of orbits. (i) Trivial orbits: these are degenerate orbits, which are single point sets {γ0 }, obtained by choosing ∗ α1 α∗ 2 γ0 = α1∗ , α2∗ ∈ R. , 0 0 The isotropy subgroup of any such point γ0 , i.e., the subgroup which leaves it invariant, is of course the entire group SIM(2). (ii) The open free orbit: this is the only nontrivial orbit of SIM(2) and is obtained by choosing

273

7.3 2-D wavelets on phase space

0 0 γ0 = , 1

(7.93)

0 or by taking for γ0 any other vector such that at least one of its last two components is nonvanishing. This is the only orbit which concerns us here and we denote it by O∗ . Note that in this case the isotropy subgroup of γ0 (i.e., the subgroup of SIM(2) elements for which this vector is a fixed point) is just the trivial subgroup consisting of the identity element of SIM(2). This also means that topologically, the orbit is a, θ) → γ = M * (b, a, θ)γ0 , homeomorphic to the group space itself (the map, (b, from the group to the orbit, is open and free).

7.3.2

The coadjoint orbit O ∗ as a phase space Since the orbit O∗ is homeomorphic to the group SIM(2) itself, wavelet transforms a, θ ) may be considered as being transforms defined on this space. In other words, S(b, a, θ in terms of which the signal is being analyzed, can be looked upon the parameters b, as phase space parameters. This is completely in line with the situation encountered earlier, for 1-D wavelets and Gabor transforms. In order to understand better, the structure of the coadjoint orbit O∗ as a physical phase space, let us first note that points on the orbit are obtained from (7.91) upon setting 1 α ∗ = 0 and taking for β∗ the two dimensional unit vector e1 = (see (7.93)). Thus 0 a generic point γ ∈ O∗ is given as σ3 rθ b α ∗ 0 1 a, θ ) , = M * (b, γ = = a r θ e1 β∗ e1 1 0 , σ3 r−θ = rθ σ3 . (7.94) σ3 = 0 −1 a, θ) written as in (7.37), we find for the general Explicitly, with the group element (b, ∗ phase space point γ ∈ O , ∗ b1 cos θ + b2 sin θ α1 −b sin θ + b cos θ α∗ 1 2 2 1 . (7.95) γ = ∗ = a β1 cos θ β2∗

sin θ

274

Matrix geometry of wavelet analysis. II

a, θ) in terms of The above relations can be solved to express the group parameters (b, ∗ ∗ the phase space variables α , β : ∗ 1 ∗ −1 β2 b = a rθ σ3 α , , θ = tan . (7.96) a= β1∗ |β∗ | This also reflects the fact that the orbit O∗ is topologically homeomorphic to the group space, and moreover, from the form of (7.95) one infers that, geometrically, O∗ is also the cotangent bundle of R2∗ (2-D plane with the origin removed), i.e., O∗ SIM(2) R2 × R2∗ = T ∗ R2∗ .

(7.97)

Indeed, consider the point cos θ 1 ∈ R2∗ . β∗ = a sin θ Differentiating with respect to a and θ, we get the two tangent vectors at β∗ : 1 cos θ 1 − sin θ . , tθ = ta = − 2 a a cos θ sin θ We may thus take, as basis for the tangent space at β∗ , the columns of the matrix, 1 cos θ − sin θ . T= a sin θ cos θ The columns of the transposed matrix: cos θ sin θ 1 1 , , tθ∗ = ta∗ = a a − sin θ cos θ then form a basis for the cotangent (i.e., dual of the tangent) space at β∗ . An arbitrary element of this dual space has, therefore, the form b1 cos θ + b2 sin θ 1 , b1 , b2 ∈ R . α ∗ = b1 ta∗ + b2 tθ∗ = a −b sin θ + b cos θ 1

2

Thus, we shall interpret β∗ in (7.95) as representing a point in the manifold R2∗ and α ∗ as a vector in its cotangent space at this point. Physically, one calls the α ∗ configuration space vectors , while the β∗ are momentum vectors. Coadjoint orbits carry natural invariant measures under the group action (see, for example, [Kir76]). Using the transformation rules (7.91) under the coadjoint action, it is straightforward to compute the invariant measure on O∗ . Expressed in terms of the α ∗ , β∗ , it is

275

7.3 2-D wavelets on phase space

d( α ∗ , β∗ ) =

d 2 α ∗ d 2 β∗ dα1∗ dα2∗ dβ1∗ dβ2∗ = . β1∗ 2 + β2∗ 2 |β∗ |2

(7.98)

(In the physical literature, this would be called the Liouville measure for this phase space.) If we express the position vector β∗ in polar coordinates (ρ = |β∗ |, θ ), the coadjointinvariant measure (7.98) transforms to d( α ∗ , ρ, θ) = d 2 α ∗

dρ dθ. ρ

(7.99)

In these coordinates it is easy to verify that the differential 2-form ω( α ∗ , ρ, θ) =

1 dα1∗ ∧ dρ + dα2∗ ∧ dθ, ρ

(7.100)

is invariant under the coadjoint action. Moreover, the coordinate transformations on phase space, α ∗ → − α∗,

ρ →

1 , ρ

θ → −θ,

(7.101)

leave this 2-form invariant, meaning that they constitute a canonical transformation of the phase space. (We might point out that the 2-form (7.100) is just the well-known Kirillov–Kostant–Souriau symplectic structure [Kir76], carried by coadjoint orbits.) a, θ ) On the other hand, if we parametrize O∗ by means of the group parameters (b, using (7.96), the coadjoint action (7.91) transforms to group multiplication from the left. In other words, the coadjoint action Ad*g0 , corresponding to the group element a, θ ) as follows: g0 = (b0 , a0 , θ0 ) transforms the point in O∗ represented by (b, a0 a, θ0 + θ), a, θ) = (b0 + a0rθ b, a, θ ) → (b , a , θ ) = (b0 , a0 , θ0 )(b, (b,

(7.102)

and thus the invariant measure (7.98) on O∗ , changes to precisely the left Haar measure dµ in (7.38) under this transformation [compare (2.51)]: a, θ) = dµ(b, a, θ) = d(b,

1 2 d b da dθ. a3

(7.103)

There is yet another parametrization of the points of the orbit O∗ , which in a sense a, θ) parametrizations. This other is more natural than either the ( α ∗ , β∗ ) or the (b, parametrization is given in terms of the so-called Darboux coordinates, which we denote by ( q , p) and which are related to the other two sets of coordinates as q1 1 q = = rθ σ3 α ∗ = b ρ q2 cos θ cos θ 1 p1 = . (7.104) = β∗ = ρ p = p2 a sin θ sin θ

276

Matrix geometry of wavelet analysis. II

The transformation properties of these coordinates under the coadjoint action are also easily obtained. Once again, if ( q , p ) is the transform of ( q , p) under Ad*g0 , g0 = (b0 , a0 , θ0 ), then q = b0 + a0 rθ0 q,

p = a0−1 rθ0 p.

(7.105)

One can verify that the differential 2-form, ω( q , p) = dq1 ∧ dp1 + dq2 ∧ dp2

(7.106)

is invariant under the above coadjoint action and hence the corresponding Liouville measure on O∗ d( q , p) = d 2 q d 2 p = dq1 dq2 dp1 dp2 ,

(7.107)

is also invariant under this action. It is not hard to see that (7.106) and (7.107) are precisely the transforms of the 2-form (7.100) and the measure (7.103), respectively, under the coordinate change (7.104). This last choice of coordinates, and in particular the differential form (7.106) makes evident the phase space structure of the coadjoint orbit O∗ . The components of q refer to the position of the system on the configuration space R2 , while at each such point the vector p ∈ R2∗ denotes its canonical momentum. On the other hand, the fact that the a, θ) can also be used as coordinates for the orbit O∗ , shows that group parameters (b, the group itself can be identified with the phase space as well. In this case, b denotes a point in configuration space and (a −1 , θ ) are the polar coordinates of a momentum vector.

7.4

The affine Poincare´ group While the SIM(2) group is the most natural generalization of the affine group for building 2-D wavelets, it is by no means the only group which could be used. As a matter of fact, any group of the type G = R2 H , where H is a group consisting of 2 × 2 matrices, which acting on R2 generates an open free orbit, can be used to build wavelets. In other words, if the group H is such that for some fixed 2-vector x, the set, Ox = {y = hT x | h ∈ H}, h x = x if and only if h is the identity matrix, is an open set in R2 and for all x = 0, then such a group G has square integrable representations and hence can be used to build wavelets. As an example, we briefly look at the affine Poincar´e group. This group, which we denote by Paff , is a semidirect product of the above type. It differs from the SIM(2) group in that the spatial rotations rθ are replaced by hyperbolic rotations #ϑ :

277

7.4 The affine Poincare´ group

#ϑ =

cosh ϑ sinh ϑ

sinh ϑ cosh ϑ

.

(7.108)

The set of matrices {#ϑ | ϑ ∈ R} (note that det #ϑ = 1 and #−1 ϑ = #−ϑ ) constitutes a group, denoted SO0 (1, 1). (In physics, this is the group of relativistic transformations of a space–time having only one spatial dimension.) In dealing with this group, we shall use the physicists’ convention of writing the components of a vector x as x0 x = , x 0 , x ∈ R, x and use the Minkowski inner product between two such vectors: 1 0

T

x ; x = x0 x0 − xx = x g x , g= , g 2 = I2 . 0 −1

(7.109)

2 will also be defined using this inner product. If x = Duality between R2 and R

#ϑ x, then x ; x = x ; x , so that hyperbolas x02 − x2 = const are mapped into themselves by SO0 (1, 1).

7.4.1

Group structure and representations A general element of Paff has the matrix representation, a#ϑ b b0 a, ϑ) = , a > 0 , ϑ ∈ R , b = ∈ R2 , (b, b 1 0 T

(7.110)

giving the group the structure of the semidirect product, Paff = R2 (R+ ∗ × SO0 (1, 1)). Topologically, Paff R2 × C, where C is any one of the four open cones: ↑

x ∈ R2 | x02 > x2 , ±x0 > 0}, C± = { ↓

x ∈ R2 | x02 < x2 , ±x0 > 0}. C± = {

(7.111)

1, 0), b ∈ R2 , is a commutative subgroup of Paff The set of elements of the type (b, a, ϑ), a > 0, ϑ ∈ R . and so also is the set of elements, (0, The affine Poincar´e group is nonunimodular; the left and right Haar measures can be computed in exactly the same way that we computed them for the similitude group in Section 7.2.1. This time the two measures are a, ϑ) = dµ(b,

1 2 d b da dϑ, a3

and

a, ϑ) = dµr (b,

1 2 d b da dϑ, a

(7.112)

278

Matrix geometry of wavelet analysis. II

which look exactly the same as those for the SIM(2) group (to be expected, since the group SO0 (1, 1) is unimodular). The group Paff acts on the plane in the manner, y → a#ϑ y + b and therefore, we can again look for its representations in the signal space L 2 (R2 , d 2 x). As already noted, the situation here is largely similar to that of the similitude group. The signal symmetries again include translations and dilations; however, we have hyperbolic rotations now and not rigid rotations of space. Such signal symmetries could be expected in problems involving the detection of extremely fast moving objects (such as occurs, for example, in high energy physical experiments). Unlike the rotations rθ , the action of #ϑ actually has the effect of deforming the shapes of objects: the disc, x02 + x2 r 2 , is transformed into the interior of the rotated ellipse, x02 + x2 + tanh(2ϑ) x0 x0 sech(2ϑ) r 2 . Thus, if images are scanned using instruments which distort them in this manner, a group such as this could be more appropriate for their analysis than the similitude group. The natural unitary representation of Paff on the signal space L 2 (R2 , d 2 x), which a, ϑ): reflects its action y → a#ϑ y + b on R2 , is carried by the unitary operators U (b, a, ϑ)s]( x − b)), [U (b, x ) = a −1 s(a −1 #ϑ (

(7.113)

an expression which should be compared to that for the similitude group in (2.13). The unitarity of these operators is straightforward to prove; however, in contrast with the a, ϑ) do not carry an irreducible representation of operators (2.13), the operators U (b, Paff . For isolating the irreducible sectors, it is best to work in the Fourier domain. In order to do this, it will be convenient to adopt the physicists’ convention for defining the Fourier transform, using the Minkowski inner product. Accordingly, we define 2 , d 2 k), F : L 2 (R2 , d 2 x) → L 2 (R = = 1 (Fs)(k) s(k) d 2 x eik ; x s( x ), s ∈ L 2 (R2 , d 2 x), 2π R2 1 2 , d 2 k). = s( s )(k) x) = s(k), s ∈ L 2 (R (7.114) d 2 k e−ik ; x (F −1 2 2π R a, ϑ) are Using this Fourier transform, and the matrix identity, g#ϑ g = #−ϑ , the U (b, 2 2 2 seen to transform into the operators U (b, a, ϑ) on L (R , d k): (b, a, ϑ) = a eik ; b U s (k) s(a#−ϑ k). (7.115) Let C be any one of the four open cones defined in (7.111). A quick computation shows (b, a, ϑ) in (7.115) that if k ∈ C, then also a#−ϑ k ∈ C. From the nature of the operator U we see that if s has support inside this cone, then so also does the transformed function 2 , d 2 k), (b, a, ϑ) which is a subspace of L 2 (R U s. Thus, the Hilbert space L 2 (C, d 2 k), carries a subrepresentation of U (b, a, ϑ), i.e., restricted to this subspace the operators (b, a, ϑ) again define a unitary representation of Paff . This representation can also be U shown to be irreducible (see, for example, [Ali00] for a detailed proof of such results). Also, we have the obvious Hilbert space decomposition,

279

7.4 The affine Poincare´ group

2 , d 2 k) = L 2 (C+↑ , d 2 k) ⊕ L 2 (C−↑ , d 2 k) ⊕ L 2 (C+↓ , d 2 k) ⊕ L 2 (C−↓ , d 2 k), L 2 (R (b, a, ϑ) to these four suband using a self-evident notation for the restrictions of U spaces, we may write (b, +↑ (b, −↑ (b, +↓ (b, −↓ (b, a, ϑ) = U a, ϑ) ⊕ U a, ϑ) ⊕ U a, ϑ) ⊕ U a, ϑ). U

(7.116)

(b, a, ϑ) is a direct sum of four irreducible repThis shows that the representation U 2 , d 2 k). Each one of these resentations, carried by four orthogonal subspaces of L 2 (R subspaces consists of signals whose supports are contained in a cone. Returning to the inverse Fourier domain, the signal space decomposes as: ↑

↑

↓

↓

L 2 (R2 , d 2 x) = H+ ⊕ H− ⊕ H+ ⊕ H− , ↓

where, for example, H+ consists of all functions in L 2 (R2 , d 2 x) whose Fourier trans↓ forms have supports contained in C+ . Correspondingly, the representation decomposes as a, ϑ) = U+↑ (b, a, ϑ) ⊕ U−↑ (b, a, ϑ) ⊕ U+↓ (b, a, ϑ) ⊕ U−↓ (b, a, ϑ) . U (b,

(7.117)

a, ϑ) and in Let us generically represent any one of these subrepresentations by UC (b, the Fourier domain by UC (b, a, ϑ). This latter representation acts on the Hilbert space of all square integrable functions supported in the cone C. L 2 (C, d 2 k)

7.4.2

Affine Poincare´ wavelets C (b, a, ϑ) is known to be square integrable [Ali00]. Indeed, if we The representation U compute the integral C (b, | ψ| 2, a, ϑ) |U a, ϑ)ψ I (ψ) = dµ(b, Paff

∈ L 2 (C, d 2 k), we easily find for some ψ d 2 k 2 = (2π )2 ψ k)| 2. I (ψ) |ψ( 2 2 C |k0 − k |

(7.118)

From this follows the admissibility condition for an affine Poincar´e wavelet: a vec ∈ L 2 (C, d 2 k) is admissible if and only if it satisfies the integrability condition tor ψ (the < ∞, i.e., if and only if it is in the domain of the unbounded operator C I (ψ) Duflo–Moore operator), where 2π ψ)( k) k). = (C ψ)( (7.119) 2 |k0 − k2 |1/2 2 , we may also write down the resolution of the identity Setting cψ = C ψ 1 ψ | = I, a, ϑ) |ψ dµ(b, ψ b,a,ϑ b,a,ϑ b,a,ϑ = UC (b, a, ϑ)ψ, cψ Paff or, more generally,

(7.120)

280

Matrix geometry of wavelet analysis. II

1 ψ ψ | C C

Paff

ψ | = I, a, ϑ) |ψ dµ(b, b,a,ϑ b,a,ϑ

(7.121)

ψ ψ | C = 0. These equations should ψ such that C for two admissible vectors ψ, be compared to (7.41), (7.42), and (7.45). We can now go ahead and define the affine Poincar´e wavelet transform of an arbitrary by the quantity signal s ∈ L 2 (C, d 2 k) s(k). (7.122) S(b, a, ϑ) = ψb,a,ϑ | s=a d 2 k e−ik ; b ψ(a# −ϑ k) C

All the analysis carried out for the 2-D wavelet transform (obtained using the similitude group), including the phase space considerations, can again be repeated in the present case. In particular, any two-dimensional signal can be decomposed using affine Poincar´e wavelets. If the signal s is, for example, the quantum mechanical wave function of a fastmoving elementary particle (in a space–time of one time and one spatial dimension), a, and ϑ could represent its position (in space–time), its the analyzing parameters, b, mass and its rapidity.

8

Minimal uncertainty and Wigner transforms

This chapter is devoted to a brief examination of two topics. The first concerns a certain minimality property of gaborettes and how it generalizes to wavelets in one and two dimensions. The second is an analysis of the Wigner transform, as an alternative to the wavelet transform. This latter transform is extensively used in certain physical computations and in the analysis of radar signals. Notice that neither of these topics is a prerequisite for the study of the more general wavelets described in Chapters 9 and 10.

8.1

Phase space distributions and minimal uncertainty gaborettes The generalized gaborettes defined in (6.123), which give rise to holomorphic Gabor transforms, have a well-known minimal uncertainty property, related to localization in phase space. In (7.23) we had introduced the localization operators aψ ( ). As discussed in Chapter 7, these operators can be used to measure the proportion of the signal transform S which is concentrated in the (phase space) region . Consider the case of Gabor wavelets and let ψq, p ∈ L 2 (R, d x) be the family of gaborettes defined in (6.109), using the window function ψ. These vectors satisfy the resolution of the identity (6.112). Assuming the normalization ψ2 = 1/2π, we see that the operators aψ ( ) = dq d p |ψq, p ψq, p |,

give rise to the probability measure s|aψ ( )s 1 p S ( ) = = dq d p |S(q, p)|2, s2 2π s2

(8.1)

for any signal s ∈ L 2 (R, d x) with Gabor transform S. In Section 6.3 it was noted that the Gabor transform S(q, p) is a time–frequency transform, with q being the time and p the frequency parameter. On the other hand, if the signal s = s(q) is given as a function over time, then its Fourier transform, s = s( p) is a function over frequency. However, as is well known, the density distribution in time, |s(q)|2 /s2 , gives no information on the frequency content of the signal, while the frequency distribution, | s( p)|2 /s2 , gives no information on the variation of the signal with time. The phase 281

282

Minimal uncertainty and Wigner transforms

space density, |S(q, p)|2 , does however, carry information on both time and frequency, but this information is not expected to be sharp, i.e., we do not expect |s(q)|2 to be the marginal density of |S(q, p)|2 in time or | s( p)|2 to be its marginal in frequency. Indeed, computing these marginal densities, we find ∞ ∞ S1 (q) = dp |S(q, p)|2 = d x χq (x) |s(x)|2 , −∞ −∞ ∞ ∞ 2 dq |S(q, p)| = dx χ q (ξ ) | s(ξ )|2 , (8.2) S2 ( p) = −∞

−∞

where χq (x) = 2π |ψ(x − q)|2

− p)|2 , and χ p (ξ ) = 2π |ψ(ξ

(8.3)

are the shifted weight functions in time and frequency, respectively, generated by the window ψ. Thus, the marginal density S1 (q) appears as a weighted average, at each point, over the sharp density |s(x)|2 in time, while S2 ( p) appears as a similarly weighted average over the sharp density | s(ξ )|2 in frequency. It is in this sense that we should 2 think of |S(q, p)| as being an averaged phase space density, the variables q and p representing averages of their “sharp” values, computed with respect to the probability )|2 , respectively. It then becomes pertinent to ask for densities 2π |ψ(x)|2 and 2π|ψ(ξ what choice of window function, this averaging would be optimal, i.e., entail a minimum of unsharpness, for it is clear from (8.2) and (8.3) that there is no L 2 -function ψ for which S1 (q) = |s(q)|2 and S2 ( p) = | s( p)|2 . Consider the standard deviations of the (sharp) time and frequency variables, measured with respect to the probability distributions )|2 , respectively: 2π |ψ(x)|2 and 2π |ψ(ξ " σψ = 2π " σψ = 2π

∞

−∞

∞

−∞

d x x |ψ(x)| − 2π 2

2

)|2 − 2π dξ ξ 2 |ψ(ξ

∞

2 # 12 2

d x x|ψ(x)| −∞

∞

−∞

)|2 dξ ξ |ψ(ξ

,

2 # 12 .

(8.4)

It is well known, from the theory of Fourier transforms, that their product satisfies the inequality σψ .σψ 12 and that equality is attained when ψ(x) is a Gaussian as in (6.116) (or one of the other modified gaborettes ψ(0,q, p) , constructed using such a Gaussian). Hence the choice of a Gaussian for the window function ψ would, in the light of the present analysis, lead to a Gabor transform S(q, p) which measures the variables q and p with optimal accuracy. In order to put the above discussion in operator terms and to make the connection with group theory again, let us assume that the chosen window function ψ satisfies the symmetry property |ψ(x)|2 = |ψ(−x)|2 , almost everywhere. A fairly straightforward computation (see, for example, [5] for details) then leads to the following interesting average values:

283

8.1 Phase space distributions and minimal uncertainty gaborettes

q=

R

s|Qs d p S (q, p) q = , 2π s2

p=

R

d p S (q, p) p =

s|Ps . 2π s2

(8.5)

where Q and P are (unbounded) self-adjoint operators, defined by the integral relations Q= dq d p q|ψq, p ψq, p |, P= dq d p p|ψq, p ψq, p |. (8.6) R2

R2

By (8.1), the phase space probability distribution p S has the density ρ(q, p) =

1 |S(q, p)|2 2π s2

and hence it follows from (8.5) that the operator Q (respectively, P) gives the mean value of the phase space position parameter q (respectively, momentum parameter p), computed using the probability distribution determined by the Gabor transform S(q, p) of the signal s. Indeed, for an arbitrary signal vector s, having Gabor transform S corresponding to the window ψ, we get s|Qs = dq d p q|S(q, p)|2 , s|Ps = dq d p p|S(q, p)|2 . (8.7) R2

R2

It is remarkable that the values of the two integrals in (8.7) depend only on the signal vector and not on the window ψ. To see this, we note that the actions of these operators on the Hilbert space L 2 (R, d x) are easily calculated. One obtains, (Qs)(x) = xs(x),

(Ps)(x) = −

i d s(x), λ dx

(8.8)

on vectors s taken from appropriate domains. Thus, for λ−1 = , Q and P are the well-known position and momentum operators of quantum mechanics, satisfying the commutation relations, [Q, P] = Q P − P Q =

i I, λ

(8.9)

but now appearing explicitly as measures of average phase space position and momentum localization. This shows that Q and P are independent of the window function ψ, although their expressions in (8.6) appear in terms of it. Once again, this brings out a point we noted earlier: although the Gabor transform (like the wavelet transform) depends on the window, intrinsic quantities, such as mean values of the parameters of the transform, turn out to be independent of its choice. It is also possible to rewrite the unitary operators U λ in (6.108), realizing an irreducible representation of the Weyl– Heisenberg group, in terms of the operators Q and P [Ali00]. We get U λ (θ, q, p) = eiλ(θ+ p Q−q P) .

(8.10)

This means that Q, P (together with the identity operator I ) also constitute a Hilbert space representation of the generators (6.106) of the Lie algebra of the Weyl–Heisenberg group.

284

Minimal uncertainty and Wigner transforms

Given any self-adjoint operator A on a Hilbert space H, we define its mean value A in the state (i.e., normalized vector) φ ∈ H by A/≡ Aφ = φ|Aφ and its standard deviation A (in the state φ) by A ≡ φ A = A2 − A2 . In standard quantum mechanical lore, it well-known [Coh89,Got66], that given two self-adjoint operators A and B, the product of their standard deviations obeys the uncertainty relation

A. B

1 |[A, B], 2

[A, B] = AB − B A.

(8.11)

The state φ is said to have minimum uncertainty if equality holds in (8.11), which happens if and only if (A − A)φ = −iλo (B − B)φ,

(8.12)

for some λo > 0. By (8.9), for the operators Q and P, the uncertainty relation (8.11) assumes the form:

Q. P

1 , 2λ

(8.13)

which in this case is exactly the same relation as (8.4). Thus, minimal uncertainty is attained for vectors of the type: λ2 φ(x) = π

! 14

q

eiλ(x− 2 ) p e−

λ2 (x−q)2 2

,

(8.14)

for fixed λ = 0 and q, p ∈ R. Of course, these vectors are precisely the generalized gaborettes defined in (6.123), which minimize the product of the standard deviations (8.4) and which give rise to holomorphic Gabor transforms. Thus, referring to our previous discussion, for a signal vector s, the absolute square of its Gabor transform |S(q, p)|2 can be interpreted as giving a sort of “unsharp” joint probability distribution of the position (which now appears as the spectrum of the operator Q) and momentum (the spectrum of P) variables, the uncertainty relation (8.13) forbidding the existence of a sharp joint distribution. However, since the uncertainty is the smallest when the window is a vector of the type (8.13), the corresponding Gabor transform is in a sense optimal. In the following section, we shall use the ideas developed here to construct minimal uncertainty wavelets. We shall proceed group theoretically and isolate two generators of the corresponding representation and use their commutation relation to compute and minimize the uncertainty.

8.2

Minimal uncertainty wavelets We turn our attention now to determining minimal uncertainty wavelets in one and two dimensions. As we know, the relevant groups are G (+) aff and SIM(2) and we have to work

285

8.2 Minimal uncertainty wavelets

with the representations of their Lie algebras on the Hilbert spaces of irreducible representations [24,116]. Consider first the one-dimensional case. We know from Section 6.2.2 that the Lie algebra of G (+) aff is two-dimensional. The two matrix generators of this + (R ) algebra, X 1 and X 2 , computed in (6.77) are represented on the Hilbert space H + (b, a) (see (6.59)–(6.61)) by the generators D and of the irreducible representation U P, of dilation and translation, respectively. They act on Hilbert space vectors in the manner 1 ξ (D s )(ξ ) = −i ([ + ξ ] s )(ξ ), 2 dξ

(P s )(ξ ) = ξ s(ξ ),

(8.15)

+ (R ) are chosen from the appropriate domains of the unbounded operators. where s∈H (Note, this is the “momentum space representation” of these operators.) The operators and P satisfy the commutation relations (see (6.78)) D P] = i P. [ D,

(8.16)

The minimal uncertainty vectors in this case are found to be [246,Pau85,305] the 1-D m (ξ ) = ξ m e−ξ , for ξ 0 (m > 0) and 0, otherwise. Note Cauchy wavelets, namely, ψ that these are also the wavelets which lead to holomorphic wavelet transforms. A similar analysis applies for 2-D wavelets. The Lie algebra is now four-dimensional. The four generators, denoted by P1 , P2 for translations, D for dilations and J for rotations, may be derived explicitly from the transformation (7.35), or its action (2.13) on signals or its equivalent (2.14) in k-space. (Note that we are using the same notation as for the corresponding matrix generators of the Lie algebra (see (7.64)).) Among these four operators, there are four nonzero commutators, namely [D, P1 ] = i P1 ;

[J, P2 ] = −i P1 ;

[D, P2 ] = i P2 ;

[J, P1 ] = i P2 ,

(8.17)

but the first two transform into the last two under a rotation by π/2. More generally, defining Pγ = P1 cos γ + P2 sin γ , we replace both pairs of commutators by the relations: [D, Pγ ] = i Pγ ;

[J, Pγ +π/2 ] = −i Pγ ,

(8.18)

and we look for wavelets which are minimal with respect to this pair. Thus, minimality has to be defined with respect to a fixed direction eγ , and it is impossible to do it for two directions at the same time, for instance, for all four relations (8.17) simultaneously [116]. For fixing ideas, we consider the uncertainty relations for the first pair in (8.17), corresponding to γ = 0:

D. P1

1 |P1 |; 2

J. P2

1 |P1 |. 2

(8.19)

saturates these inequalities iff it satisfies the Then, according to (8.12), a vector ψ following system of equations

286

Minimal uncertainty and Wigner transforms

k) k) = (D + iλ1 P1 )ψ( (D + iλ1 P1 )ψ( k) k) = (J + iλ2 P1 )ψ( (J + iλ2 P2 )ψ(

(λ1 , λ2 > 0) .

(8.20)

This system of partial differential equations may be solved, in polar coordinates k = (ρ, φ), imposing successively five conditions: (i) Integrability of the system requires λ1 = λ2 = λ > 0. (ii) 2π -periodicity in φ implies that P2 = 0 and J = m ∈ Z. (iii) ∈ L 2 , implies that the support of ψ is restricted to a convex Square integrability, ψ cone in the right half-plane. (iv) Admissibility of ψ implies λP1 > 1. (v) Finally, k) be real implies that J = D = 0. imposing the condition that ψ( is minimal with respect to the first pair of the The result is that a real wavelet ψ commutation relations (8.17) iff it vanishes outside some convex cone C in the halfplane k x > 0 and is exponentially decreasing inside: κ −λkx (κ > 0, λ > 0), k ∈ C, k) = c |k| e (8.21) ψ( 0, otherwise. k) |k| κ e−λ kx = c χC (k) ψ(

(κ > 0, λ > 0),

(8.22)

where χC is the characteristic function of C, or a smoothened version thereof. More generally, if one chooses the commutation relations in (8.18), one obtains a similar result, rotated by γ , that is, a wavelet supported in a convex cone Cγ with axis in the direction eγ , and exponentially decreasing inside. We may now impose some degree of regularity (vanishing moments) at the boundary of the cone, by taking an appropriate linear superposition of such minimal wavelets ψ. Thus we obtain finally: C (k) = c χC (k) F(k) e−λkx , ψ

(λ > 0)

(8.23)

is a polynomial in k x , k y , vanishing at the boundaries of the cone C, where F(k) including the origin. Clearly a Cauchy wavelet is of this type but, of course, one uses in practice a narrow support cone C, in order to obtain good directional selectivity, as discussed in Section 3.3.4. Other minimal wavelets may be obtained if one includes commutators with elements of the enveloping algebra, i.e., polynomials in the generators. For instance, if one requires the wavelet to be rotation invariant, one may start from the commutator between D and the Laplacian − = P12 + P22 . Then one finds a whole family of minimal isotropic wavelets, among them all powers of the Laplacian, n , acting on a Gaussian, i.e., the wavelets (3.7) [22]. For n = 2, this gives the 2-D isotropic Mexican hat (3.6) [116]. There exist more general solutions of the minimizing equations, but most of them are not square integrable. We ought to emphasize at this point, that the property of minimality for wavelets is a mathematical one, and it is not clear whether it implies an operational meaning in the same way as was discussed for gaborettes. Cauchy wavelets are linear combinations of

287

8.3 Wigner functions

minimal wavelets, but they are not the most efficient conical wavelets for directional analysis. This is not new: in 1-D too, the Cauchy–Paul wavelet ( [Pau85]) is minimal, but many others are as least as useful in practice, for instance the derivatives of the Gaussian or the Morlet wavelet. As a last remark, it may be interesting to note that a concept closely related to minimality has been developed by Simoncelli et al. [341] under the name of jointly shiftable filters. First, shiftable filters are the natural generalization of steerable filters to other variables than rotations, such as translation or scaling. Then a filter is jointly shiftable in two variables simultaneously if and only if the corresponding operations commute (i.e., “are independent”). Thus strict joint shiftability is impossible for position and spatial frequency; only approximately shiftable filters exist and the optimal ones, that is, those that minimize the “joint aliasing”, are the same as our minimal wavelets.

8.3

Wigner functions In Chapter 1, we mentioned the Wigner–Ville transform as an example of a signal transform which is quadratic (or generally, sesquilinear) in the signal vector(s). We now take a closer look at this transform. Wigner functions have long been used in signal analysis as phase space transforms of, generally, a pair of signal vectors. The original transform, due to Wigner [373], was introduced as a phase space quasi-probability density for computations in atomic physics. We begin by introducing the cross-Wigner function, W λ (ψ, φ|q, p), of two signal vectors ψ, φ ∈ L 2 (R, d x): λ x x λ W (ψ, φ|q, p) = d x ψ(q − ) e−iλ x p φ(q + ), λ > 0, (8.24) 2π R 2 2 adopting a somewhat different notational convention than used in (1.5). The phase space here is R2 , with the variables q and p identified either as time and frequency or position and momentum. The superscript λ will eventually be identified with the parameter labeling the representation of the Weyl–Heisenberg group (see (6.108)). The form of the expression for the cross-Wigner function makes it plausible to look upon it as a mapping of the rank-one operator ρ = |φψ| to a function of the phase space variables q, p. (Recall that the operator ρ acts on an arbitrary vector χ ∈ L 2 (R, d x) by ρχ = ψ|χ φ.) Indeed, the integral on the right-hand side of (8.24) can be manipulated to be brought into the form W λ (ψ, φ|q, p) ≡ W λ (ρ|q, p) λ

= dq dp eiλ(qp − pq ) Tr[e−iλ(Qp −Pq ) ρ], 2π R2

(8.25)

where Q and P are the operators defined in (8.8), and “Tr” denotes the trace of an operator:

288

Minimal uncertainty and Wigner transforms

Trρ =

∞

φk |ρφk ,

i=k 2 {φk }∞ k=1 being an orthonormal basis of L (R, d x). A fairly straightforward computation also shows that the unitary operator appearing within the square brackets on the right-hand side of (8.25) is none other than the representation operator of the Weyl–Heisenberg group, appearing in (6.108). Indeed,

eiλ(Qp−Pq) = U λ (0, q, p), and hence

λ W (ρ|q, p) = 2π λ

R2

dq dp eiλ(qp − pq ) Tr[U λ (0, q , p )∗ ρ],

(8.26)

which clarifies the relationship of the superscript λ, in the definition of the Wigner function, to representations of the Weyl–Heisenberg group. Let us introduce the symplectic Fourier transform, f of a function f ∈ L 2 (R2 , dq d p): λ

f (q, p) = dq dp eiλ(qp − pq ) f (q , p ), 2π R2 with inverse, λ f (q, p) = 2π

R2

f (q , p ). dq dp eiλ(qp − pq )

Clearly, the symplectic Fourier transform is a Hilbert space isometry. Then (assuming for the moment that W λ (ρ|q, p) is an L 2 -function), we may write λ (ρ|q, p). Tr[U λ (0, q, p)∗ ρ] = W

(8.27)

Comparing with (6.125), we see that the symplectic Fourier transform of the crossWigner function W λ (ψ, s|q, p) is just the generalized Gabor transform of the signal s, computed using the window ψ: λ (ψ, s|q, p). S(q, p) = W

(8.28)

However, it ought to be emphasized that, while the Gabor transform S(q, p) is an L 2 transform, on phase space, of the signal s, the cross-Wigner function W λ (ψ, s|q, p) is to be looked upon as an L 2 -transform, on the same phase space, but of the rank-one operator ρ = |sψ|. The original Wigner function [373] was defined with ψ = φ and ρ = |ψψ|. For this case, we shall use the simpler notation W λ (ψ|q, p), and call it the Wigner function for the wave function ψ (or more accurately, for the operator ρ). This function has a number of well-known properties, which make it resemble a probability distribution. However, as is clear from its definition, for a general ψ, its Wigner function W λ (ψ|q, p) is not positive for all q, p. In fact, it is only when ψ is a Gaussian of the type in (6.122), that W λ (ψ|q, p) is everywhere positive. It is for this reason that the Wigner function

289

8.3 Wigner functions

is also called a quasi-probability distribution. Nevertheless, it is still possible to think of it as a phase space transform, which in a certain sense, is the signature of the signal vector. A few properties of the Wigner function W λ (s|q, p) of a signal vector s are now in order. (i) Reality: The Wigner function is real-valued, W λ (s|q, p) = W λ (s|q, p).

(8.29)

This is not generally true of the cross-Wigner function, for which one has the hermiticity condition, W λ (ψ, φ|q, p) = W λ (φ, ψ|q, p).

(8.30)

(ii) Trace condition: dq d p W λ (s|q, p) = Trρ = s2 R2

ρ = |ss|.

(8.31)

This condition is reminiscent of the fact that for a probability distribution, the total probability equals one (as would be the case when s2 = 1). (iii) Marginality: λ 2 d p W (s|q, p) = |s(q)| , dq W λ (s|q, p) = | s( p)|2 . (8.32) R

R

If we think of |s(q)|2 as giving the distribution of the signal in time (or position) and | s(q)|2 its distribution in frequency (or momentum), then the above relations make W λ (s|q, p) formally look like a joint time–frequency (or position–momentum) distribution. Again, the fact that this “joint distribution” is not everywhere positive is a reflection of the uncertainty relations (8.13). (iv) Covariance: An important property, which the Wigner function inherits from the Weyl–Heisenberg group, is reflected in the covariance relation: W λ (U λ (0, q0 , p0 )s|q, p) = W λ (s|q − q0 , p − p0 ).

(8.33)

Identifying the phase space variables with coordinates on a coadjoint orbit of the Weyl–Heisenberg group, the transformation q → q − q0 , p → p − p0 is the coadjoint action discussed in Section 6.3.2. Thus, the above relation expresses symmetry under the natural phase space transformations. We come now to the problem of reconstructibility. In general it is not possible to reconstruct the signal s itself from its Wigner function; however, the operator |ss| can be recovered from it. More generally, the operator |φψ| can be reconstructed from the cross-Wigner function W λ (ψ, φ|q, p). In order to obtain an inversion formula, we go back to the general orthogonality relation (7.12) and see that in the present case it leads to the relation

290

Minimal uncertainty and Wigner transforms

λ 2π

R2

Sψ (q, p) dq d p Sψ (q, p) = ψ|ψ s |s,

(8.34)

where Sψ (q, p) is the generalized Gabor transform of the signal s, computed using the window ψ and Sψ (q, p) that of s computed using ψ. We note, however, that in this case, the above relation holds for any four vectors ψ, ψ , s, and s . Moreover, we immediately get from it the resolution of the identity λ

dq d p |ψ(0,q, (8.35) p) ψ(0,q, p) | = ψ|ψ I, 2π R2 again for arbitrary ψ, ψ ∈ L 2 (R, d x), and where, of course, ψ(0,q, p) = U λ (0, q, p)ψ,

λ

ψ(0,q, p) = U (0, q, p)ψ .

λ (ρ|q, p), the symplectic Fourier transform of its crossLet ρ = |ψφ| and consider W Wigner function. Then, for any ψ ∈ L 2 (R, d x) we have, ! λ λ (ρ|q, p)U λ (0, q, p) ψ dq d p W 2π R2 λ

= dq d p|ψ(0,q, p) φ(0,q, p) |ψ = φ|ψ ψ 2 2π R = ρψ . Since ψ is arbitrary in the above expression, we obtain the reconstruction formula, λ λ (ρ|q, p)U λ (0, q, p), dq d p W (8.36) ρ= 2 2π R with λ (ρ|q, p) = λ W 2π

R2

dq dp eiλ(qp − pq ) W λ (ρ|q , p ).

We can exploit the orthogonality relations (8.34) to extend the definition of the crossWigner function to arbitrary Hilbert–Schmidt operators on L 2 (R, d x). Recall that these operators form a Hilbert space with respect to the scalar product ρ1 |ρ2 2 = Tr[ρ1∗ ρ2 ]. Let us denote this Hilbert space by B2 (R). It is well-known that finite linear combinations of rank-one operators form a dense set in B2 (R). Next note that using (8.28) and the fact that the symplectic Fourier transform is an isometry, we may transform (8.34) into λ dq d p W λ (ρ |q, p) W λ (ρ|q, p) = Tr[ρ ∗ ρ], (8.37) 2 2π R where ρ = |sψ|,

ρ = |s ψ |.

The relation (8.37) remains valid if we replace ρ, ρ by finite linear combinations of rank-one operators, meaning that the map given by the integral on the right-hand

291

8.4 Wigner functions for the wavelet groups

side of (8.25), associating an operator ρ ∈ B2 (R), taken from this dense set, to the function W λ (ρ|q, p), is linear and an isometry (up to a factor of λ/2π). Using this fact we can extend the map to a unitary transformation between the Hilbert spaces B2 (R) and L 2 (R, dq d p). Thus, to any Hilbert–Schmidt operator ρ, we can associate a general Wigner function W λ (ρ|q, p). However, the explicit expression for this function is given by the integral in (8.25) only for operators with a well-defined trace. Otherwise, it has to be obtained as an L 2 -limit of such functions. Conversely, every function f in L 2 (R, dq d p) is the general Wigner function of a unique Hilbert–Schmidt operator. If f is also L 1 -integrable, then this operator is given by (see (8.36)) λ ρ= dq d p f (q, p)U λ (0, q, p), (8.38) 2π R2 where again f is the symplectic Fourier transform of f . If f is not L 1 -integrable, then the corresponding Hilbert–Schmidt operator is obtained as a Hilbert space limit (in B2 (R)) of operators obtained using (8.38). The Wigner function of a general Hilbert– Schmidt operator probably does not have a natural meaning in signal analysis. It is only Wigner functions of rank-one operators that have been used directly, as signature functions of signals. In quantum mechanics, however, a Hilbert–Schmidt operator, which is trace-class and of unit trace, represents a mixed state and its Wigner function again has the interpretation of a phase space quasi-probability distribution for this state. We end this section by repeating what we said already in Chapter 1, that the Wigner function W λ (s|q, p) of a signal is in a sense more intrinsic than its Gabor transform, since the former does not depend on an arbitrarily chosen window. On the other hand, the reconstruction formula (8.36) only gives back the operator |ss|, and not the function s itself. Still, both are transforms carrying information about the signal in terms of phase space variables. A second point to be borne in mind is that while the cross-Wigner function is sesquilinear (see (8.30)), and the Wigner function is quadratic, when looked upon as a transform on signal vectors, it is in fact linear when looked upon as a transform on the space of Hilbert–Schmidt operators.

8.4

Wigner functions for the wavelet groups In view of the fact that the Wigner function has proved itself to be an extremely useful tool, both in signal analysis and in atomic and quantum optical computations [63,64,376], it makes sense to look for similar signal transforms related to groups other than the Weyl–Heisenberg group. In particular, one would like to construct such transforms for the one- and two-dimensional wavelet groups. These could then provide one with alternatives to the wavelet transforms discussed in the previous two chapters. We now proceed to obtain such transforms, with the proviso that these generalized Wigner functions should also be phase space functions (i.e., functions defined

292

Minimal uncertainty and Wigner transforms

on coadjoint orbits of the relevant groups). Also, we would like to preserve as many of the properties (8.29)–(8.33) as possible. A general procedure for constructing such functions has been proposed in [7– 9,248]. (See also [72– 74].) Again, the objective is to find a one-to-one linear correspondence between Hilbert–Schmidt operators on the carrier space of an irreducible representation of the group and L 2 -functions on phase space. As with wavelet transforms, the square integrability of the group representations, used in the construction, will turn out to be of crucial importance and as before, the orthogonality relations (7.12) will guarantee a reconstructibility condition. We shall keep the discussion here mainly descriptive, without venturing into too many mathematical details. Just as while constructing two- (or higher) dimensional wavelet transforms, we exploited the symmetry groups of signals, so also for constructing generalized Wigner transforms we look at semi-direct product groups of a particular type. Suppose that our signal vectors are elements of L 2 (Rn , d n x), where for the rest of this discussion n = 1 or 2. Following our discussion in Section 7.1.1, we assume that the allowable transformation symmetries of our signals are of the following types: r Translations: s( x ) → s( x + b), r

Dilations: s( x ) → s(a x),

r

b ∈ Rn .

a > 0.

Matrix transformations of Rn : s( x ) → s(h x ), where h is an n × n nonsingular matrix. We shall assume that the set of all such admissible matrices form a group H and, furthermore, that the following technical condition is satisfied: there exists a vector x ∈ Rn such that the set, Ox = {y = ahT x | a > 0 , h ∈ H },

is true only when h is the identity matrix. is open in Rn and y = hy , for any y = 0, In this case Ox is called an open free orbit. (This condition is satisfied by all the wavelet groups used in the current literature, including of course, the affine group, the 2-D wavelets group and the affine Poincar´e group, studied in the previous two chapters.) Thus, we are assuming that the full symmetry group of allowed transformations on our signals is G = Rn (R+ ∗ × H ). Elements of this group can be conveniently represented by the matrices ah b a, h) = , a > 0 , h ∈ H, b ∈ Rn. g ≡ (b, (8.39) 0 T 1

293

8.4 Wigner functions for the wavelet groups

A point x ∈ Rn undergoes the transformation x → ah x + b under its action. This group is nonunimodular and its left invariant (Haar) measure is, a, h) = dµG (b,

1 a n+1 det h

d n b da dµ H (h),

(8.40)

where d n b is the Lebesgue measure on Rn and dµ H (h) the left invariant measure of the group H . The Hilbert space of signals L 2 (Rn , d n x) then carries a natural unitary repre a, h), reflecting the transformation sentation of the group G by unitary operators U (b, properties of the signals under the group action 1 −1 −1 a, h)s ( U (b, x) = h ( x − b)). (8.41) 1 s(a n (a det h) 2 n , d n k), the corresponding operators act in the manner In the Fourier space L 2 (R (b, a, h) = (a n det h) 12 e−i b· k s(a −1 hT k). (8.42) U s (k) (b, a, h) is in general not irreducible. However, if O is an open The representation U n 2 , and L (O, d n k) n , d n k), the subspace of L 2 (R consisting of functions free orbit in R supported on this orbit, then the representation U (b, a, h), restricted to this subspace, is irreducible and square-integrable; hence it is appropriate for constructing generalized wavelet transforms. It can also be shown [69] that all irreducible subrepresentations of (b, a, h) are of this type and that there is only a finite number of them (the connected U affine group, G + aff , has two, while the SIM(2) group has only one, etc.). Generalized Wigner functions, which bear a strong resemblance to the ones defined in (8.24) and (8.26), can also be constructed now using these same irreducible representations. The phase space on which the Wigner functions are defined is O∗ = Rn × O, which can be identified with a coadjoint orbit of the group G and is associated to an irreducible representation. The dimension of the phase space is 2n. The exact construction relies rather heavily on the properties of the Lie algebra g of G and its dual g∗ , which we prefer to omit, displaying and discussing only the final expressions for the two wavelet groups G + aff and SIM(2). For details, the reader may refer to [7,8,248]. Let X 1 , X 2 , . . . , X 2n be a (vector space) basis of the Lie algebra and X 1 ∗ , X 2 ∗ , . . . , X 2n ∗ the dual basis for the vector space g∗ . The basis elements are (n + 1) × (n + 1) matrices and we assume that the last n elements, X n+1 , X n+2 , . . . , X 2n correspond to translations of Rn (i.e., they are the generators of the one-parameter translation subgroups). A general element in the Lie algebra is a linear combination

2n i X = i=1 x X i = x · X , in an obvious notation. Similarly, a general element in the

2n dual is a linear combination, X ∗ = i=1 γi X i ∗ = γ · X ∗ . The identification of the 2n coadjoint orbit O∗ with Rn × O ⊂ R2n is then done with respect to the basis {X i ∗ }i=1 T and we denote a point in the orbit by a column vector γ = (γ1 , γ2 , . . . , γ2n ) . Let U (g), g ∈ G, be an irreducible (square-integrable) representation of G on the Hilbert space H. Denote by C the Duflo–Moore operator defined in (7.5) (see also (6.16) and

294

Minimal uncertainty and Wigner transforms

(7.42)) and by X i the self-adjoint operator which represents X i on the Hilbert space H will denote a vector operator with components, (via the representation U (g)). Also, X X i . The generalized Wigner function, corresponding to the irreducible representation U (g) is then explicitly given, for any Hilbert–Schmidt operator ρ on H, such that the operator ρC −1 is of trace class, by the expression 1 1 2n i( X −γ )· x −1 W (ρ | γ ) = d x Tr e ρC (8.43) [σ (γ ) m( x )] 2 , n (2π ) N0 0 is an appropriate subset of R2n , depending on certain properties of the Lie where N 0 = R2n . The function m( algebra. In the two examples given below, N x ) is needed to transform the Haar measure on the group to a measure on the Lie algebra parameters xi , while the function σ (γ ) is a density which converts the invariant measure d(γ ) on the phase space O∗ to the Lebesgue measure on R2n . The orthogonality relations (7.6) play a crucial role in the derivation of the above expression for the Wigner function, which can in fact be extended (by taking limits) to all Hilbert–Schmidt operators on H. In other words, the generalized Wigner function is again a transform of a Hilbert–Schmidt operator (on the Hilbert space of signals) to a function on phase space. If ρ = |φψ|, we obtain the cross-Wigner function, 1 1 W (ψ, φ | γ ) = d 2n x C −1 ψ | ei( X −γ )·x φ[σ (γ ) m( x )] 2 , (8.44) n (2π ) N0 provided the vector ψ is chosen from the domain of the operator C −1 , implying thereby an admissibility condition on the operator ρ. The inversion formula is then, 1 1 ρ= d(γ ) d 2n x e−i( X −γ )·x C −1 W (ρ|γ ) [σ (γ ) m( x )] 2 . (8.45) n (2π) O∗ 0 N

8.4.1

Wigner functions for the affine group + (b, a) of the connected affine group G + We take the irreducible representation U aff , defined in Section 6.2 (see (6.61)–(6.62)). The Hilbert space of this representation is + (R ) of L 2 -functions supported on the positive real axis (see (6.62)). The open free H ∗ orbit is O = R∗+ and the phase space is the coadjoint orbit O+ = R × R∗+ , identified in (6.89). This orbit is equipped with the measure (6.92), which is invariant under + (R ), φ ∈ H the coadjoint action spelled out in (6.87). If (γ1 , γ2 ) ∈ R × R∗+ , and ψ, inserting into (8.44) the cross-Wigner function becomes, ∞ x x 2 γ2 e−iγ1 x γ2 e− 2 γ2 e φ|γ 1 , γ2 ) = 1 1 φ dx ψ , (8.46) W (ψ, sinch x2 sinch x2 sinch x2 (2π ) 2 −∞ where (see (7.72)), sinch u =

sinh u . u

295

8.4 Wigner functions for the wavelet groups

For the purposes of signal analysis, the parameters γ1 and γ2 would be identified with time and frequency, respectively. The following facts should be noted about this function. r W is sesquilinear in φ, ψ and indeed, it is again a transform of the rank-one operator ψ| to a phase space function. Moreover, the support of this function is ρ = |φ contained entirely in the coadjoint orbit O∗ . r Writing W (ψ|γ 1 , γ2 ) for the case when φ = ψ, this function is seen to be real, although in general not everywhere positive, and again, the trace condition is satisfied: ∞ ∞ 1 dγ2 1 , γ2 ) = Tr [ρ], ψ|. dγ W (ψ|γ ρ = |ψ (8.47) 1 1 γ2 2 (2π) −∞ 0 r

As anticipated, the covariance condition assumes the form: + (b, a)ψ| γ ) = W (ψ|M * (b, a)−1 γ ) = W (ψ|γ 1 − bγ2 , aγ2 ), W (U

(8.48)

with M * (b, a) being the matrix of the coadjoint action of the group on the phase space, obtained in (6.85)–(6.87). r The definition of the Wigner function can be extended, using the trace condition, to + (b, a) all Hilbert–Schmidt operators ρ, on the Hilbert space of the representation U and then, the orthogonality relation ∞ ∞ 1 dγ2 dγ W (ρ1 |γ1 , γ2 ) W (ρ2 |γ1 , γ2 ) = Tr [ρ1∗ ρ2 ], (8.49) 1 1 γ 2 2 (2π) −∞ 0 holds. From this, or using (8.45) it is easy to write down a reconstruction formula for ρ, given its Wigner function. (For details, the reader is referred to [7].) r Unlike the original Wigner function (8.24), only one of the two marginality conditions (8.32) is satisfied in this case. One gets, ∞ 1 dγ1 1 , γ2 ) = |ψ(γ 2 )|2 . W (ψ|γ (8.50) 1 (2π) 2 −∞ γ2 The nonexistence of a simple form for the second marginality condition is related to the fact that there does not seem to be a “natural” choice of coordinates on the phase space, in terms of which the condition could be stated.

8.4.2

Wigner functions for the similitude group To construct Wigner functions for the SIM(2) group, we use its only irreducible a, θ), carried by the Hilbert space L 2 (R2 , d 2 x) (see (2.13)) or representation U (b, (b, a, θ ), carried by the Hilbert space equivalently, its Fourier-transformed version U 2 2 2 L (R , d k) (see (2.14)). The Duflo–Moore operator is the one given in (7.42) and we adhere to the notation and terminology introduced in Sections 7.3.1 and 7.3.2. The computation of the cross-Wigner function follows in several steps, which we now summarize.

296

Minimal uncertainty and Wigner transforms

(i) Taking a ψ in the domain of C −1 and an arbitrary vector φ in the Hilbert space we may write

C −1 ψ|ei( X −γ )·x φ = e−i[α·α

∗

β∗ ] +β·

U (e X )C −1 ψ|φ,

(8.51)

T . (Recall that in the above equation, e X indicates an where we have set x = ( α , β) element of the group SIM(2), while e−i X ·x is a unitary operator giving its Hilbert space representation.) (ii) Using the explicit form of the representation, the expression for the Duflo–Moore operator and equations (7.70) and (7.82), we obtain, = (U (e X )C −1 ψ)(k)

e2λ −i[F(s(α))β]· k |k| e ψ(eλ r−θ k). 2π

(8.52)

(iii) Next we note that, if V is a 2 × 2 real matrix of the type in (7.67) and u a 2-vector, 1 then |V u| = [detV ] 2 | u |. Thus, ! 12 λ2 + θ 2 −1 α )) k| = |k|. (8.53) |F(s( 2eλ (cosh λ − cos θ ) Using this fact to effect a change of variables in the integration involved in the scalar product, we obtain ! 32 e2λ (λ2 + θ 2 ) X −1 U (e )C φ|ψ = d 2 k ei β·k |k| 2 2π 2eλ (cosh λ − cos θ ) R ! T T /2 −s( s ( α ) e α) /2 e k ψ k . (8.54) ×φ sinch(s( sinch(s( α )T /2) α )T /2) (iv) Inserting the above expression into the formula for the Wigner function in (8.44) and using the expressions (see [8] for details of derivation), cosh λ − cos θ , λ2 + θ 2 σ (γ ) = σ ( α ∗ , β∗ ) = |β∗ |2 = (β1∗ )2 + (β2∗ )2 ,

= 2e−λ m( x ) = m( α , β)

leads to W (ψ, φ| α ∗ , β∗ ) =

1 2(2π )3

d 2 k

R2

∞

−∞

2π

dλ

(8.55)

dϑ 0

R2

d 2 β

λ2 + ϑ 2 cosh λ − cos ϑ ! T −s( α )T /2 es(α) /2 e k φ k , ×ψ sinch(s( α )T /2) sinch(s( α )T /2) λ ϑ λ s( α )T = , α = . (8.56) −ϑ λ ϑ e

β∗ )·β −i α · α i(k− ∗

e

|β∗ | |k|

297

8.4 Wigner functions for the wavelet groups

The integration over β represents a delta measure in k − β∗ . Using this fact to perform the k-integration, we obtain finally ∞ 2π λ2 + ϑ 2 1 ∗ ∗ W (ψ, φ| α ∗ , β∗ ) = dλ dϑ e−i(α1 λ+α2 ϑ) |β∗ |2 4π −∞ cosh λ − cos ϑ 0 ! T e−s(α) /2 es(α)T /2 ∗ ∗ ×ψ β φ β . sinch(s( α )T /2) sinch(s( α )T /2) (8.57) This then is the final expression of the cross-Wigner function for SIM(2) at a phase space point, written in the ( α ∗ , β∗ ) coordinates. More compactly, using the matrices defined in (7.79)–(7.80), we may write |β∗ |2 ∗ ∗ ∗ W (ψ, φ| α ,β ) = d 2 α e−i α ·α ψ(eλ r−ϑ F(s( α )T )−1 β∗ ) 2π R2 α )T )]−1 φ(F(s( α )T )−1 β∗ ). (8.58) × eλ [detF(s( Also, it is straightforward to express this function in terms of the other coordinates discussed at the end of Section 7.3.2. For example, in the Darboux (or canonical) coordinates, introduced in (7.104), the cross-Wigner function assumes the form: ρ2 W (ψ, φ| q , p) = d 2 α e−i[ ρ rϕ σ3 q ]·α ψ(eλ r−ϑ F(s( α )T )−1 p) 2π R2 α )T )]−1 φ(F(s( α )T )−1 p), × eλ [detF(s(

(8.59)

where we have written p = (ρ, ϕ) in polar coordinates while, as before, α has the Cartesian coordinates λ, ϑ. By its very construction, for each pair of wave functions ψ, φ, the cross-Wigner function W is a function on phase space. The exact variables of phase space may however be chosen to suit our convenience. Moreover, it is clear, that even when φ = ψ, the Wigner function may not be an everywhere positive function. The covariance property, in the present context, can be expressed as W (U (g0 )ψ, U (g0 )φ|γ ) = W (ψ, φ|M * (g0 )−1 γ ), g0 = (a0 , θ0 , b0 ).

(8.60)

Using (7.85), this can be written more explicitly as W (U (a0 , θ0 , b0 )ψ, U (a0 , θ0 , b0 )φ | α ∗ , β∗ ) = W (ψ, φ| α ∗ , β∗ ) α ∗ = α ∗ − s(b0 )T β∗ , β∗ = a0 r−θ0 β∗ .

(8.61)

The orthogonality condition is reflected in the cross-Wigner function through the following relation: d 2 α ∗ d 2 β∗ W (ψ1 , φ1 | α ∗ , β∗ ) W (ψ2 , φ2 | α ∗ , β∗ ) = ψ2 | ψ1 φ1 | φ2 . (8.62) ∗ 2 ∗ |β | O

298

Minimal uncertainty and Wigner transforms

Although this relation is guaranteed by the very way the Wigner function is constructed, it is still instructive to demonstrate it directly, as we now proceed to do. Denoting the left-hand member of (8.62) by I (ψ1 , φ1 , ψ2 , φ2 ) and substituting from (8.58) we get ∗

I (ψ1 , φ1 , ψ2 , φ2 ) = d 2 α ∗ d 2 β∗ |β∗ |2 d 2 α d 2 α ei α ·(α −α) O∗

R 2 ×R 2

× ψ1 (eλ r−ϑ F(s( α )T )−1 β∗ ) φ1 (F(s( α )T )−1 β∗ )

× e(λ +λ) [detF(s( α ))T ]−1 [detF(s( α ))T ]−1

! × ψ2 (eλ r−ϑ F(s( α )T )−1 β∗ ) φ2 (F(s( α )T )−1 β∗ ) .

(8.63)

The integral in α ∗ represents a delta measure in α − α which allows us to perform the α -integration. Next changing variables, β∗ → F(s( α )T )−1 β∗ , we obtain I (ψ1 , φ1 , ψ2 , φ2 ) = d 2 α ∗ d 2 β φ1 (β∗ ) φ2 (β∗ ) |β∗ |2 e2λ O∗

λ

× ψ1 (e r−ϑ

! ∗ λ ∗ β ) ψ2 (e r−ϑ β ) .

(8.64)

Noting that λ, ϑ are the components of α , we change variables as α → eλ r−ϑ β∗ to rewrite the α ∗ -integration. The determinant of this transformation is precisely |β∗ |2 e2λ , implying that I (ψ1 , φ1 , ψ2 , φ2 ) = ψ2 |ψ1 φ1 |φ2 ,

(8.65)

as asserted in (8.62). Finally, we again obtain in this case only one marginality condition for the Wigner function W (ψ| α ∗ , β∗ ) (obtained by taking φ = ψ in the cross-Wigner function): d 2 α ∗ 1 W (ψ | α ∗ , β∗ ) = |ψ(β∗ )|2 . (8.66) ∗ |2 2π R2 |β Note also the trace condition, d 2 α ∗ d 2 β∗ 1 W (ψ | α ∗ , β∗ ) = |ψ|2 . 2π O∗ β∗ 2

8.4.3

(8.67)

The Wigner function and the wavelet transform As already noted, there is a close connection between the Wigner function and the wavelet transform arising from a general Lie group. In this section we explicitly demonstrate this relationship for the SIM(2) group. The starting point is the resolution of the identity (7.45): 1 U (a, θ, b)|ψψ|U ∗ = I, dµ(a, θ, b) (a, θ, b) cψ = Cψ2 . (8.68) cψ SIM(2)

299

8.4 Wigner functions for the wavelet groups

Let s be a signal in the Fourier domain, i.e., an element of the Hilbert space L 2 (R2∗ , d 2 k). Its wavelet transform, corresponding to the mother wavelet ψ, is a, θ) = ψa,θ,b | s = U (a, θ, b)ψ | s. Sψ (b,

(8.69)

Using the group parameters as the phase space variables for the Wigner function (see (7.96)), we easily deduce √ 2π 2 i a, θ ) = W (Cψ, s|b, dλ dϑ d 2 β e− a [α·(rθ σ3 b) + β·eθ ] 2 (2π )2 a R 0 R ! 1 (cosh λ − cos ϑ) 2 λ × (8.70) Sψ (F(s( α ))β, e , ϑ) , | α| T T with eθ = cos θ, sin θ , α = ϑ, λ and F(s( α )) given, as before, by (7.71). Thus, the relationship between the wavelet transform and the Wigner function is an integral transform on the Hilbert space L 2 (SIM(2), dµ) of all square integrable (with respect to the Haar measure) functions of the group, and it is then straightforward to check that a, θ ) |Sψ (b, a, θ )|2 = a, θ) |W (Cψ, s | b, a, θ)|2 . (8.71) dµ(b, dµ(b, SIM(2)

SIM(2)

If the signal s is a Hilbert space vector which is in the domain of the inverse operator C −1 , then taking s = Cψ we also obtain √ 2π 2 i a, θ ) = W (s|b, dλ dϑ d 2 β e− a [α·(rθ σ3 b) + β·eθ ] 2 (2π ) a R 0 R2 ! 1 (cosh λ − cos ϑ) 2 λ e , ϑ) . SC −1 s (F(s( × α ))β, (8.72) | α| The above results show that any information about the signal which can be obtained using the wavelet transform, can also be derived using the Wigner function. The use of one rather than the other is therefore more a matter of practical convenience or computational ease, in any given situation.

9

Higher-dimensional wavelets

In the previous chapters, we have thoroughly discussed the 2-D CWT and some of its applications. Then we have made the connection with the group theoretical origins of the method, thus establishing a general framework, based on the coherent state formalism. In the present chapter, we will apply the same technique to a number of different situations involving higher dimensions: wavelets in 3-D space R3 , wavelets in Rn (n > 3), and wavelets on the 2-sphere S 2 . Then, in the next chapter, we will treat time-dependent wavelets, that is, wavelets on space–time, designed for motion analysis. In all cases, the technique is the same. First one identifies the manifold on which the signals are defined and the appropriate group of transformations acting on the latter. Next one chooses a square integrable representation of that group, possibly modulo some subgroup. Then one constructs wavelets as admissible vectors and derives the corresponding wavelet transform.

9.1

Three-dimensional wavelets Some physical phenomena are intrinsically multiscale and three-dimensional. Typical examples may be found in fluid dynamics, for instance the appearance of coherent structures in turbulent flows, or the disentangling of a wave train in (mostly underwater) acoustics, as discussed above. In such cases, a 3-D wavelet analysis is clearly more adequate and likely to yield a deeper understanding [56]. The same is true for many problems in astrophysics, such as galaxy/void counting or grouping, or cluster structure analysis. Hence we will also describe briefly the 3-D CWT, following the general pattern of the previous section.

9.1.1

Constructing 3-D wavelets Given a 3-D signal s ∈ L 2 (R3 , d 3 x), with finite energy, one may act on it by translation, dilation and rotation: a, ,)s ( sb,a,, ( x ) ≡ U (b, x ) = a −3/2 s(a −1 ,−1 ( x − b)), (9.1)

300

301

9.1 Three-dimensional wavelets

where b ∈ R3 , a > 0 and , ∈ SO(3). The 3 × 3 rotation matrix , ∈ SO(3) may be parametrized, for instance, in terms of three Euler angles. These three operations generate the 3-D Euclidean group with dilations, i.e., the similitude group of R3 , SIM(3) = R3 (R+ ∗ × SO(3)). Then (9.1) is a unitary representation of SIM(3) in L 2 (R3 , d 3 x), which is irreducible and square integrable, hence it generates a CWT exactly as before. Wavelets are taken in L 2 (R3 , d 3 x) and the admissibility condition is now d 3 k 2 |ψ(k)| < ∞. (9.2) 3 R3 |k| As in 2-D, a necessary, and almost sufficient, condition for admissibility is the familiar zero mean condition: d 3 x ψ( x ) = 0. (9.3) R3

The two standard wavelets have a 3-D realization. r The 3-D Mexican hat is given by ψH ( x ) = (3 − |A x|2 ) exp(− 12 |A x|2 ).

(9.4)

where A = diag[$1 −1/2 , $2 −1/2 , 1], $1 1, $2 1, is a 3 × 3 anisotropy matrix. We distinguish three cases: (1) If $1 = $2 = 1, one has the fully anisotropic 3-D Mexican hat (the stability subgroup Hψ is trivial); (2) If $1 = $2 = 1, one has the isotropic, SO(3)-invariant, 3-D Mexican hat (Hψ = SO(3)); (3) If $1 = $2 ≡ $ = 1, the wavelet is axisymmetric, i.e., SO(2)-invariant, but not isotropic (Hψ = SO(2)). Thus, in this case, wavelets are coherent states of Gilmore–Perelomov type, whose parameter space is indeed the quotient SO(3)/SO(2) S 2 (see the discussion in Section 7.1.5). r

The 3-D Morlet wavelet is given by ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) + corr.,

(9.5)

where A is the same 3 × 3 anisotropy matrix as in the first example. Here again, for $1 = $2 ≡ $ = 1 and ko along the z-axis, the wavelet ψ is invariant under SO(2) and we obtain coherent states of Gilmore–Perelomov type. r A 3-D Cauchy wavelet is defined by a straightforward generalization of the 2-D case, as follows. Consider the convex simplicial (or pyramidal) cone C(α, β, γ ) defined by the three unit vectors eα , eβ , eγ , the angle between any two of them being smaller ˜ γ˜ ), where eα˜ = eβ ∧ eγ than π. The dual cone is also simplicial, namely C = C(α, ˜ β, is orthogonal to the β-γ face, etc. With these notations, given a vector η ∈ C and l, m, n ∈ N∗ , we define a 3-D Cauchy wavelet in spatial frequency space as: (k · eα˜ )l (k · eβ˜ )m (k · eγ˜ )n e−k·η , k ∈ C(α, β, γ ), (C,η) (k) (9.6) = ψ lmn 0, otherwise.

302

Higher-dimensional wavelets

As in the 2-D case, the factors (k · eα˜ )l , etc. represent vanishing moments on the faces of the cone, and thus determine the regularity of the wavelet in k-space. Here too the expression for the 3-D wavelet in position space may be obtained explicitly: l+m+n vol [eα , eβ , eγ ] i l+m+n+3 (C,η) ψlmn ( x) = , (9.7) l! m! n! · det A · 2π (z · eα )l+1 (z · eβ )m+1 (z · eγ )n+1 where A is the matrix that transforms the unit vectors e1 , e2 , e3 into the triple eα , eβ , eγ , vol [·, ·, ·] denotes the volume of the parallelepiped generated by the three vectors, and we have written z = x + i η. Note that the direct calculation, following the pattern of the 2-D case (Section 3.3.4) yields for the numerator in (9.7) the expression (eα˜ · eα )l (eβ˜ · eβ )m (eγ˜ · eγ )n , but then one has eα˜ · eα = eβ ∧ eγ · eα = eβ · eγ ∧ eα = eβ · eβ˜ = eγ · eγ˜ = vol [eα , eβ , eγ ], which proves (9.7). From the expressions (9.6) and (9.7), one may then obtain other 3-D Cauchy wavelets, for instance, one supported in a circular cone. Take a circular convex cone, aligned on the positive k z -axis, with total opening angle 2θo (0 < θo < π/2). In spherical polar coordinates k = (|k|, θ, φ), the interior of the cone is simply C(θo ) = {k ∈ R3 | θ θo }. Then an axisymmetric 3-D Cauchy wavelet supported in this cone may be defined, for instance, by |k|l (tan2 θo − tan2 θ)m e−|k| cos θ , 0 θ θo ; (θo ) ψm (k) = (9.8) 0, otherwise. Again m ∈ N defines the number of vanishing moments on the surface of the cone, that is, the regularity of the wavelet. For θo very small, this wavelet lives inside a narrow pencil: it clearly evokes the beam of a searchlight – a vivid illustration of the wavelet as a directional probe! If we note that the expression on the right-hand side of (9.8) may be written as

|k|l (tan2 θo k z2 − k x2 + k 2y )m e−kz , we see that all these wavelets are built on the same = 0 is the equation of the cone. m e−kz , where F(k) model, namely F(k)

9.1.2

The 3-D continuous wavelet transform Then, given a signal s ∈ L 2 (R3 ), its CWT with respect to the admissible wavelet ψ is given as a, ,) = a −3/2 s( S(b, d 3 x ψ(a −1 ,−1 ( x − b)) x) . (9.9) R3

As compared with (2.19), the only differences are in the normalization factors and the rotation matrices. Since the structure of the formulas is the same as before, so are the

303

9.1 Three-dimensional wavelets

(a)

(b)

Fig. 9.1. The 3-D cube and its Fourier transform.

interpretation and the consequences (local filtering, reproducing kernel, reconstruction formula, etc.). Thus the CWT (9.9) may be interpreted as a mathematical camera with magnification 1/a, position b and directional selectivity given, in the axisymmetric case, a, ,) by the rotation parameters ζ ≡ (θ, ϕ). As for the visualization, the full CWT S(b, is a function of seven variables. However, if the wavelet ψ is chosen axisymmetric, i.e., SO(2)-invariant, S depends on six variables only, b ∈ R3 , a > 0, and ζ ∈ S 2 SO(3)/SO(2), the unit sphere in R3 . In this case again, (a −1 , ζ ) may be interpreted as polar coordinates in spatial frequency space. This is in fact true in any number of dimensions. It follows that, here too, there are two natural representations for the visualization of the WT: the position representation (a, ζ fixed) and the scale-orientation (or spatial frequency) representation (b fixed). Of course, there are many other possible representations that may be useful. As an example, we present in Figure 9.1 (the characteristic function of) a 3-D cube and its Fourier transform (note the occurrence of sinc functions among all three coordinate axes), then in Figure 9.2 the wavelet transform of the cube signal, with a Morlet wavelet, at four successively finer scales, a = 1, 0.5, 0.25, 0.125. As in 2-D, the net result, for a sufficiently fine scale, is the contour of the cube. The latter has become totally transparent, only the edges survive! A similar, slightly more complicated example is that of a cube with a small box removed, see Figure 9.3. Here one faces the well-known visual ambiguity: depending of the angle of view, the small cube appears either as removed (concave) or added (convex). As shown in Figure 9.4, the ambiguity is resolved with the WT, the part that has been removed yields a negative WT, whereas an added, convex, portion would yield a positive contribution to the WT. This is true, of course, if one uses a real wavelet (here the 3-D Mexican hat) and plots the WT itself, not its modulus. We recover exactly the 2-D situation, discussed in Section 4.1.2.

304

Higher-dimensional wavelets

(a)

(c)

(b)

(d)

Fig. 9.2. Wavelet transform of the 3-D cube at four successively finer scales.

(a) Fig. 9.3. The 3-D cube minus a box and its Fourier transform.

(b)

305

9.1 Three-dimensional wavelets

(a)

(c)

(b)

(d)

Fig. 9.4. Wavelet transform of the 3-D cube minus a box at four successively finer scales (a = 1,

0.5, 0.25, 0.125).

A practical application of this 3-D CWT is the detection of 3-D objects in a cluttered medium. We consider a scene with 3-D objects (targets) immersed in a cluttered medium, modeled by the signal: s( x) =

L

sl ( x ) + n( x ),

(9.10)

l=1

where sl ( x ) denotes the density of the target l, and n( x ) the density of the medium. Since the density of the targets is very different from that of the medium, there will be a high density gradient at the boundary between the objects and the medium. In this a, ζ ) may be used for extracting the 3-D objects and situation, the wavelet transform S(b, determining their characteristics, position (range and orientation) and spatial frequency. This is, of course, nothing but a 3-D version of the ATR problem discussed already in Section 4.2 (see also [16]). In particular, 3-D directional wavelets (e.g. Morlet or Cauchy) will behave as their 2-D counterparts. If a 3-D image s( x ) contains elongated

306

Higher-dimensional wavelets

a, ζ ) shows “sausage” images for all the objects which objects, a visualization of S(b, have the same direction as the wavelet, and small spheres at the tips of all the objects which are misaligned. With an appropriate thresholding, we may filter out all the objects which are not in the desired direction. Another domain of application is in computational fluid dynamics. For instance, the technique has been used in [56] for analyzing two types of simulated 3-D flows: (i) detection of elongated vortex structures in homogeneous 3-D turbulence, and (ii) interaction of a shock wave with a 3-D supersonic mixing layer between two parallel streams. In both situations, the use of scale as a fundamental variable offers distinct advantages over traditional methods based on sophisticated graphic tools. In the first case, the selection of structures is based on characteristic scales rather than on vorticity values. For the shear/shock interaction, the simultaneous detection of main steady flow features with unsteady multiscale mixing features is a practical way to the study of compressibility controlled mixing phenomena. Clearly this opens interesting perspectives for wavelet analysis in fluid dynamics. As a final remark, we may note that the 3-D discrete WT has also been used successfully in several applications. One of them is the analysis of the substructure of galaxy clusters [300]. Another one is the lossless compression of 3-D images produced by computer tomography or radar scanning [81]. The lossless character of the transform is based, as in 1-D or 2-D, on the use of an integer WT (see Section 2.5.2.4).

9.1.3

Extension to higher dimensions From the analysis of the previous section, it should be clear that the extension to higher dimensions is straightforward – although it has probably an academic interest only (except perhaps in quantum mechanics, since an N -particle wave function belongs to L 2 (R3N , d 3N x)). Given a finite energy signal s ∈ L 2 (Rn , d n x), the transformation group to consider is again the similitude group, now SIM(n) = Rn (R+ ∗ × SO(n)). This group has a (unique) natural unitary irreducible representation in L 2 (Rn , d n x), given by sb,a,, ( x ) ≡ U ( b, a, ,)s ( x ) = a −n/2 s(a −1 ,−1 ( x − b)), (9.11) where b ∈ Rn is a translation, a > 0 is a global dilation and , ∈ SO(n) is a rotation. This representation is square integrable, with admissibility condition d n k 2 |ψ(k)| < ∞ , (9.12) n Rn |k| reducing as usual to the necessary condition Rn d n x ψ( x ) = 0. The n-D wavelet transform is, as before, s( a, ,) = a −n/2 d n x ψ(a −1 ,−1 ( x − b)) x) . (9.13) S(b, Rn

307

9.1 Three-dimensional wavelets

Standard wavelets are the same, namely, r The n-D Mexican hat: ψH ( x ) = (n − |A x|2 ) exp(− 12 |A x|2 ). r

The n-D Morlet wavelet: ψM ( x ) = exp(i ko · x) exp(− 12 |A x|2 ) + corr.

Once again, the anisotropy matrix A = diag[$1 −1/2 , $2 −1/2 , . . . , $n−1 −1/2 , 1], $1 1, $2 1, . . . , $n−1 1, leads to various situations, depending on the number of different values taken by the parameters $ j . The interesting case is, of course, the axisymmetric or SO(n − 1)-invariant case, where all the $ j coincide and $ j = $ = 1. n−1 Then, the parameter space reduces to Rn × R+ Rn × Rn∗ and the WT reads ∗ ×S a, ζ ), ζ ∈ S n−1 . As before, (a −1 , ζ ) may be interpreted as polar coordinates in S(b, spatial frequency space, and thus the n-D CWT yields a phase space representation of signals. The rest is as before. Cauchy wavelets also extend to Rn , as follows. We start with the n-simplex C(e1 , e2 , . . . , en ), generated by the unit vectors e1 , e2 , . . . , en . Define successively e1˜ = e2 ∧ e3 ∧ . . . ∧ en−1 ∧ en e2˜ = e3 ∧ e4 ∧ . . . ∧ en ∧ e1 ... en˜ = e1 ∧ e2 ∧ . . . ∧ en−2 ∧ en−1 . Then the only nonzero inner products are e1˜ · e1 = e2˜ · e2 = . . . = en˜ · en = vol [e1 , e2 , . . . , en ]. ˜ we define the n-dimensional Cauchy wavelet as Thus, for η ∈ C, 6n η j˜ )l j e−k· , k ∈ C(e1 , e2 , . . . , en ), j=1 (k · e (C,η) ψ ( k) = l1 l2 ...ln 0, otherwise,

(9.14)

and, in position space,

(C,η) ψl1 l2 ...ln ( x)

l +l +...+ln vol [e1 , e2 , . . . , en ] 1 2 6n = const . z · e j )l j +1 j=1 (

(9.15)

In these formulas, l1 , l2 , . . . , ln ∈ N∗ and the factors (k · e j˜ )l j represent vanishing moments on the faces of the cone. Exactly as in three dimensions, one may also construct Cauchy or conical wavelets = 0 and, in particular, n-D axisymmetric adapted to a general cone of equation F(k) wavelets in the case of a circular cone.

308

Higher-dimensional wavelets

9.2

Wavelets on the 2-sphere and other manifolds

9.2.1

The problem In most cases of physical interest, experimental data are given on the line (signal processing), on the plane (image analysis), or occasionally in R3 (e.g., in fluid dynamics). However, there are situations where data are given on a sphere, for instance, in geophysics or astronomy, of course, but also in statistical problems [Fis87], computer vision or medical imaging (see [216] for precise references). If one is interested only in very local features, one may forget the curvature and work on the tangent plane with standard methods, that is, Fourier or time–frequency analysis, in particular the CWT. However, when global aspects become important (description of plate tectonics on the Earth, for instance, or structure of the Universe as a whole), curvature can no longer be ignored, so that one needs a genuine generalization of wavelet analysis to the sphere (or more general manifolds). Let us first make that statement precise. We may speak of a genuine spherical CWT if (i) the signals and the wavelets live on the sphere; (ii) the transform involves (local) dilations of some kind; and (iii) for small scales, the spherical CWT reduces to the usual CWT on the (tangent) plane (Euclidean limit). Several authors have studied this problem, with various techniques, mostly discrete. For instance: r One may extend to S 2 the discrete wavelet scheme based on a multiresolution analysis, but this approach leads often to numerical difficulties around the poles [114,197,318]. r A different technique is to use the lifting scheme and second generation wavelets, as described in Chapter 2, Section 2.5.2. An efficient solution has been obtained in this way by Schr¨oder and Sweldens [336], but this obviously misses the particular symmetry of the sphere. r One may exploit the geometry of the sphere, as encoded in the system of spherical harmonics [Fre97,Fre99,169,294,319], but the resulting analyzing functions are poorly localized (in fact they do not really resemble wavelets). On the other hand, this approach leads to good approximation methods for spherical functions; we shall come back to this problem in Section 9.3. However, to fully preserve the rotational invariance of the sphere, a continuous approach is clearly necessary. Here too, several authors have proposed solutions. r Considering the fact that the sphere does not admit global dilations, since it is compact, one resorts to a wavelet transform on the tangent bundle of the sphere [116] or, instead, to a Gabor transform on the sphere itself [355,356]. r The most satisfactory approach is that of Holschneider [225], who produces a CWT on the sphere that satisfies the three criteria above. However, the rˆole of dilation is

309

9.2 Wavelets on the 2-sphere and other manifolds

played by an abstract parameter that satisfies a number of ad hoc assumptions. The correct Euclidean limit is obtained, but it is essentially put by hand. As can be seen from this brief description, none of the proposed solutions fully qualifies for a genuine CWT on S 2 . It turns out that the general formalism developed in [Ali00,6] and sketched in Chapter 7 yields an elegant solution to the problem, entirely derived from group theory, and in particular allows one to derive all the assumptions of [225]. Although the discussion is too technical to be given here in detail, it is interesting to outline the main ideas, because they lead to significant generalizations. A detailed treatment may be found in [23,28,29].

9.2.2

The continuous wavelet transform on S 2

9.2.2.1

Affine transformations on the sphere S2 We consider the 2-sphere S 2 , with polar spherical coordinates ζ = (θ, ϕ). As usual, finite energy signals are taken as square integrable functions on the 2-sphere, s ∈ H = L 2 (S 2 , dµ), where dµ(ζ ) ≡ dµ(θ, ϕ) = sin θ dθ dϕ is the usual (rotation invariant) measure on S 2 . The first step for constructing a CWT on S 2 is to identify the natural operations on such signals. These are of two types: (i) Motions or displacements, given by elements of the rotation group SO(3), which indeed acts transitively on S 2 , and S 2 SO(3)/SO(2). (ii) Dilations, that may be derived in two steps. First, dilations around the North Pole are obtained by considering usual dilations in the tangent plane and lifting them to S 2 by inverse stereographic projection from the South Pole. This gives: Da(N ) (θ, ϕ) = (θa , ϕ),

with

tan

θ θa = a · tan . 2 2

(9.16)

Then a dilation around any other point ζ ∈ S 2 is obtained by moving ζ to the North Pole by a rotation , ∈ SO(3), performing a dilation Da(N ) as before and going back (ζ ) by the inverse rotation: Da = ,−1 Da(N ) ,. Clearly the dilations act also transitively on S 2 . Next we have to identify a group of affine transformations on S 2 . First we note that motions , ∈ SO(3) and dilations by a ∈ R+ ∗ do not commute. Also it is impossible to build a semidirect product from SO(3) and R+ ∗ , and therefore the only extension of + SO(3) by R∗ is their direct product. A way out is to embed the two factors into the Lorentz group S Oo (3, 1), by the so-called Iwasawa decomposition: SOo (3, 1) = SO(3) · A · N ,

(9.17)

where A ∼ SOo (1, 1) ∼ R ∼ R+ ∗ is the subgroup of Lorentz boosts in the z-direction and N ∼ C is two-dimensional and abelian (under the stereographic projections, N corresponds to translations in the tangent plane). The appearance of the Lorentz group

310

Higher-dimensional wavelets

SOo (3, 1) in this context is not fortuitous, since it is the conformal group of the sphere S 2 (and of the plane R2 as well). It turns out that the stability subgroup of the North Pole is the so-called minimal parabolic subgroup P = M = SO(2) · A · N , where M = SO(2) is the subgroup of rotations around the z-axis. Thus we get S 2 SOo (3, 1)/P SO(3)/SO(2).

(9.18)

This shows that SOo (3,1) acts transitively on S 2 as well. This action may be computed explicitly using the Iwasawa decomposition (9.17). For a pure dilation by a, the result is precisely the usual dilation lifted on S 2 by inverse stereographic projection, given in (9.16).

9.2.2.2

Spherical wavelets The next step towards constructing wavelets (affine coherent states) on S 2 is to find a suitable unitary irreducible representation (UIR) of the Lorentz group SOo (3, 1) acting in the Hilbert space L 2 (S 2 , dµ). Natural candidates are the representations of the continuous principal series, also called class I representations [Kna96,351]. The simplest one, that we shall use, is given by the operators: (9.19) [U (g) f ] (ζ ) = λ(g, ζ )1/2 f g −1 ζ , g ∈ SOo (3, 1), f ∈ L 2 (S 2 , dµ), where g = ,an by the Iwasawa decomposition and the multiplier λ (g, ζ ) is a Radon– Nikodym derivative (or a 1-cocycle), expressing the fact that the measure dµ is not invariant under the full group SOo (3, 1): dµ g −1 ζ λ (g, ζ ) = (9.20) , g ∈ SOo (3, 1). dµ (ζ ) This representation U of SOo (3, 1) is unitary and irreducible. Since we are only interested in the action of dilations and motions, we quotient out the subgroup N . In other words, the parameter space of the spherical wavelets is X = SOo (3, 1)/N SO(3) · R+ ∗ . Then, introducing a suitable section σ : X = SOo (3, 1)/N → SOo (3, 1), we concentrate on the reduced expression (9.21) [U (σ (x)) f ] (ζ ) = λ(σ (x), ζ )1/2 f σ (x)−1 ζ , x ≡ (,, a). We choose the natural (Iwasawa) section σ (,, a) = , a, , ∈ SO(3), a ∈ A. Using the action (9.16) of dilations, one gets easily 4a 2 λ(σ (,, a), ζ ) ≡ λ(a, θ) = $ %2 , a 2 − 1 cos θ + a 2 + 1

ζ = (θ, ϕ).

(9.22)

The function λ(a, θ) satisfies the so-called cocycle relation (which guarantees that U is indeed a representation): λ(a −1 , θ ) λ(a, θa ) = λ(1, θ) = 1.

(9.23)

311

9.2 Wavelets on the 2-sphere and other manifolds

In addition, from the choice of the section, we have U (σ (,, a)) = U (, a) = U (,)U (a), and therefore the representation (9.21) factorizes as U (σ (,, a)) = R, Da .

(9.24)

In this relation, R, ≡ Uqr (,), , ∈ SO(3), where Uqr is the quasi-regular representation of S O(3) in L 2 (S 2 , dµ), and Da , a ∈ R+ ∗ , is an operator of pure dilation, that is, (Da f ) (ζ ) = λ(a, θ)1/2 f (ζ1/a ), with ζa ≡ (θa , ϕ). The quasi-regular representation of SO(3), (Uqr (,) f )(ζ ) = f (,−1 ζ ), is infinite dimensional and decomposes into the direct sum of all the familiar (2l + 1)-dimensional representations, l = 0, 1, 2, . . . . Following the general approach of [Ali00,6], we build now a system of spherical wavelets, realized as coherent states for the Lorentz group, indexed by points of X = SOo (3, 1)/N . Since N is not the isotropy subgroup of a particular vector in the representation Hilbert space, the resulting coherent states are not of the Gilmore– Perelomov type [Per86] (see Section 7.1.5). First we show that the UIR (9.21) is indeed square integrable on X , that is, we check that there exists a nonzero vector ψ ∈ L 2 (S 2 , dµ) such that dν(,, a) |U (σ (,, a))ψ|φ|2 < ∞, ∀ φ ∈ L 2 (S 2 , dµ), X

where dν(,, a) = a −3 d, da, and d, is the invariant (Haar) measure on SO(3). Proposition 9.2.1 The representation U given in (9.21) is square integrable modulo the subgroup N and the section σ . A nonzero vector ψ ∈ L 2 (S 2 , dµ) is admissible mod(N , σ ) iff there exists c > 0, independent of l, such that, for all l 0: l ∞ da ( 8π 2 |ψa (l, m)|ψa |2 < c . Gl ≡ 2l + 1 m=−l 0 a 3

(9.25)

Here f (l, m) ≡ Ylm | f denotes a Fourier coefficient of f ∈ L 2 (S 2 ), with Ylm the usual spherical harmonic, and ψa = Da ψ = U (σ (e, a))ψ corresponds to a pure dilation. The proof, as usual, consists in an explicit calculation, using the properties of Fourier analysis on the sphere. Thus any admissible ψ generates a continuous family {ψa,, ≡ U (σ (,, a))ψ, (,, a) ∈ X } of spherical wavelets, but in fact we have more. 2π Proposition 9.2.2 For any admissible vector ψ such that 0 dϕ ψ(θ, ϕ) ≡ 0 (for instance, axisymmetric), the family {ψa,, , (,, a) ∈ X } is a continuous frame, that is, there exist constants A > 0 and B < ∞ such that 2 A φ dν(,, a) |ψa,, |φ|2 B φ2 , ∀ φ ∈ L 2 (S 2 , dµ). (9.26) X

312

Higher-dimensional wavelets

Thus, for most admissible vectors ψ, we get a continuous frame, but not necessarily a tight frame. We conjecture that the resulting frame is never tight, that is, A = B.

9.2.2.3

The spherical wavelet transform Proposition 9.2.1 yields the basic ingredient for writing the CWT on S 2 . Given an admissible vector ψ ∈ L 2 (S 2 , dµ), our wavelets on the sphere are the functions ψa,, = U (σ (,, a))ψ = R, Da ψ = R, ψa .† Then, the spherical CWT of a signal s ∈ L 2 (S 2 ) is defined as S(,, a) = ψa,, |s = dµ(ζ ) ψa,, (ζ ) s(ζ ) 2 S dµ(ζ ) ψa (,−1 ζ ) s(ζ ). =

(9.27)

S2

This relation gives the spherical CWT as a convolution on the sphere S 2 : ψa (ζ ) and s(ζ ) are two functions on S 2 , and their convolution ψa % s is a function on SO(3). Such a formulation leads both to mathematical subtleties and to numerical difficulties. The former will be treated in Section 9.3 (see also Appendix A) and the latter in Section 9.2.5, where we will discuss the numerical implementation of the spherical CWT. According to the general theory, the admissibility of the wavelet ψ is sufficient to guarantee the invertibility of the spherical CWT on its range, that is, we can reconstruct the signal s(ζ ) from its wavelet transform S(,, a): s(ζ ) = dν(,, a) S(,, a) A−1 (9.28) σ ψa,, (ζ ) . X

In this relation, Aσ denotes the resolution operator (7.34), whose action is a multiplication in Fourier space (as it is the case with most Duflo–Moore operators, for instance, (7.42)), A σ f (l, m) = G l f (l, m), with G l defined in the admissibility condition (9.25). As usual, the integral in (9.28) is to be taken in the weak sense. Of course, as in the flat case, one may consider more general reconstruction formulas, with two different wavelets for the analysis and the synthesis (see (2.31) in Section 2.2). These formulas gets simpler if the wavelet ψ is axisymmetric, i.e., ψ(ζ ) ≡ χ(θ), for, then, we may exploit the fact that S 2 SO(3)/SO(2). First, since the dilation is purely radial, ψa is also axisymmetric and ψa (ζ ) = χa (θ ). Then the action of R, on ψa has for sole effect to transport its center from the North Pole ζo to some point ζ = , ζo . †

As the notation ψa,, suggests, all the operations involved in the CWT consist in manipulating the function ψa at a fixed scale a. This is consistent with [Fre97,169] and [35], but not with [29], where the same wavelets were denoted ψ,,a .

313

9.2 Wavelets on the 2-sphere and other manifolds

We can thus characterize R, ψa by its center point ζ , which is independent of a, and we may write ψa,, ≡ ψa,ζ . A more precise way of achieving this is to split , ∈ SO(3) into , = (χ , [ζ ]) with χ ∈ SO(2) and ζ ∈ S 2 . This is formally done through a projection , → ζ (,) in the fiber bundle SO(3) → S 2 SO(3)/SO(2), followed by an arbitrary choice of section ζ → [ζ ] in S O(3). The splitting corresponds to decomposing the motion R, of the wavelet ψa into an initial rotation of angle χ around the North Pole ζ0 followed by a transport to the point ζ = , ζo on the sphere. In other words, R, ψa (ζ ) = Rχ ψa ([ζ ]−1 ζ ) where Rχ is a rotation around the North Pole. Accordingly, the spherical wavelet transform will also be denoted by S(χ, ζ , a). This amounts simply to decomposing the parameter space SO(3) × R+ ∗ into a more appropriate form, including “translations” ζ , dilations a, and “rotations” χ , exactly as in the Euclidean case. Of course, when the wavelet ψ is axisymmetric, the dependence on χ can be dropped and the spherical wavelet transform will be written simply as S(ζ , a): dµ(ζ ) ψa,ζ (ζ ) s(ζ ). (9.29) S(ζ , a) = S2

(A related statement is given in Proposition 9.3.1, in Section 9.3.) In that case, the −3 parameter space of the spherical CWT reduces to S 2 × R+ dµ(ζ ) da. ∗ , with measure a Hence, the integral over SO(3) in the reconstruction formula is replaced by an integral over S 2 : dµ(ζ ) da s(ζ ) = S(ζ , a) A−1 (9.30) σ ψa,ζ (ζ ). 3 + 2 a S R∗ According to the general coherent state formalism, the reconstruction formulas (9.28) and (9.30) are valid only in the weak sense. In the flat case, however, we have seen in Section 2.6 that the corresponding formula holds in the strong L 2 sense. This guarantees that it can be used for approximating functions on the plane through an approximate identity. That means, the approximating function is obtained by convolution with a smoothing kernel, which tends to the identity (δ function) as the parameter goes to 0. We will show in Section 9.3 that exactly the same situation prevails on the sphere. In order to prove these results, we will have to switch to an L 1 formalism (as already mentioned in [29]), by introducing a modified dilation operator D a that preserves the L 1 norm of functions. As a consequence, we will have at our disposal two types of spherical CWT. An important aspect of the flat space CWT is its covariance property (Proposition 2.2.3). In the present case, an explicit calculation [29] shows that the spherical CWT (9.27) is covariant under motions on S 2 , but not covariant under dilations. r It is covariant under motions on S 2 , namely, for any , ∈ SO(3), the transform of the o rotated signal s(,o−1 ζ ) is the function S(,o−1 ,, a).

314

Higher-dimensional wavelets r

But it is not covariant under dilations. Indeed the wavelet transform of the dilated signal λ(ao , ζ )1/2 s(ao−1 ζ ) is U (g)ψ|s, with g = ao−1 ,a, and the latter, while a welldefined element of SOo (3, 1), is not of the form σ (, , a ). For applications, of course, it is the covariance under motions that is essential, since it reduces to translation covariance in the Euclidean limit, as we shall see in Section 9.2.3. As for dilations, the negative result reflects the fact that the parameter space X SO(3) × R+ ∗ of the spherical CWT is not a group. The condition (9.25), which was derived in [225] in a different way, is necessary and sufficient for the admissibility of ψ, but it is somewhat complicated to use in practice, since it requires the evaluation of nontrivial Fourier coefficients. Instead, there is a simpler, although only necessary, condition. Proposition 9.2.3 A function ψ ∈ L 2 (S 2 , dµ) is admissible only if it satisfies the condition ψ(θ, ϕ) dµ(ζ ) = 0, ζ ≡ (θ, ϕ). (9.31) 2 1 + cos θ S This necessary condition is the exact equivalent of the usual necessary condition for wavelets in the plane, d 2 x ψ( x ) = 0, and it reduces to the latter in the Euclidean limit (see Section 9.2.3). The interesting point is that (9.31) is a zero mean condition, as in the flat case. As such it ensures that the CWT on S 2 given in (9.27) acts as a local filter. This is crucial for applications and it is one of the main reasons of the efficiency of the CWT, and the same holds here. Using Proposition 9.2.3, it is easy to build explicit wavelets on the sphere, namely “Difference wavelets”, similar to the ones described for the flat case in Chapter 3, Section 3.2.2. For that purpose, we notice the following easy result [28]. Proposition 9.2.4 Let φ ∈ L 2 (S 2 , dµ). Then 1 Da φ(θ, ϕ) φ(θ, ϕ) = , dµ(ζ ) dµ(ζ ) a S2 1 + cos θ 1 + cos θ S2 where Da = U (σ (e, a)) is again a pure (covariant) dilation. Given a square integrable (smoothing) function φ, we define ψφ(α) (θ, ϕ) = φ(θ, ϕ) −

1 Dα φ(θ, ϕ) α

(α > 1).

(9.32)

By Proposition 9.2.4, ψφ(α) satisfies the admissibility condition (9.31), that is, it is a spherical wavelet, and it is fully admissible if φ is sufficiently regular at the poles. The simplest difference wavelet is obtained with the choice φG (θ, ϕ) = exp(− tan2 θ2 ), which is essentially the inverse stereographic projection of a Gaussian in the tangent plane. The corresponding spherical wavelet, called the spherical DOG wavelet and

9.2 Wavelets on the 2-sphere and other manifolds

1 1

0.5

0.5

Z

Z 0

0

−0.5

−0.5

−1

−1 −1

−0.5

−0.5

−0.5

0 0.5

−1 0

0

−0.5 0.5

0.5

0 1

0.5

X

Y

X

Y

1

0.5

Z

315

0

−0.5

−1 −1 −0.5

−0.5 0

0 0.5

0.5 1

X Y

(α)

Fig. 9.5. The spherical DOG ψG wavelet, for α = 1.25. (Top line) The wavelet at scale a = 0.125

and positioned at the North Pole θ = 0◦ (left) and on the equator θ = 90◦ , ϕ = 90◦ (right); (Bottom) The same at scale a = 0.0625 and position θ = 90◦ , ϕ = 0◦ . As mentioned in the text, “at scale a” means the function Da ψφ(α) .

denoted ψG(α) in the sequel, is an axisymmetric wavelet, which is shown in Figure 9.5, for different values of the scale a and in various positions (θ, ϕ) on the sphere. Note that here “ψ at scale a” means that the function being plotted is Da ψ, i.e., one must always use the covariant dilation operator Da . This wavelet yields an efficient detection of discontinuities on the sphere [28]. Explicit examples of spherical wavelet transforms based on it will be given in Section 9.2.5.3. Now, the construction of the scaling function φG by inverse stereographic projection from the tangent plane suggests a general procedure for generating spherical wavelets.

316

Higher-dimensional wavelets

We will implement it in Section 9.2.4, where we will construct directional wavelets on the sphere by the same method. Before going into that, let us discuss the Euclidean limit of our spherical wavelet transform.

9.2.3

The Euclidean limit As said above, a good wavelet transform on the sphere should be asymptotically Euclidean, that is, the spherical WT should match the usual CWT in the plane (in this case, the tangent plane at the North Pole) at small scales or, what amounts to the same, for large values of the radius of curvature. This statement may be given a precise mathematical meaning, using the technique of group contractions (or deformations). Without entering into technical details, we sketch the successive steps. First, we reformulate the theory on a sphere of radius R and let R → ∞. Then SR2 becomes the plane R2 , the group SO(3) contracts into the Euclidean group of R2 and the Lorentz group SOo (3, 1) into the (semidirect) product G E = R2 SIM(2), where 2 SIM(2) = R2 (R+ ∗ × SO(2)) is the similitude group of R , that is, the invariance group of the Euclidean CWT, discussed in Chapter 7, Section 7.1.2. Notice that the contraction preserves the minimal parabolic subgroup P = M AN ∼ SIM(2), and, in particular, the subgroup SO(2) of rotations around the z-axis. The next step is to transfer the contraction process to the relevant homogeneous spaces. On one hand, the manifolds S 2 = SOo (3, 1)/M AN and R2 = G E /M AN , that carry the respective CWT, are related through contraction. On the other hand, since the abelian subgroup N is preserved under the contraction, the parameter space X = SOo (3, 1)/N SO(3) × A of the spherical CWT goes into that of the Euclidean CWT, namely SIM(2) = G E /N . Notice that the former is not a group (and this forces us to use the general formalism of [Ali00,6], described in Chapter 7), whereas, after contraction, we get SIM(2), that is, the missing group structure is restored by the contraction! In this geometrical context, the Euclidean limit itself can be formulated as a contraction at the level of group representations. Whereas contractions of Lie algebras and Lie groups are relatively ancient and well-known [232,331], the extension of the procedure to group representations is rather recent [272]. A rigorous version has been given by Dooley [150–152], that was followed in [23,29]. The additional difficulty here is that the representation space itself varies during the procedure. Let H R = L 2 (S R2 , dµ R ) be the Hilbert space of square integrable functions on a sphere of radius R (where dµ R = R 2 dµ) and H = L 2 (R2 , d 2 x). The two spaces are related by the unitary map I R : H R → H, given as (I R f ) (r, ϕ) =

4R 2 r , ϕ , f 2 arctan 4R 2 + r 2 2R

(9.33)

317

9.2 Wavelets on the 2-sphere and other manifolds

where (r, ϕ) are polar coordinates in the plane. Clearly, the map I R just describes the stereographic projection of the sphere S R2 onto its tangent plane at the North Pole. The inverse map, that is, the inverse stereographic projection, reads as (I R−1 f )(θ, ϕ) =

θ 2 f (2R tan , ϕ), 1 + cos θ 2

(9.34)

(in the case R = 1, this map was given already in (5.1)). Now let U be the usual wavelet representation (2.13) of SIM(2) in H and U R the representation (9.21) of SOo (3, 1) realized in H R . For each R, we choose the corresponding representation U R . Then the precise statement is that the representation U of SIM(2) is a contraction, in the sense of Dooley [150–152] of the family of representations U R of SOo (3, 1) as R → ∞. This means that, for every g ∈ SIM(2), the following strong limit holds in H: ˜ R (g) I R−1 φ − U (g)φH = 0, lim I R U R (9.35) R→∞

˜ R : SIM(2) → X is the so-called reduced contraction map (see [29] for details). where This theorem yields the expected result that local wavelet analysis on the sphere as defined here is equivalent to local wavelet analysis in flat space. Indeed the whole structure on the sphere S R2 goes into the corresponding one in R2 as R → ∞. Since U R → U , the corresponding matrix elements converge to one another, and so the square integrability condition (9.25) converges into the corresponding one for the CWT in R2 , namely k)| 2 |ψ( d 2 k < ∞. 2 |k| R2 Admissible wavelets on S 2 converge to admissible wavelets on R2 (Proposition 9.2.5), and the necessary condition (9.31) also goes into the usual one in the plane, namely d 2 x ψ( x ) = 0. R2

9.2.4

Directional wavelets on the sphere The general mathematical setting for designing directional wavelets on S 2 has been introduced in the previous sections, but the construction and the properties of these particular spherical wavelets have not been discussed. Yet these wavelets are quite important in practice, since directional features (roads, streams, geological faults, . . . ) abound on the spherical Earth! Thus one really needs the additional degree of freedom they offer for characterizing signals.

9.2.4.1

General remarks Whenever the wavelet ψ is not axisymmetric, the continuous spherical wavelet transform depends on the additional parameter χ , the rotation angle around the North Pole, thus (9.27) may be rewritten as

318

Higher-dimensional wavelets

S(χ , ζ , a) =

S2

dµ(ζ ) Rχ ψa ([ζ ]−1 ζ ) s(ζ ).

(9.36)

In this formula, there is an arbitrariness in the way the rotation [ζ ] of S O(3) is associated to the point ζ on the sphere. The section [·] : S 2 → S O(3) can be depicted as mapping the sphere to a tangent vector field of unit length defined on it. Intuitively, one would like this mapping to be smooth to correspond to the idea of direction defined on the sphere, and the arbitrariness in the choice of the section may be exploited to that effect, at least locally. When this is the case, we expect that the values of the wavelet transform correspond to filtering in a given direction χ and at a given scale a as in the case of the 2-D wavelet transform in the plane. Some caution must be exercised, however, when dealing with directions on the sphere. Indeed, it is a classical result in topology that there exists no differentiable vector field of constant norm on S 2 , which means there is no global way of defining directions. There will always be some singular point where the definition fails.† In other words, one cannot comb a perfectly spherical porcupine! Nevertheless, testing orientations using directional wavelets is a small-scale operation, that is, a local procedure [24]. Then, around any point on S 2 , the support of the wavelet at small scale defines a neighborhood in which the following reasoning holds. This ability to perform local analysis is definitely one the most important properties of wavelet analysis. We will see more about this in the examples below. From now on, we will make use of the classical parametrization of S O(3) in terms of Euler angles, , ≡ (χ, θ , ϕ ), which corresponds to the choice of section (θ , ϕ ) → (0, θ , ϕ ), which in turn defines a direction on the sphere. The singular points are the North and South Poles: it makes no sense to define cardinal points at the poles! For this choice of parametrization, we may write Rχ ψa ([ζ ]−1 ζ ) = ψa,χ,ζ (ζ ) ≡ ψa,χ,θ ,ϕ (θ, ϕ),

(9.37)

which implies ψa,χ ,θ ,ϕ (θ, ϕ) = ψa,χ,θ ,0 (θ, ϕ − ϕ ).

(9.38)

Therefore, (9.36) becomes a convolution in ϕ which, by means of the convolution theorem, takes the form π 2π S(χ , θ , ϕ , a) = ψa,χ,θ ,0 (θ, ϕ − ϕ ) s(θ, ϕ) sin θ dθ dϕ (9.39) 0

= 2π

0

∞ k=−∞

†

e

i kϕ

π

ψˇ a,χ ,θ ,0 (θ )[k] sˇ (θ)[k] sin θ dθ,

(9.40)

0

This statement is valid for S 2 , but not in the case of the circle S 1 and the higher dimensional spheres S 3 and S 7 .

319

9.2 Wavelets on the 2-sphere and other manifolds

ˇ )[k] denotes its Fourier coefficient with where, for any function h on the sphere, h(θ respect to the longitudinal coordinate ϕ: 2π ˇ )[k] = h(θ dϕ h(θ, ϕ) e−i kϕ . (9.41) 0

In the discretization method of Section 9.2.5, the relations (9.39)–(9.40) will give us a tool for reducing the computational time of the spherical CWT. Indeed, they will allow us to use the fast Fourier transform (FFT).

9.2.4.2

Estimating the angular selectivity of a wavelet Given a wavelet ψ, it is very important in practice to know how well it will discriminate between two close directions. In other words, we would like to quantify the angular resolving power (ARP) of a spherical wavelet (see Section 3.4.1). A tempting definition is simply to look at the correlation between ψ and its rotated version: ψ|Rχ ψ K ψ (χ ) = = ψ−2 ψ|Ylm Ylm |Rχ ψ ψ|ψ l 0 |m|l −2 = ψ ψ|Ylm Rχ Ylm |ψ l 0 |m|l

= ψ =

−2

l 0 |m|l

ψ−2

=

m)|2 eimχ |ψ(l,

l |m|

m∈Z

ψ|Ylm eimχ Ylm |ψ

am e

imχ

.

m∈Z

As a function of χ, this expression reduces to a Fourier series where the coefficients are the same as in the spherical harmonic expansion of ψ. If the wavelet is highly sensitive to changes in the orientation, K ψ should be peaked around χ = 0. As we shall see later, the problem with this definition is that it does not depend on the scaling parameter a. This is not a problem for wavelets in R2 , but is rather counterintuitive on the sphere, because we know that a direction cannot be defined at large scales. This motivates the study of a more general indicator. Let us then introduce the following operator-valued function: Rψ,a (χ ) = dµ(ζ ) |ψa,χ,ζ ψa,χ,ζ |, (9.42) S2

where we choose again for [·] the Euler angles section. A good candidate for the ARP would then be the mean value: Rψ,a (χ )ψa =

ψa |Rψ,a (χ )ψa . ψa |ψa

(9.43)

320

Higher-dimensional wavelets

This time the inspection of Rψ,a (χ )ψa for different values of a should reveal a lack of angular precision at large scales. This is precisely what is observed in the case of the spherical Morlet wavelet, as shown later in Figure 9.8.

9.2.4.3

Designing directional spherical wavelets We have not yet addressed the problem of constructing good directional wavelets on S 2 . We will show now that this job is very naturally handled in our framework. First of all, we recall that the very definition of a direction on S 2 forces us to work at small scales. As is well known, the geometry of S 2 at small scales, or for large radii of the sphere, is closer and closer to that of R2 . As discussed in Section 9.2.3, the spherical wavelet transform respects one’s intuition by nicely approximating the Euclidean wavelet transform at small scales (the Euclidean limit property). We may remark that the notation used in (9.37) is consistent with it: roughly speaking, as the radius of the sphere goes to infinity, ψa,χ ,ζ (ζ ) goes to ψa,χ,b ( x ), where b ∈ R2 is the translation parameter. Moreover, it is a simple application of the Euclidean limit that small-scale Euclidean wavelets can be mapped to the sphere and yield small-scale admissible spherical wavelets. These can then be dilated to larger scales using the spherical dilation. This is neatly summarized by the following result, where we repeat for convenience the definition (5.1) of the inverse stereographic projection, namely, in polar coordinates, (I −1 f )(θ, ϕ) =

θ 2 f (2 tan , ϕ). 1 + cos θ 2

(9.44)

Proposition 9.2.5 The inverse stereographic projection (9.44) is a unitary map I −1 : L 2 (R2 , d x) → L 2 (S 2 , sin θ dθ dϕ) between the respective Hilbert spaces. Moreover, if ψ ∈ L 2 (R2 ) is an admissible 2-D Euclidean wavelet, then the function I −1 ψ is an admissible spherical wavelet. This results tells us that we can construct a spherical wavelet starting from any Euclidean wavelet. Now what does this tell us about directional wavelets? Since directional sensitivity is a local or small-scale attribute, it should intuitively survive this process. Yet there is more than intuition in this result. The stereographic projection and both spherical and Euclidean dilations are conformal mappings. Thus Proposition 9.2.5 defines a conformal application that, by definition, preserves angles. The directional sensitivity of the Euclidean wavelet is thus transported to the spherical wavelet. A natural candidate for building a directional spherical wavelet is to start with the (truncated) Euclidean Morlet or Gabor wavelet (3.19):

ψG ( x ) = ei k0 ·x e−|x | . 2

(9.45)

Using Proposition 9.2.5, we find the following spherical wavelet: θ

2 θ 2

eik0 tan 2 cos (φ0 −φ) e− 2 tan ψG (θ, φ) = 1 + cos θ 1

.

(9.46)

321

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.6. Real part of the spherical Morlet wavelet (9.46) at scale (a) a = 0.03 and (b) a = 0.3

(from [35]).

Fig. 9.7. Real part of the spherical Morlet wavelet (9.46) at scale a = 0.03 and centered at

(π/3, π/3): (a) θ = 0 and (b) θ = π/2 (from [35]).

This function is represented in Figures 9.6 and 9.7 for various values of the scale and rotation parameters. An example illustrating the directional sensitivity of this wavelet will be presented in Section 9.2.5.3 below (Figure 9.14), where we will compare it with that of the spherical DOG wavelet function. Of course, a spherical wavelet analysis should always be performed at small scales (see Section 9.2.5.2). As a confirmation, we show in Figure 9.8 the ARP (9.43) of

322

Higher-dimensional wavelets

1 0.9 0.8

a

〈 Rψ, (χ) 〉ψ

a

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.2

0.4

0.6

0.8

χ

1

1.2

1.4

1.6

1.8

Fig. 9.8. Angular resolving power (9.43) of the spherical Morlet wavelet (9.46) for various values of

the scale parameter: a = 0.03 (plain), a = 0.1 (dotted), a = 0.3 (dashed-dotted) and a = 1 (dashed).

the spherical Morlet wavelet for different values of the scale parameter. One clearly notices that, as the scale gets bigger, angles that are far apart become more and more correlated.

9.2.4.4

Representation of object surfaces The spherical wavelets just described are the basis of a novel representation of object surfaces [89] – a longstanding problem in computer graphics. The idea is the following. The surface of a 3-D (star-shaped) object is treated as a function on the 2-sphere (a human head is a good example, with a lot of neurophysiological interest). In a first step, the coarse structure of the surface is described by a truncated expansion of that function into spherical harmonics. Thus we get a band-limited function: f c (θ, ϕ) =

L

f (l, m) Ylm (θ, ϕ),

(9.47)

l=0 |m|l

for some small value of L (the authors choose L = 5). The justification is that the small l terms in the Fourier expansion are low-frequency components, which are well represented by functions supported on the whole sphere, such as spherical harmonics. Then the remainder f res = f − f c , which represents the fine structure of the surface, is described by well-localized spherical wavelets, in particular, spherical Morlet wavelets, with the help of an appropriate optimization procedure. The crucial step here is to

323

9.2 Wavelets on the 2-sphere and other manifolds

treat the low- and high-frequency components separately. We will see in Section 9.3 a mathematically precise version of this procedure. Namely, in the L 1 formalism, the reconstruction formula for the spherical CWT is based on precisely the same idea, with the scale variable a as the relevant parameter to distinguish between low and high frequencies (Theorems 9.3.8 and 9.3.9, which yield indeed approximation schemes for functions on the sphere). The same technique may be used also [90] to define smoothing on S 2 by a spherical Gaussian kernel, obtained again by inverse stereographical projection from a plane Gaussian (essentially our function φG of Section 9.2.2.3). Actually this Gaussian is nothing but a heat kernel, which generates a diffusion process on S 2 . This approach opens interesting new perspectives in object deformation (“morphing”), along the lines of the work of Sweldens et al. on digital geometry processing [242].

9.2.5

Implementation of the spherical wavelet transform The spherical CWT (9.27) is given as a convolution over the sphere S 2 . This creates numerical difficulties, for no really fast algorithm exists today. In particular, it is difficult to find an appropriate discretization of the sphere. Several methods have been proposed in the literature, mostly based on Fourier and spherical harmonics techniques [Moh97,154,169,216,274], but none of them is fully satisfactory. A possible exception is the method introduced by Wandelt and G´orski [367] and based on the use of the FFT. Interestingly enough, this approach was motivated by the analysis of cosmic microwave background (CMB) data, that we have mentioned in Section 5.1.2 – and for which spherical wavelets have been proposed! The new algorithm presented here, however, seems to answer the question rather well [35]. For a practical implementation of the spherical CWT, the first step is that of discretization. This means finding a suitable grid in the parameter space, so as to allow a fast calculation and a good approximation of the continuous theory. As we shall see, the key to the algorithm presented below is to use an FFT in the (periodic) longitude angle ϕ. Actually we also need some sort of criterion on the grid density for controlling aliasing problems, as indicated already in [13]. More precisely, we have to specify the scale interval in which the spherical wavelet transform makes sense. A possible answer will be suggested in Section 9.2.5.2. Then several examples will be discussed, both academic and real-life.

9.2.5.1

Discretization and algorithm Following an approach similar to that in [Win95], the first step is to discretize the integral (9.39) on a regular spherical grid M × N , π 2π t, ϕ p = p) | 0 t M − 1, 0 p N − 1}, M N by a weighted sum (χ and a are fixed throughout) G = {(θt =

(9.48)

324

Higher-dimensional wavelets

S(χ , θt , ϕ p , a) ≡ S[χ, t , p , a] = ψa,χ,t [t, p − p ] s[t, p] wt ,

(9.49) (9.50)

0 t M−1 0 p N −1

where: r s[t, p] ≡ s(θ , ϕ ); t p

r ψ

[t, p − p ] ≡ ψa,χ,θt ,0 (θt , ϕ p− p ); a,χ ,t r the index of ϕ is extended to Z by angular periodicity with the rule ϕ p+N = ϕ p ; r w = (2π 2 /M N ) sin θ are the weights suggested in [Win95] for the discretization t t of the Lebesgue measure on the particular grid G. Evaluating the sums in (9.50) requires M N additions and multiplications for each (t , p ), that is, M 2 N 2 operations altogether. However, an easy simplification can be obtained for the longitudinal coordinates by the use of a Fourier series and the Plancherel formula. Indeed, denoting by 2π ˇ k] = h[t, h[t, p] exp(−i kp ), (9.51) N 0 p N −1 the longitudinal Fourier coefficients of a given discrete function h, we obtain S[χ , t , p , a] = 2π wt F[χ, t , p , a, t]

(9.52)

0t M−1

with F[χ , t , p , a, t] =

2π ψˇ a,χ,t [t, k] sˇ [t, k] exp(i kp ). N 0k N −1

(9.53)

The quantity F may be computed with the inverse fast Fourier transform (IFFT), which leads to a reduction of the computational time from O(M 2 N 2 ) to O(M 2 N log N ). On a grid G of 256 × 256, the gain is a factor of 46. Notice that other discretization techniques than a plain Riemann sum, as used in (9.50), would be beneficial only if one imposes additional regularity conditions on the signal s. Furthermore, other weights wt could be chosen to achieve a better approximation of (9.49). An example of a different choice, both for the weights and for the discretization technique, is that of a band-limited spherical function, as considered in [216] and [367], that is, a function f (θ, ϕ) whose expansion in spherical harmonics has only finitely many nonzero terms, f (l, m) = 0, ∀ l > L. One can also ask whether a better algorithm could be designed on another type of grid. For instance, what is the use in the continuous case of an icosahedral grid, as introduced in [336,337] (see Figure 9.9) for the starting point of a lifting procedure? The great advantage of such a grid (or one built on a fullerene type of grid) is its better isotropy. The point is that the usual spherical coordinates necessarily introduce a preferred direction, the polar axis, and this is at the origin of the various tentatives reported in the literature

325

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.9. The geodesic sphere construction, starting with the icosahedron on the left (subdivision

level 0) and the next two subdivision levels (from [336]).

of designing an equidistribution of points on the sphere [113,169,330]. The last paper, in particular, contains many references to related works, from pure mathematics to physics (electrostatics).

9.2.5.2

Numerical criterion for the scale range The discretization of the continuous spherical wavelet transform gives rise to a sampling problem. Since the grid G is fixed, if we contract or dilate our wavelet too much, we obtain a function which is very different from the original ψ. In other words, aliasing occurs and the wavelet is not numerically admissible. This problem has been noted also in the flat space CWT (see Section 4.1.1), but only an empirical criterion was given [13]. We have seen in Proposition 9.2.3 that a function ψ ∈ L 2 (S 2 , dµ) is admissible only if it satisfies the zero mean condition (9.31). Approximating the integral by its Riemann sum, we get the quantity C[ψ] =

1t M−1 1 p N −1

ψ(θt , ϕ p ) µ(θt , ϕ p ). 1 + cos θt

(9.54)

Because of the discretization, even if ψ verifies (9.31), it is not necessarily true that C[ψ] vanishes. However, we may suppose that this quantity is very close to zero when ψ is sampled sufficiently, that is, if the grid G is fine enough. Nevertheless, it is difficult to give a quantitative meaning to the value of C[ψ]. How small is “very close to zero”? Here is a possible solution to this problem. Since the spherical measure µ and the function 1 + cos θ are positive, it is clear that C[ψ] C[|ψ|]

(9.55)

for any ψ ∈ L 2 (S 2 , dµ). So we can define a numerical normalized admissibility by C[ψ] , C[ψ] = C[|ψ|] a quantity always contained in the interval [-1,1]. We can now give a precise definition of numerical admissibility.

(9.56)

Higher-dimensional wavelets

Definition 9.2.6 A spherical wavelet of L 2 (S 2 , dµ) is numerically admissible with threshold p% or p% - admissible, if the numerical normalized admissibility (9.56) is smaller than (100 − p)/100 in absolute value: |C[ψ]|

100 − p . 100

(9.57)

As an example, we present in Figure 9.10 the behavior of the dilated spherical DOG wavelet, Da ψG(α) (α = 1.25), as a function of a > 0, discretized on a 128 × 128 grid (notice that, in the flat case, α = 1.6 is the value for which the DOG wavelet is almost indistinguishable from the Mexican hat). According to this plot, the wavelet Da ψG(α) is 99 % - admissible on the scale interval a ∈ [0.072, 24.71]. The lower limit is due to the fact that, for small a, Da ψ is not sampled enough. The upper limit comes from the subsampling of the area far from the North Pole which, according to the spherical dilation, gets more and more contracted. Figure 9.11 presents three typical behaviors of Da ψ discretized on a 22 point θ sampling. For a = 0.5, the sampling is correct. For a = 0.05, that is, below the lower admissibility bound, subsampling occurs, so that negative parts of Da ψ are completely missed. Clearly, this discretized wavelet is no longer admissible. Exactly the same effect was observed long ago in the flat case [13]. The third case, with a = 3.5, thus a=0.072

a=24.71

1

0.8

0.6

0.4

Num. Norm. Admiss.

326

0.2

0

−0.2

−0.4

−0.6

−0.8

−1 −5

−4

−3

−2

−1

0

1

2

log a

a ψG(α) ] as a function of log a for α = 1.25 (from [35]). Fig. 9.10. C[D

3

4

5

327

9.2 Wavelets on the 2-sphere and other manifolds

0.8

(a)

(α) 0.7

DaψG

0.6 0.5 0.4 0.3 0.2 0.1 0 −0.1 −0.2 −4

−3

−2

−1

0

1

8

DaψG(α)

2

3

4

θ 0.2

(b)

7

(c)

(α) 0.15

D a ψG

6

0.1

5

0.05

4

0 −0.05

3 2

−0.1

1

−0.15 −0.2

0 −1 −4

−3

−2

−1

0

1

2

3

4

θ

−0.25 −4

−3

−2

−1

0

1

2

3

4

θ

(α)

Fig. 9.11. Three typical behaviors of Da ψG discretized on a 22 × 22 grid G. (a) For a = 0.5, the

sampling is correct; (b) for a = 0.05, subsampling occurs, negative parts of Da ψG(α) are completely missed; (c) subsampling on the negative parts of Da ψG(α) for a = 3.5. Notice the minimum at θ = 0 (from [35]).

beyond the upper bound, is less intuitive. Here the subsampling takes place for large values of θ, that is, close to the South Pole, but the result is the same; the discretized wavelet does not have a zero mean, it is not admissible. In addition, the curve presents a minimum at θ = 0. This somewhat unexpected effect is in fact due to the cocycle, as is the dependence of the height on a. Indeed, if one performs the same calculation without the cocycle, all curves show a maximum at θ = 0, with the same height. Here again we see that curvature, which requires the presence of the cocycle, has a nontrivial effect. As a further illustration of this behavior, it is instructive to consider the function ι identically equal to 1. In the flat case, this function has a vanishing WT, by the admissibility condition R2 d 2 x ψ( x ) = 0 on the wavelet, but it is not square integrable and thus cannot be reconstructed. In the present case, however, the situation is different. The function ι is square integrable, since the sphere S 2 is compact, but its WT does not vanish, because of the presence of the cocycle. Indeed, the function ι is invariant under rotation, but not under dilation:

328

Higher-dimensional wavelets

(Da ι)(θ, φ) = λ(a, θ )1/2 ≡ 1,

(9.58)

and, therefore, I (,, a) = R, Da ψ|ι = ψ|Da ι ≡ I (a) = dµ(ζ ) ψ(ζ ) λ(a, θ)1/2 = 0.

(9.59)

S2

Thus, for fixed a, the WT I (a) of the unit function is constant, and essentially negligible for a 1. Significant values appear only for a > 2, and these scales are irrelevant for the analysis of signals such as contours. This behavior has been checked numerically [35], with the familiar DOG wavelet ψG(α) , discretized on a 128 × 128 grid. Because of the discretization, however, the function I (,, a) ≡ I (θ, a) does depend on a, but very little, and its value remains extremely small. As a matter of fact, for a < 0.1, the WT of ι is numerically negligible over the whole sphere, and may be taken as zero to a very good approximation. As a consequence, the spherical CWT does have the familiar local filtering effect, provided small scales are considered. This will be confirmed by the examples below. Once again, we see that the CWT is useful only as a local analysis.

9.2.5.3

Examples of spherical wavelet transforms As a first example, we analyze in Figure 9.12 an academic picture, namely, (the characteristic function of) a spherical triangle on S 2 , with one of the corners sitting at the North Pole. The triangle is given by 0◦ θ 70◦ , 0◦ ϕ 90◦ and is discretized on a 128 × 128 grid in (θ, ϕ). We use again the spherical Gaussian wavelet ψG(α) , for α = 1.25, discretized on the same grid. According to the admissibility analysis presented above (Figure 9.10), the wavelet is 95 % - admissible on the scale interval a ∈ [0.033, 29.27]. Thus we can evaluate the spherical CWT of this picture for various scales in the allowed range, and we have chosen four successive scales from a = 0.5 to a = 0.035. Figure 9.12 shows that the spherical WT behaves here exactly as, in the flat case, the WT of the characteristic function of a square, as shown in Figure 4.1, in Section 4.1.1. For large a, the WT sees only the object as a whole, thus allowing us to determine its position on the sphere. When a decreases, increasingly finer details appear; in this simple case, eventually only the contour remains, and it is perfectly seen at a = 0.035. The transform vanishes in the interior of the triangle, as it should, only the “walls” remain, with a negative value (black) just outside, a zero-crossing right on the boundary and a sharp positive maximum (white) just inside. In addition, each corner gives a neat peak, which is positive, since the corner is convex [13]. Notice that the three corners are alike, so that indeed the poles play no special rˆole in our spherical WT, contrary to what occurs often in the classical spherical analysis based on spherical harmonics [Fre97,Fre99,169,294,319]. On the other hand, the scales are different in the four cases, since the amplitude of the transform diminishes for decreasing values of a.

1

0.5

Z

0

−0.5

−1 −1 −1

−0.5

−0.5

0

0 0.5

0.5 1

1

X

Y

(a) 0.3

0.2

0.25 1

1 0.2

0.5

0.15

0.5

0

Z

Z

0.1

0.15

0.1

−0.5

− 0.5

0.05

−1 −1

0

−1 −1

−1

−0.5

−0.5

0

0.05

−0.5

0

0 0.5 Y

X

0

−1

0 0.5

0.5 1 1

−0.5

0

Y

0.5 1 1

−0.05 X

(c)

(b) 0.07

0.02

0.06 1

1

0.05 0.04

0.5

0.015

0.5

0.01

0.02 0.01

−0.5

Z

Z

0.03 0

0

0.005

−0.5

0

0 −1 −1

−0.01 −0.5

−0.5

0

0 0.5 Y

0.5 1 1

(d)

X

−1

−0.02 −0.03

−1 −1

−0.005 −1

−0.5

−0.5

0

0

0.5 Y

0.01

0.5 1 1

X

(e)

Fig. 9.12. Spherical wavelet transform of the characteristic function of a spherical triangle with apex at the North Pole, 0◦ θ 70◦ , 0◦ ϕ 90◦ , obtained with the difference wavelet (9.32), at scale a = 0.25. (a) Original image. The transform is shown at four successive scales, (b) a = 0.5; (c) a = 0.2; (d) a = 0.1; and (e) a = 0.035. As expected, it vanishes inside the triangle, and presents a “wall” along the contour, with sharp peaks at each summit. Notice that the amplitudes are different in the four cases (from [35]).

330

Higher-dimensional wavelets

(a)

(b)

(c)

(d)

Fig. 9.13. Spherical wavelet transform of the spherical map of the European area, computed with the spherical DOG wavelet for α = 1.25. (a) The original picture; (b) wavelet transform at a = 0.032; (c) the same at a = 0.016; (d) the same at a = 0.0082 (from [35]).

As a second, real life example, we present in Figure 9.13 the wavelet transform of a significant piece of the terrestrial globe, covering Europe, Greenland and North Africa. As before, we use the spherical DOG wavelet ψG(α) for α = 1.25. The transforms are shown again at three successive scales, a = 0.032, 0.016, 0.0082 (the grid used here is finer than the one used in the previous examples, so that smaller values of a are admissible). As expected, the resolution improves with diminishing a. However, at a = 0.0082, the discretization grid used for the computation of the transform coincides with that of the original picture, so that one sees exactly the same artifacts, such as a closed strait of Gibraltar, an unresolved complex Corsica–Sardinia, ragged coastlines, etc. Of course, we cannot hope to improve on the resolution of the original!

331

9.2 Wavelets on the 2-sphere and other manifolds

Fig. 9.14. Spherical CWT of the spherical triangle at scale a = 0.03, using spherical Morlet wavelet

(9.46): (a) θ = 0 and (b) θ = π/2 (from [35]).

As a final example, we take again the spherical triangle of Figure 9.12(a) and analyze it with the spherical Morlet wavelet (9.46). The results confirm our analysis: the spherical CWT is able to detect the local orientation of the edges of the triangle (Figure 9.14).

9.2.6

Extension to other manifolds The whole construction made so far extends almost verbatim to the (n − 1)-dimensional sphere S n−1 = SO(n)/SO(n − 1), with help of a similar class I representation of the generalized Lorentz group SOo (n, 1) [23]. Although the spheres are the manifolds on which a CWT is most desirable for applications, the mathematical analysis made here invites to consider other manifolds with similar geometrical properties. We take first n = 3. The sphere S 2 = SO(3)/SO(2) is a compact Riemannian symmetric space of constant curvature κ = 1. It has a noncompact dual, H 2 = SOo (2, 1)/SO(2), of constant curvature κ = −1 [Hel78]. H 2 is a two-sheeted hyperboloid, symmetric around the x3 -axis. Duality corresponds to the fact that SO(3) and SOo (2, 1) are the two real forms of the complex group SO(3)C ∼ S L(2, C). Exactly as in the case of the sphere, we can perform a stereographic projection from the South Pole onto the equatorial plane x3 = 0 (or, equivalently, to the plane tangent at the other pole). Then maps the upper sheet H+2 onto the interior D+ of the unit disk, and the lower sheet H−2 onto the exterior. The domain D+ , called the Lobachewskian disk, is conformally equivalent to H+2 , and both manifolds have SOo (2, 1) as isometry group.

332

Higher-dimensional wavelets

As we did for the sphere, dilations on H+2 may be obtained by lifting dilations in the equatorial plane by inverse stereographic projection. The resulting map has all the required properties for a dilation, but does not come directly from a linear group action. Thus it can only be used for constructing wavelets on H+2 if one puts it by hand. It remains to obtain a suitable representation of the resulting set SOo (2, 1) · R+ ∗ in L 2 (H+2 , dµ), where dµ is the SOo (2, 1)-invariant measure, and to show that it is square integrable in a suitable sense. Now this suggests a further generalization. In both cases, S 2 as well as H 2 , the unit disk, image of one sheet or one hemisphere, is a classical domain. Also the stereographic projection has a group-theoretical origin [Per86]. This paves the way to the generalization of the CWT to a whole class of homogeneous spaces (Riemannian symmetric spaces). For instance, S n−1 = SO(n)/SO(n − 1)

and

H n−1 = SOo (n − 1, 1)/SO(n − 1)

are dual Riemannian symmetric spaces, with constant curvature κ = ±1, respectively. Again SO(n) and SOo (n − 1, 1) are two real forms of the complexified group SO(n)C . Their isometry groups are SO(n) and SOo (n − 1, 1), respectively, so that suitable representations of the generalized Lorentz group SOo (n, 1) should provide the corresponding wavelets. While this has been obtained explicitly in the case of the (n − 1)sphere S n−1 in [23], the problem is still open in the noncompact case, i.e., the hyperboloid H n−1 .

9.3

Wavelet approximations on the sphere The central theme of approximation theory is the representation of a function by a truncated series expansion into a family of basis functions, for instance, the elements of a frame. Thus, in the flat case, one- or two-dimensional wavelets are widely used for approximation in various function spaces [Mal99]. The crucial advantage is their multiresolution character, which is optimally adapted to local perturbations. A natural framework is given by the Lebesgue spaces L p (Rn ), 1 p < ∞. One of the reasons is that approximation is often formulated in terms of convolution with an approximate identity, and many useful convolution identities are available in L p . Therefore, in order to apply these considerations to the sphere S 2 , it is necessary to have a good notion of convolution on S 2 . For that purpose, it is useful to represent the sphere as the quotient SO(3)/SO(2), since the convolution machinery extends almost verbatim to locally compact groups, and then partly to homogeneous spaces. For the convenience of the reader, we have collected in the Appendix the main definitions and essential properties of convolution on a locally compact group. In what follows, we will need two different cases. For simplicity, we write L 2 (SO(3)) ≡ L 2 (SO(3), d,), where d, is the Haar measure on SO(3), and L p (S 2 ) ≡ L p (S 2 , dµ).

333

9.3 Wavelet approximations on the sphere r

If f ∈ L 2 (SO(3)) and g ∈ L 1 (S 2 ), then their spherical convolution is the function on S 2 defined as f (,) g(,−1 ζ ) d,. (9.60) ( f % g)(ζ ) = SO(3)

Then f % g ∈ L 2 (S 2 ) and one has f % g2 f 2 g1 , r

(9.61)

where the norms refer to the corresponding spaces. If f ∈ L 2 (S 2 ) and g ∈ L 1 (S 2 ), their spherical convolution is the function on SO(3) defined as (f % g)(,) = f (,−1 ζ ) g(ζ ) dµ(ζ ). (9.62) S2

Then f % g ∈ L 2 (SO(3)) and f % g2 f 2 g1 ,

(9.63)

Here, however, we are only interested in functions on the sphere S 2 , that is, functions on SO(3) that are SO(2)-invariant. In particular, we will deal mostly with axisymmetric, or zonal, functions on S 2 , that is, functions of θ alone. Thus, we will focus on elements of L 2 ([−1, +1], dt), where t = cos θ, for which the Fourier series reduces to a Legendre expansion: ψ(t) =

∞ 2l + 1 l=0

= 2π ψ(l)

4π

+1

−1

Pl (t), ψ(l)

dt Pl (t) ψ(t) =

*

4π ψ(l, 0), 2l + 1

m) ≡ Y m |ψ denotes the Fourier coefficient of ψ. If f is an axisymmetric where ψ(l, l function, the spherical convolution (9.62) takes a simpler form [Fre97]: Proposition 9.3.1 Let f and g be two measurable functions on S 2 . If f is axisymmetric, the spherical convolution of f and g is a function on S 2 , which can be written:

dµ(ζ ) f ( ζ · ζ ) g(ζ ), (9.64) ( f % g)(ζ ) = S2

ζ is the R3 scalar product of unit vectors of directions ζ and ζ . where ζ · The proof amounts to a straightforward application of harmonic analysis (Fourier series) on S 2 and of the addition theorem for spherical harmonics, to the effect that f (,−1 ζ ) = f ( ζ · ζ ), where ζ ≡ ,˙ ∈ S 2 denotes the left coset of , ∈ SO(3) (see the geometrical discussion in Section 9.2.4.1).

334

Higher-dimensional wavelets

Now we may turn to the approximation problem proper. As in the Euclidean case [Lie97,Ste71], a convenient technique is to perform a convolution with a smoothing kernel, that acts as an approximate identity, that is, a kernel which tends to the identity (δ function) as the parameter goes to 0. For the sake of simplicity, we will only deal with zonal kernels, following mainly [Fre97]. Definition 9.3.2 Let Kτ , τ ∈ (0, τo ], τo ∈ R+ ∗ , be a family of elements of 1 ( L ([−1, +1], dt) satisfying Kτ (0) = 1. The functional Sτ [ f ] defined by Sτ [ f ] = Kτ % f,

f ∈ L p (S 2 ),

1 p < ∞,

is called a singular integral. It is called an approximate identity of L p (S 2 ) if lim f − Sτ [ f ] p = 0,

τ →0 τ >0

∀ f ∈ L p (S 2 ) .

(9.65)

The following theorem characterizes those spherical kernels which are associated with an approximate identity. Theorem 9.3.3 Let {Kτ } be a uniformly bounded spherical kernel, that is, there exists a constant M, independent of τ , such that +1 dt |Kτ (t)| M, ∀τ ∈ (0, τo ] . −1

Then the associated singular integral is an approximate identity of L p (S 2 ) if and only if (τ (n) = 1, lim K

τ →0 τ >0

∀n 0.

(9.66)

A proof may be found in [Fre97]. A particularly interesting case is given by positive definite kernels. In this case, since |Pl (t)| 1, {Kτ } is uniformly bounded, with (, (0). The following theorem gives a nice characterization of bound M = supτ ∈(0,τ0 ] K approximate identities associated with positive kernels. Theorem 9.3.4 Let {Kτ }, τ ∈ (0, τo ], be a positive kernel associated with a singular integral of L p (S 2 ). Then each of the following conditions is equivalent to (9.65) and (9.66), which means that {Kτ } is the kernel of an approximate identity: (τ (0) = 1, (i) lim K τ →0 τ >0

δ

(ii) lim

τ →0 τ >0

−1

dt Kτ (t) = 0, δ ∈ (−1, +1).

335

9.3 Wavelet approximations on the sphere

Condition (ii) is in fact a constraint on the localization of the kernel, as we shall see in an explicit example below (see Figure 9.15). Approximate identities are a very useful tool for harmonic analysis on the sphere and many applications can be found in [Fre97]. Thus it is gratifying that the spherical wavelet transform naturally yields a systematic way of deriving approximate identities on the sphere. This is actually an interesting way of handling functions on the sphere, because it allows us to represent information by means of localized, and hierarchically organized, coefficients. With such a representation, a local modification of the function would only result in a slight local perturbation of the original coefficients, a definite advantage over Fourier series (exactly as in the case of flat space). As a matter of fact, many examples of approximate identities are given in the textbook of Freeden et al. [Fre97], and they are applied extensively by these authors to geophysical data [Fre99]. Most of these examples are based on families of kernels indexed by a parameter which behaves like a dilation. However, since the latter is introduced directly as a parameter in those kernels, there is no unique way of generating approximate identities, as in Rn [Ste71]. But this problem disappears naturally if one uses the spherical dilation, since, as we shall see, the dilation operator generates an approximate identity in L 2 (S 2 ). However, we have to modify it first and adapt it to the L 1 environment. Using the notation of Section 9.2.2, we define, instead of Da , a new dilation operator: (D a f )(ζ ) ≡ f a (ζ ) = λ(a, θ) f (ζ1/a ),

(9.67)

and this operator clearly conserves the L 1 norm. Notice that the situation is more complicated here than in the flat case. There, indeed, changing the dilation operator from L 2 to L 1 simply amounts to changing the power of a in front of the transform [29]. Here, one replaces the factor λ(a, θ)1/2 by its square λ(a, θ), but this modifies the CWT itself in a nontrivial way. In particular, the admissibility condition (9.25) becomes ∞ da (a 8π 2 |ψ (l, m)|2 < c . (9.68) 2l + 1 |m| l 0 a First, we notice that our new dilation operator does not change the mean of a function, thus simplifying the statement of Proposition 9.2.4. Proposition 9.3.5 If ψ ∈ L 1 (S 2 ), then dµ(ζ ) ψ a (ζ ) = dµ(ζ ) ψ(ζ ). S2

(9.69)

S2

The proof reduces to a simple change of variables, taking into account the cocycle relation (9.23).

336

Higher-dimensional wavelets

Acting with the new dilation D a on a suitable function, one can now easily construct an approximate identity, as shown in the next proposition. Proposition 9.3.6 Let f ∈ C([−1, +1]), with f (0) = 1. Then the family { f a ≡ D a f, a > 0} is the kernel of an approximate identity. In view of Theorem 9.3.3, the proof consists in two steps. First, one shows that the fam +1 ily { f a } , a ∈ (0, 1], is uniformly bounded, which is obvious since −1 dt | f a (t)| = f 1 . Next, it remains to verify that f a (l) = 1, lim (

a→0 a>0

and this is done again by a change of variables and applying the cocycle relation (9.23) [35]. This technique is applied in Figure 9.15 to a zonal function of Gaussian shape, namely the mother wavelet of the spherical DOG wavelet, φG (θ, φ) = exp(− tan2 (θ/2)), θ ∈ [−π, π]. One clearly sees how dilation localizes the kernel better and better as a → 0. In the L 1 formalism, we recall from [29] that the necessary condition for admissibility becomes a genuine zero mean condition, exactly as in the flat case: 12

φ (θ, ϕ)

a = 0.7

G

a = 0.5

10

a = 0.3 8

6

4

2

0

π

−3

−2

−1

0

θ

1

2

3

−π

Fig. 9.15. Kernel of an approximate identity obtained by dilating a Gaussian mother function with scaling factor a = 0.7, 0.5 and 0.3 (from [35]).

337

9.3 Wavelet approximations on the sphere

S2

dµ(θ, ϕ) ψ(θ, ϕ) = 0,

(9.70)

Correspondingly, in view of Proposition 9.3.5, the difference wavelet ψφ(α) given in (9.32) is replaced by ψ˘ φ(α) (θ, ϕ) = φ(θ, ϕ) − D α φ(θ, ϕ)

(α > 1).

Now, combining the modified dilation operator D a with the usual rotation operator R, , we define a new set of spherical wavelets, starting from an admissible ψ, namely, ψ,a ≡ R, D a ψ = R, ψ a . Accordingly, we redefine as follows the spherical wavelet transform of a signal s ∈ L 2 (S 2 ): S˘ψ (,, a) =

S2

dµ(ζ ) ψ,a (ζ ) s(ζ ).

(9.71)

In particular, if the wavelet ψ is zonal, we get S˘ψ (ζ, a) =

S2

dµ(ζ ) ψ a ( ζ · ζ ) s(ζ ).

(9.72)

In addition, the correspondence between spherical wavelets and their stereographic projections on the tangent plane (Proposition 9.2.5) is modified, as follows. Proposition 9.3.7 If ψ ∈ L 2 (R2 ) is an admissible 2-D Euclidean wavelet, and I −1 denotes the inverse stereographic projection (9.44), then the function (1 + cos θ )−1 I −1 ψ is an admissible spherical wavelet for the transform defined with the L 1 normpreserving dilation operator D a . After this preparation, we proceed to show that the spherical CWT admits a reconstruction formula, valid in the strong L 2 topology, exactly as the usual CWT in Rn . As in the flat case, described in Section 2.6.1, we may distinguish between a bilinear and a linear formalism [Tor95]. But there is a crucial difference. In the flat case, it is advantageous, but not compulsory, to treat the large scales or low frequencies separately, in terms of a scaling function. Here, however, we are forced to do it. The reason is that, geometrically, only small scales are relevant and lead to the expected filtering behavior, as discussed in Section 9.2.5.2. We arbitrarily choose a = ao as reference scale and define the scales a > ao as large. Notice that we recover here, in precise mathematical terms, the argument behind the mixed spherical harmonics/spherical wavelets representation of object surfaces from [89], discussed in Section 9.2.4.4. Let us begin with the bilinear analysis. Given a wavelet ψ ∈ L 1 (S 2 ), we define the corresponding scaling function ≡ (ao ) by its Fourier coefficients:

338

Higher-dimensional wavelets

m)|2 = | (l,

∞

ao

da (a |ψ (l, m)|2 , a

l 1,

(9.73)

1 (9.74) 8π 2 (the integral in (9.73) converges in virtue of the admissibility condition (9.68) satisfied by ψ). Of course, (9.73) does not define the function uniquely. We can, for instance, m) 0, ∀ l, m, as in [Fre97]. Corresponding to (9.71), assume in addition that (l, we define the large-scale part of a signal s as o) ˘ (,, ao ) = / dµ(ζ ) (a (9.75) , (ζ ) s(ζ ), 0)|2 = | (0,

S2

(ao ) −1 o) where we have put (a (, ζ ). , (ζ ) ≡

Theorem 9.3.8 (Bilinear analysis) Let ψ ∈ L 1 (S 2 ) be a spherical wavelet and let ≡ (ao ) , ao > 0, denote the associated scaling function. Assume the following two conditions are satisfied: r for all l = 1, 2, . . . , 8π 2 ∞ da (a |ψ (l, m)|2 = 1, (9.76) 2l + 1 |m| l 0 a r

for all $ ∈ (0, ao ), there is a constant M > 0, independent of $, such that ao da ψ a 2 M. a $

Then, for all s ∈ L 2 (S 2 ), we have the equality ao da o) ˘ (,, ao ) (a s= d, S˘ψ (,, a) ψ,a + d, / , , a SO(3) 0 SO(3)

(9.77)

(9.78)

˘ is the large-scale where S˘ψ is the spherical CWT of s with respect to the wavelet ψ, / 2 2 part of s and the integral is understood in the strong sense in L (S ). Proof . We consider the first term in (9.78). Since ψ ∈ L 1 (S 2 ) and s ∈ L 2 (S 2 ), Young’s convolution inequality (9.62) shows that S˘ψ ∈ L 2 (SO(3)). As in the flat case, we define the infinitesimal detail at scale a: a d (ζ ) = d, S˘ψ (,, a) ψ,a (ζ ). SO(3)

This is a convolution on SO(3) and Young’s inequality (9.61) shows that d a ∈ L 2 (S 2 ). Explicitly, we have d a (ζ ) = dµ(ζ ) s(ζ ) d, ψ a (,−1 ζ ) ψ a (,−1 ζ ). (9.79) S2

SO(3)

339

9.3 Wavelet approximations on the sphere

Using the relation ψ a (,−1 ζ ) =

∞

l (a (l, n) Ylm (ζ ), Dmn (,) ψ

(9.80)

l=0 |m| l |n| l l where Dmn (,) denotes a Wigner function [Ros57,Tal68], we find

(a (l, n) ψ (a (l , n ) d a (ζ ) = dµ(ζ ) s(ζ ) Ylm (ζ ) Ylm (ζ ) ψ S2

lmn l m n

× SO(3)

l (,) D l (,). d, Dmn m n

Using the orthogonality of Wigner functions and the addition theorem for spherical harmonics (see Section A.4), this gives: ∞ a (a (l, m)|2 . d (ζ ) = 2π ζ · ζ |ψ dµ(ζ ) s(ζ ) Pl S2

l=0 |m| l

Now consider the following expression: ao da a d (ζ ) s$, (ζ ) = a $ ao ∞ da

(a (l, m)|2 . = 2π ζ · ζ |ψ dµ(ζ ) s(ζ ) Pl a l=0 |m| l S2 $ In virtue of condition (9.77), the double summation on the right-hand side of this equation is absolutely and uniformly convergent, since it is majorized by ao ao ∞ da da (a (l, m)|2 = |ψ ψ a 2 . a a $ $ l=0 |m| l Now let us introduce the quantity: ∞ ao da (a (ao ) 2 K$ (t) = 2π |ψ (l, m)| Pl (t), a $ l=0 |m| l so that s$(ao ) = K$(ao ) % s. By (9.77), we see that K$(ao ) ∈ L 1 ([−1, +1]), for all 0 < $ ao , and K$(ao ) 1 2π M. Next, we show in the same way that the second term in (9.78) equals H(ao ) % s, where H(ao ) (t) = 2π

∞

m)|2 Pl (t). | (l,

l=0 |m| l

Again, H(ao ) ∈ L 1 ([−1, +1]). Finally, we define the kernel K$ = K$(ao ) + H(ao ) , which also belongs to L 1 ([−1, +1]). Condition (9.77) shows that K$ is a uniformly bounded

340

Higher-dimensional wavelets

m), we deduce kernel. In addition, from (9.76) and the definition (9.73)–(9.74) of (l, the following constraint on its Legendre coefficients : ao 8π 2 da (a 2 2 ( lim K$ (l) = |ψ (l, m)| + | (l, m)| $→0 2l + 1 |m| l 0 a 8π 2 ∞ da (a |ψ (l, m)|2 = 1, l 1, 2l + 1 a |m| l 0 = 2 8π | (0, 0)|2 = 1, l = 0. Then Theorem 9.3.3 shows that K$ is the kernel of an approximate identity, which proves the strong convergence in L 2 (S 2 ) of the approximation: lim (K$ % s) = s.

$→0

As a check of the reconstruction formula (9.78), let us consider the unit function ι. Contrary to the case of the L 2 formalism, the L 1 -normalized CWT of ι vanishes identically, as a consequence of Proposition 9.3.5: a ˘Iψ (,, a) = dµ(ζ ) ψ (ζ ) = dµ(ζ ) ψ(ζ ) = 0. S2

S2

Hence only the second term, the large-scale part, subsists in (9.78). Using again the expansion (9.80), we find successively: 0), dµ(ζ ) (,−1 ζ ) = (0, I˘ (,, ao ) = S2

and, for (9.78), ι(ζ ) = (0, 0)

0)|2 = 1. d, (,−1 ζ ) = 8π 2 | (0, SO(3)

This result shows that the large-scale part of a signal must be treated separately, because constant functions on the sphere are square integrable, and hence must be reconstructible, although their CWT vanishes identically. In practice, of course, large scales should be irrelevant, since wavelet analysis is local, and we expect the second term in (9.78) to be numerically negligible (that is, one must choose ao large enough for this to be true). Theorem 9.3.8 applies, in particular, to a zonal wavelet. The only change is the parameter space of the spherical CWT which takes the form of the product S 2 × R+ ∗, −1 with the measure a da dµ(ζ ). A further simplification yet is to consider a singular reconstruction wavelet and build a framework similar to the Morlet linear analysis. As in the bilinear case, we begin by defining, through its Legendre coefficients, a scaling function φ ≡ φ (ao ) that takes care of the large scales :

341

9.3 Wavelet approximations on the sphere

= φ(l)

∞

ao

da (a ψ (l), a

l 1,

= 1. φ(0)

(9.81) (9.82)

The corresponding large part of a signal s is then σ˘ φ (ζ, ao ) = dµ(ζ ) φ( ζ · ζ ) s(ζ ).

(9.83)

S2

In these notations, the linear reconstruction formula is given by the following theorem. Theorem 9.3.9 (Linear analysis) Let ψ ∈ L 1 (S 2 ) be a zonal spherical wavelet satisfying the following two conditions: r for all l = 1, 2, . . . , ∞ da (a (9.84) ψ (l) = 1, a 0 r

for all $ ∈ (0, ao ), ∞ 2l + 1 ao da (a ψ (l) < ∞ . 4π a $ l=0

(9.85)

Then, for all s ∈ L 2 (S 2 ), we have the equality ao da ˘ Sψ (ζ, a) + σ˘ φ (ζ, ao ), s(ζ ) = a 0 the integral being again understood in the strong sense in L 2 . Proof . The same arguments as in the proof of Theorem 9.3.8 show that the partial sum ao da ˘ (ao ) s$ (ζ ) = Sψ (ζ, a) a $ belongs to L 2 (S 2 ). Expanding this expression and adding the large-scale term, we find ao da a s$ (ζ ) = dµ(ζ ) ψ (ζ · ζ ) s(ζ ) + σ˘ φ (ζ, ao ) a S2 $ ao da a

= dµ(ζ ) s(ζ ) ψ (ζ · ζ ) + φ(ζ · ζ ) a S2 $ ao ∞ 2l + 1 da (a

ψ (l) + φ(l) Pl ( = dµ(ζ ) s(ζ ) ζ · ζ ) 4π a S2 $ l=0 = (κ$ % s)(ζ ),

342

Higher-dimensional wavelets

where we have used (9.85) and set ao ∞ 2l + 1 da (a ψ (l) + φ(l) Pl (t). κ$ (t) = 4π a $ l=0 The Legendre coefficients of this kernel are ao da (a (l) = ψ (l) + φ(l). κ( $ a $ As in the proof of Theorem 9.3.8, we deduce from condition (9.84) that lim$→0 κ( $ (l) = 1, ∀ l = 0, 1, . . . . Thus we have again an approximate identity, which allows us to conclude that lim s − κ$ % s2 = 0 .

$→0

The conclusion of this analysis is that our spherical CWT, with the modified dilation operator D a , leads to the same approximation scheme as that developed by Freeden [Fre97,Fre99]. The present approach, however, has the additional advantage of giving a clear geometric meaning to the approximation parameter a. By the same token, it intuitively explains the validity of the Euclidean limit established in [29]. Indeed, taking a → 0 means going to the pointwise limit where curvature becomes unimportant, that is, going to the tangent plane and recovering the flat CWT.

10

Spatio-temporal wavelets and motion estimation

10.1

Introduction We live in a world where objects (cars, animals, men, birds, aeroplanes, the Sun, etc.) that surround us are constantly in relative motion. One would like to extract the motion information from the observation of the scene and use it for various purposes, such as detection, tracking and identification. In particular, tracking of multiple objects is of great importance in many real world scenarios. The examples include traffic monitoring, autonomous vehicle navigation, and tracking of ballistic missile warheads. Tracking is a complex problem, often requiring to estimate motion parameters – such as position, velocity – under very challenging situations. Algorithms of this type typically have difficulty in the presence of noise, when the object is obscured, in situations including crossing trajectories, and when highly maneuvering objects are present. Most motion estimation (ME) techniques such as the ones based on block matching, optical flow, and phase difference [Jah97,280,281] assume that the object is constant from frame to frame. That is, the signature of the object does not change with time. Consequently, these techniques tend to have difficulty handling complex motion, particularly when noise is present. The time-dependent continuous wavelet transform (CWT) is attractive as a tool for analysis, in that important motion parameters can be compactly and clearly represented. The CWT maps a given image sequence into a six-dimensional representation in which position, time, scale, and velocity (speed and orientation of the velocity) are explicit parameters. In effect, this transform provides a multiscale description of motion – the conventional continuous wavelet transform in two dimensions – with additional operators to provide control over the speed and orientation (i.e., velocity) in space– time [158]. It is a multidimensional filtering in all six variables, position, time, scale, speed, and orientation. Several authors have already shown how band-pass filtering can be used to evaluate the speed in an image. A large class of models of human motion sensing use this approach [3,218,371]. The early mechanisms involved in human perception of motion appeared to be sensitive to spatial and temporal frequencies. Some neurons of the

343

344

Spatio-temporal wavelets and motion estimation

visual cortex were found to respond best when they were stimulated with temporally modulated stimuli, the temporal frequency being included in a given range. One of these neurons may respond to the presentation of a slowly moving bar, and yet not respond to the same bar moving faster in its receptive field. The organization of these filters in spatial and temporal frequencies show some special characteristics related with our spatio-temporal sensitivity. In fact, our visual system seems to make a trade-off between its spatial and temporal resolution. For instance, if one stares at a train leaving the station, at low speed it is still possible to read the destination of the cars, but at a higher speed it becomes impossible to read those details, only global shapes are still available for our vision. Actually, the interpretation of this example is complicated by the fact that our eyes (and head) may be following the train. The organization of this chapter is parallel to that of Chapter 2. We start by describing the spatio-temporal signals and motions, emphasizing that motion is in fact orientation in space–time. Next we describe the elementary operations applied to spatio-temporal filters for motion extraction. Five of these operations are the same as the ones introduced in the 1-D and the 2-D CWT (but in a spatio-temporal setting). The sixth, called speed tuning operator, is different and acts on space and time in a way that allows us to cope with the presence of motion in a spatio-temporal signal. This is then cast into the group-theoretical language, in particular, we define the appropriate unitary irreducible representation. Since the latter is square integrable, wavelets in the usual sense may be constructed. Thus we define the (2+1)-D CWT and give its properties. Finally, we describe the CWT tracking algorithm, referred to here as the Mujica–Murenzi–Leduc–Smith (MMLS) tracking algorithm, and apply it to two test sequences that reflect difficulties associated with noise, accelerated motion, temporary occlusion, and time-varying signatures. The first scene corresponds to four moving objects with linear and nonlinear motion. The experiments are performed both for noiseless and for noisy cases. The results are compared with those of block matching algorithms (BMA). The second sequence corresponds to a circularly moving object under the conditions of increasing velocity and acceleration.

10.2

Spatio-temporal signals and their transformations We will consider (2+1)-dimensional signals s (image sequences) of finite energy, represented by square integrable complex valued functions on R2 × R, i.e., s ∈ L 2 (R2 × R, d 2 x dt): d 2 x dt |s( x , t)|2 < ∞ , (10.1) s2 = R2 ×R

where x is a vector of R2 and t is the time.

345

10.2 Spatio-temporal signals and their transformations

In practice, a black and white sequence of images will be represented by a bounded non-negative function: 0 s < M < ∞. As in the spatial case, we will also consider as admissible signals some generalized functions (distributions), such as a delta function δ( x − vo t) or a plane wave exp i(ko · x + ωo t). The Fourier transform of s is defined, as usual, by −3/2 s(k, ω) = (2π ) d 2 x dt e−i(k·x −ωt) s( x , t), (10.2) R2 ×R

where k is the spatial frequency, ω is the temporal frequency, and k · x is the Euclidean scalar product. If s moves with a constant speed vo , the resulting signal s may be written as s( x , t) = s( x − vo t, t),

(10.3)

and its Fourier transform becomes ω) = ω − k · vo ) . s(k, s(k,

(10.4)

This constant speed motion does not affect the spatial frequency of the signal, but the components of its Fourier transform corresponding to the spatial frequency k are translated by −k · vo in the direction of the ω axis. This operation transforms the plane defined by ω = 0 into a new plane defined by ω = k · vo . Consider now a static signal which, in Fourier space, is roughly concentrated around the plane defined by ω = 0. When the same signal moves with constant speed vo , its Fourier transform is concentrated around the plane defined by ω − k · vo = 0

voT

1

k ω

=0

(10.5)

where voT denotes the transposed vector. Thus the velocity plane defined by (10.5) is perpendicular to the velocity vector vo . Therefore a filter which is concentrated around this plane will be sensitive to the moving components of the image which have the same speed vo . This idea is the cornerstone of the band-pass filtering approach to visual motion perception. It is directly used in spatio-temporal energy models, such as that of Adelson and Bergen [3]. Their filters can see a pattern moving with a fixed speed if the ratio of their temporal and spatial mean frequency fits with the value of the speed and if the direction of the speed and the spatial frequency are close enough. As in Chapter 2, we begin by introducing the elementary operations that we want to apply to spatio-temporal signals for motion extraction. From our past experience, we know, however, this is equivalent to transforming the filters, in particular, the wavelets, and to analyzing a given object with help of the transformed filter (this is the so-called

346

Spatio-temporal wavelets and motion estimation

passive point of view described in Section 7.1). Thus, from now on, we consider transformations to be applied to a given spatio-temporal filter (in particular, a spatio-temporal wavelet), in order to have a basis on which to decompose a given spatio-temporal signal. As we will see in next section, appropriate filters, which are wavelets in our case, must vanish at zero temporal frequency and zero spatial frequency, (10.35)–(10.36). A fortiori, a good wavelet for motion analysis will have its support in the Fourier domain concentrated in a convex cone in the half space R2 × R+ ∗ , with apex at the origin. It is also supposed to be concentrated around the plane corresponding to a fixed speed vector vo . Let us consider a spatio-temporal filter ψ (of finite energy, as usual). We consider the following elementary operations. r Spatio-temporal translations The wavelet is shifted to a given point of space and time (R2 × R). This transformation in the spatio-temporal Fourier domain: is denoted by T in direct space and T t − τ) , x , t) = ψ( x − b, (Tb,τ ψ)(

x +ωτ ) ψ)( k, ω) = e−i(k· ω). ψ(k, (T b,τ

(10.6)

τ ), is used to detect the location of obThis transformation, with parameter q = (b, k, ω), remains jects. In the Fourier domain, the wavenumber–frequency spectrum, ψ( concentrated around the vo velocity plane, only a linear phase term is introduced. r Rotation This transformation, denoted by Rθ , rotates the wavelet in spatial coordinates around the temporal (or frequency) axis. In this way, filters can be tuned to a particular orientation associated with the velocity vo . It is defined by x , t) = ψ(r−θ ( x ), t) , (Rθ ψ)(

θ ψ)( k, −θ (k), ω) = ψ(r ω), (R

where rθ is the usual 2 × 2 rotation matrix cos θ − sin θ rθ = , 0 θ < 2π. sin θ cos θ

(10.7)

(10.8)

The effect of this transformation is to change the vo -plane into a new velocity plane associated with velocity rθ ( vo ), that is, · vo = k · rθ ( vo ). ω = r−θ (k)

(10.9)

Thus, the parameter θ allows for orientation changes and is used to estimate velocities in the context of motion estimation, together with the speed tuning parameter c, to be defined next. r Scaling in wave-number– This transformation, denoted by D in space–time domain and D frequency domain, is the analog of the one we used in Chapter 2 and is defined by

347

10.2 Spatio-temporal signals and their transformations

(Da ψ)( x , t) = a −3/2 ψ(a −1 x, a −1 t),

a ψ)( k, ω) = a 3/2 ψ(a k, aω), (D

a > 0. (10.10)

The dilation preserves the norm of L 2 (R2 × R, d 2 x dt). Since it operates in the same way on both time and space, the speed of the object is not affected by the dilation. In the Fourier domain, this means that the dilation keeps the wavelet on the same constant-speed plane. r Speed tuning transformation Unlike the dilation and the shift transformations, which do not affect the concentration of the energy of the wavelet around a plane of constant speed in the Fourier space, the speed tuning transformation allows the concentration of the energy of the wavelet to move from one velocity plane with speed | vo | to a velocity plane with a different speed. in wavenumber The transformation, denoted by # in the space–time domain and # domain, can be seen as an anisotropic scaling on space and time, that is, c ψ)(k, c−β ω), ω) = ψ(cα k, (#

c > 0.

(10.11)

We require the transformation to be unitary and to map the vo -plane into the c vo plane. This results in a system of two linear equations, that fixes α and β. The first constraint gives k, k, c ψ)( 2 ω)2 = (# ω)2 = c−2α+β ψ ψ(

(10.12)

that is, 2α = β. The second constraint gives c−β ω = cα k · vo = k · cα+β vo = k · c vo .

(10.13)

This requires that α + β = 1, which together with the unitarity constraint requires that α = 23 and β = 13 . The speed tuning transformation is thus explicitly defined by (#c ψ)( x , t) = ψ(c−1/3 x, c2/3 t), c ψ)( k, 1/3 k, c−2/3 ω). ω) = ψ(c (#

(10.14) (10.15)

The speed tuning transforms a wavelet which can see a speed vo into a wavelet sensitive to c vo . Therefore, the transformation allows a direct and natural tuning of the value of the speed that the wavelet is able to analyze. In Fourier space, the transformation generates a family of wavelets which are distorted and shifted along hyperbolas defined by the ω. The parameter c of the speed tuning allows to adapt the constancy of the product |k| speed analysis independently of the scale analysis. When analyzing a moving pattern, the dilation may be used to adapt the scale of the wavelet to the spatial extension of the pattern without affecting the speed that the wavelet analysis will detect. Therefore, the τ ) can be directly values of the wavelet transform for a fixed point in time and space (b, interpreted in terms of scale (with parameter a) and speed (c and θ).

348

Spatio-temporal wavelets and motion estimation

τ, θ; a, c) ≡ Combining all these operators, one obtains the operator U (b, Tb,τ Rθ Da #c , namely, a −1 c2/3 (t − τ )). τ, θ; a, c)ψ ( x − b), (10.16) U (b, x , t) = a −3/2 ψ(a −1 c−1/3r−θ ( τ, θ; a, c)ψ}, in terms of From the wavelet ψ, one obtains a family of wavelets {U (b, which one can decompose any spatio-temporal signal s.

10.3

The transformation group and its representations Clearly, the transformations presented in the previous section form a group. In this section we will develop the wavelet machinery associated with that group. As in the case of 2-D spatial wavelets studied in Chapter 2 and in Section 7.2, the key to the construction of spatio-temporal wavelets is to have at one’s disposal a unitary representation of this transformation group in the natural space of finite energy signals, namely, L 2 (R2 × R, d 2 x dt). We are looking for a group G acting on space–time R2 × R, whose restriction to space variables coincides with the usual continuous wavelet group in two dimensions, i.e., the similitude group SIM(2), while the restriction to the time variable coincides with the usual continuous wavelet group in one dimension, i.e., the affine group G + aff . Let us consider first the group G 1 = E(2) × R, the Euclidean group (rotations and translations) acting on the (2+1)-dimensional space–time. The elements of G 1 are τ, θ ), where b is the translation vector, τ is the time translation, and θ is the triplets (b, τ, θ) ∈ G 1 acts on the point ( rotation parameter. The element (b, x , t) in the space–time 2 R × R in a natural way: t + τ ), x ∈ R2 , t ∈ R. τ, θ) : ( x ) + b, (b, x , t) → (rθ (

(10.17)

+ Next, we consider the product G 2 = R+ ∗ × R∗ of two dilation groups with the following action on space–time: + (a, c) : ( x , t) → (ac3/2 x, act), a ∈ R+ ∗ , c ∈ R∗ .

(10.18)

This action allows one to define the semidirect product G mv ≡ G 1 G 2 = {g ≡ τ, θ; a, c)}, with multiplication: (b, τ, θ; a, c)(b , τ , θ ; a , c ) = (b + ac1/3rθ (b ), τ + ac−2/3 τ , θ + θ ; aa , cc ). (b, (10.19) The inverse is given by −a −1 c2/3 τ, −θ; a −1 , c−1 ), τ, θ; a, c)−1 = (−a −1 c−1/3r−θ (b), (b,

(10.20)

349

10.3 The transformation group and its representations

0, 0; 1, 1). In 4 × 4 matrix notation, we may write and the unit element is e = (0, (compare (7.37)): 1/3 ac rθ 0 b τ, θ; a, c) ≡ (b, (10.21) ac2/3 τ . 0 0 0 1 The group G mv is locally compact with right Haar measure dµR and left Haar measure dµL : da dc da dc , dµL = d 2 b dτ dθ 4 . (10.22) a c a c In the sequel, we will use systematically the left Haar measure dµL and, correspondingly, we will write 2π ∞ da ∞ dc 2 . (10.23) dµL (b, τ, θ; a, c) ≡ d b dτ dθ a4 0 c G mv R2 R 0 0 dµR = d 2 b dτ dθ

Having identified the appropriate transformation group G mv , we proceed to find a square integrable representation of it, in the Hilbert space of finite energy signals, ac τ, θ; a, c) ∈ cording to the general formalism sketched in Section 7.1. For every g ≡ (b, 2 2 2 G mv , consider the operator U (g) on L (R × R, d x dt) defined in (10.16), namely, a −1 c2/3 (t − τ )). x − b), x , t) = a −3/2 ψ(a −1 c−1/3r−θ ( [U (g)ψ] (

(10.24)

This may be written in the Fourier space as $ % b+ωτ ) (g)ψ (k, ac−2/3 ω). ω) = a 3/2 e−i(k· ψ(ac−1/3r−θ (k), U

(10.25)

Proposition 10.3.1 The operator U defines a unitary representation of G mv in L 2 (R2 × R, d 2 x dt). This is proven by a straightforward verification. defines two unitary irreducible representations of Proposition 10.3.2 The operator U 2 G mv in H+ ≡ H+ (R × R) and H− ≡ H− (R2 × R), respectively, where k, ω) = 0, ω < 0} H+ (R2 × R) = {ψ ∈ L 2 (R2 × R, d 2 x dt) : ψ( k, ω) = 0, ω > 0} H− (R2 × R) = {ψ ∈ L 2 (R2 × R, d 2 x dt) : ψ(

(10.26) (10.27)

with L 2 (R2 × R, d 2 x dt) = H+ (R2 × R) ⊕ H− (R2 × R).

(10.28)

Proof . The proof follows the same line as that of Proposition 2.1.2. Let ψ ∈ H+ be an arbitrary nonzero vector. We are going to show that ψ is cyclic for the representation

350

Spatio-temporal wavelets and motion estimation

U , i.e., the linear span of the orbit of ψ is dense in H+ . Let f be orthogonal to the span of {U (g)ψ}, that is, U (g)ψ| f = 0, ∀ g ∈ G mv . We show that f = 0. We have τ, θ; a, c) ∈ G mv , indeed, for all (b, (g)ψ| U (g)ψ| f = U f −1/3r−θ (k), ac−2/3 ω) d 2 k dω ei(k·b+ωτ ) ψ(ac = a 3/2 R 2 ×R

= 0. −1/3r−θ (k), ac−2/3 ω) ω) vanishes This means that the Fourier transform of ψ(ac f (k, (almost everywhere) for all θ, a, c. Then the transitivity of the action of (θ, a, c) ∈ ω) = 0 (almost everywhere). SO(2) × R∗+ × R∗+ on R2 × R∗+ implies that f (k, The same holds true for H− . Proposition 10.3.3 The representation U is square integrable, that is, there exists a function ψ ∈ L 2 (R × R, d 2 xdt), ψ = 0, such that the matrix ele τ, θ; a, c)ψ|ψ is square integrable with respect to the measure ment U (b, τ, θ; a, c), that is, in the notation of (10.23), dµL (b,

τ, θ; a, c)|U (b, τ, θ; a, c)ψ|ψ|2 < ∞. dµL (b,

I =

(10.29)

G mv

In addition, I = cψ ψ2

(10.30)

where cψ = (2π)

3 R 2 ×R

d 2 k dω |ψ(k, ω)|2 . 2 |ω| |k|

Proof . The result follows from a direct calculation: I =

τ, θ; a, c)ψ|ψ|2 dµL |U (b, G mv

(b, ψ| 2 τ, θ; a, c)ψ| dµL |U b+ωτ ) −1/3r−θ (k), ac−2/3 ω) e−i(k· ψ(k, ω) = dµL a 3 d 2 k dω ψ(ac =

G mv

G mv

×

R2 ×R

R2 ×R

τ ) −1/3r−θ (k ), ac−2/3 ω ) ei(k ·b+ω d 2 k dω ψ(ac ψ(k , ω )

(10.31)

351

10.3 The transformation group and its representations

=

2π

0

∞

0

∞

0

da dc dθ a c

×

2

d bdτ e

2

d k dω R2 ×R

k )·b+(ω−ω −i{(k− )τ }

d 2 k dω

R 2 ×R

R2 ×R

k, k , ω ) ω) ψ( ψ(

−1/3r−θ (k ), ac−2/3 ω ) −1/3r−θ (k), ac−2/3 ω) ψ(ac × ψ(ac = (2π)3

2π

∞

∞

dθ 0

0

0

× = (2π)3

R2 ×R

da dc a c

−1/3r−θ (k), k, ac−2/3 ω)|2 |ψ( ω)|2 d 2 k dω |ψ(ac

R2 ×R

k, ω)|2 d 2 k dω|ψ(

2π

× 0

0

∞

0

∞

! da dc −1/3 −2/3 2 dθ r−θ (k), ac ω)| . |ψ(ac a c

Perform now the following change of variables: k = ac−1/3r−θ (k),

−2/3 ω = ac ω,

(10.32)

which is equivalent to −1/3 k1 ac k1 k2 0 cos θ k2 = −k2 k1 0 ac−1/3 sin θ .

−2/3 0 0 ω ac ω Thus the measure dθ

(10.33)

d k dω da dc becomes and we have a c |k |2 |ω |

k , ω )|2 |ψ( k, ω)|2 d 2 k dω |ψ(

2

2 2 | k | |ω | R ×R R ×R 2 k, | ψ( ω)| = (2π )3 d k dω ψ2 . 2 2 | k| |ω| R ×R

I = (2π )3

d k dω

It is clear that the integral I converges if and only if ψ satisfies the admissibility condition (10.29).

352

Spatio-temporal wavelets and motion estimation

10.4

The spatio-temporal wavelet transform

10.4.1 Spatio-temporal wavelets: definition and examples By definition, a spatio-temporal wavelet [158,Duv91] is a complex-valued function ψ ∈ L 2 (R2 × R, d 2 x dt) satisfying the condition: d 2 k dω 3 (10.34) |ψ(k, ω)|2 < ∞ cψ = (2π) 2 |ω| R2 ×R |k| is the Fourier transform of ψ. where ψ If ψ is regular enough, the admissibility condition (10.34) simply means that the wavelet must be of zero mean with respect to space and time independently: ψ(0, ω) = 0 ⇐⇒ d 2 x ψ( x , t) = 0. (10.35) R2

and k, 0) = 0 ⇐⇒ ψ(

R

dt ψ( x , t) = 0.

(10.36)

Clearly the unitary operators T(b,τ ) , Rθ , Da , #c preserve the admissibility condition, τ, θ; a, c). Hence any function ψb,τ,θ and so does therefore U (b, ;a,c = U (b, τ, θ ; a, c)ψ obtained from a wavelet ψ by translation, dilation, rotation, or speed tuning is again a wavelet. Thus the given wavelet ψ generates the whole family {ψb,τ,θ ;a,c }, indexed by the 2 elements (b, τ ) ∈ R × R, θ ∈ [0, 2π ), a > 0, c > 0, that is, by the elements of G mv . We will consider only one example of spatio-temporal wavelet, the spatiotemporal Morlet wavelet, characterized by the wavenumber–frequency (k0 , ω0 ) and the anisotropy parameter $ and defined by −1 1 1 1 2 −1 2 −1 2 ψ$ ( x , t) = ei k0 ·A x e− 2 |A x| − e− 2 |A x| e− 2 |k0 | 1 2 1 2 1 2 (10.37) × eiω0 t e− 2 t − e− 2 t e− 2 ω0 in the spatio-temporal domain and by 1 1 1 2 2 2 k0 |2 2 2 $ (k, ω) = e− 12 |Ak− − e− 2 (|Ak| +|k0 | ) e− 2 (ω−ω0 ) − e− 2 (ω +ω0 ) ψ

(10.38)

in the wavenumber–frequency domain, where A = diag[$ −1/2 , 1], $ 1, is the usual anisotropy matrix. As in the pure spatial case, for large |k0 | and |ω0 |, typically |k0 | 6, |ω0 | 6, the counterterms (the second term in each factor of (10.37) and (10.38)) are small enough to be neglected. This wavelet is a good candidate for motion estimation applications. Its region of support can be appropriately located around a particular plane, vo , typically vo = (1, 0), for example in the case where k0 = (k0 , 0), k0 = ω0 . Figure 10.1 shows half-energy equisurfaces of the squared amplitude of Morlet wavelets for different values of the parameters, θ, a and c. This illustrates how the

353

10.4 The spatio-temporal wavelet transform

ω

ω

kx

kx ky

ky

(a)

(b)

ω

kx ky

(c)

(d)

Fig. 10.1. Wavenumber–frequency domain coverage of the spatio-temporal Morlet wavelet for different parameters of the CWT. (a) Rotation θ; (b) scale a; (c) speed tuning c; (d) rotation, scale, and speed tuning θ, a, c.

wavelet parameters distribute the energy of the resulting filter on the wavenumber– frequency domain. The energy is distributed around a circle by the rotation parameter, along a conic volume by the scale parameter, and along a hyperbolic-like path by the speed tuning parameter. In order to control the variance of the wavelet with respect to the reference velocity plane, an anisotropy parameter $ applied to the spatial variables is used. This indeed controls its temporal support. As it can be seen from Figure 10.1, the region of support of the filters is an ellipsoidal cone concentrated around a particular velocity plane. Thus, velocity detection and filtering are possible.

10.4.2 The spatio-temporal wavelet transform Let now s ∈ L 2 (R2 × R, d 2 xdt) be an image sequence. Its (2+1)-D continuous wavelet transform (with respect to the fixed wavelet ψ), S ≡ Wψ s is the scalar product of s τ, θ; a, c)ψ, considered as a function with the transformed wavelet ψb,τ,a,c,θ ≡ U (b, of (b, τ, a, c, θ ):

354

Spatio-temporal wavelets and motion estimation

τ, θ; a, c) = √1 < ψb,τ,θ;a,c S(b, |s > (10.39) cψ 1 a −1 c2/3 (t − τ )) s( =√ x − b), x , t) d 2 x dt a −3/2 ψ(a −1 c−1/3 rθ ( cψ (10.40) 1 1/3 rθ (k)ac −2/3 ω) ω). d 2 k dω a 3/2 e−i(k·b+ωτ ) ψ(ac =√ s(k, (10.41) cψ The main properties of the (2+1)-D CWT Wψ : s → S may be summarized as follows [Com89,Mey91]: (i) As the purely spatial version, Wψ is linear in the signal s; (ii) Wψ is covariant under all the operations considered, namely [Com89,Mur90,13]. Proposition 10.4.1 The map Wψ is covariant under translations, rotations and τ, θ; a, c) dilations, which means that the correspondence Wψ : s( x , t) → S(b, implies the following ones: s( x − bo , t) → S(b − bo , τ, θ; a, c) τ − to θ; a, c) s( x , t − to ) → S(b,

(10.42)

τ, θ − θo ; a, c). s(r−θo ( x ), t) → S(r−θo (b), −1 −1 −1 ao−1 τ, θ ; ao−1 a, c) ao s(ao x, ao t) → S(ao−1 b,

(10.44)

co−2/3 τ, , θ ; a, co−1 c) S(co1/3 b,

(10.46)

s(co−1/3 x, co2/3 )

→

(10.43) (10.45)

It is worth noting that, conversely, the wavelet transform is uniquely determined by the three conditions of linearity, covariance and energy conservation, plus some continuity [Mur90]. (iii) Energy conservation: 1 2 2 τ, θ; a, c)|U (b, τ, θ; a, c)ψ | s|2 . d x dt |s( x , t)| = dµL (b, cψ G mv R 2 ×R (10.47) Thus, Wψ is an isometry from the space of signals into the space of transforms. (iv) As a consequence, Wψ is invertible on its range and the inverse transformation is simply the adjoint of Wψ . Thus one has an exact reconstruction formula: 1 τ, θ; a, c)ψb,τ,θ τ, θ; a, c). dµL (b, x , t) S(b, (10.48) s( x , t) = ;a,c ( cψ G mv In other words, the (2+1)-D wavelet transform, like its 1-D and 2-D counterpart, provides a decomposition of the signal in terms of the analyzing wavelets ψb,τ,θ ;a,c , τ, θ; a, c). with coefficients S(b, (v) Redundancy: Exactly as in the 2-D case discussed in Chapter 2, one has: Proposition 10.4.2 The projection from L 2 (G mv , dµL ) onto the range Hψ of

355

10.4 The spatio-temporal wavelet transform

Wψ , the space of wavelet transforms, is an integral operator whose kernel τ, θ; a, c) is the autocorrelation function of ψ, also called K (b , τ , θ ; a , c | b, reproducing kernel: τ, θ; a, c) = cψ−1 ψb ,τ ,θ ;a ,c | ψb,τ,θ K (b , τ , θ ; a , c | b, ;a,c .

(10.49)

Therefore, a function f ∈ L 2 (G, dg) is the wavelet transform of a certain signal iff it satisfies the reproduction property: f (b , τ , θ ; a , c ) τ, θ; a, c)K (b , τ , θ ; a , c | b, τ, θ; a, c) f (b, τ, θ; a, c). dµL (b, = G mv

(10.50)

10.4.3 An alternative: relativistic wavelets The spatio-temporal wavelets just described, which could be called kinematical, may not always be sufficient, depending on the type of signal to be analyzed. One may wish to consider a specific form of movement, i.e., choose a particular relativity group. Three examples may be of interest (we begin again with one space dimension). (i) Galilean wavelets Here we add to the transformations discussed above the Galilei boosts, thus getting (x, t) → (a1 x + vt + b1 , a0 t + b0 ). The resulting group G aff 1 , called the affine Galilei group, is quite complicated. It has a natural unitary representation in the space of finite energy signals, which splits into the direct sum of four irreducible ones, and each of these is square integrable, so that wavelets may be constructed in the usual way [25]. In addition, more restricted wavelets may be constructed by taking as parameter space various quotient spaces G aff 1 /H , where H is not the stability subgroup of the basic wavelet. (ii) Schr¨odinger wavelets One obtains an interesting subclass of the previous one by imposing the relation a0 = a12 , so that the transformations leave invariant the Schr¨odinger (or the heat) equation. Then the unitary irreducible representation UG of G aff 1 splits into the direct sum of two square integrable ones of the (Schr¨odinger) subgroup. Thus again a CWT is at hand, which may prove useful for describing, for instance, the motion of quantum particles on the line. (iii) Poincar´e wavelets In order to get a CWT in the relativistic regime, it suffices to replace Galilei transformations by Poincar´e ones, while of course imposing the relation a0 = a1 to space and time dilations. The result is the affine Poincar´e group, that we have discussed at length in Section 7.4. The Poincar´e wavelets might be useful, for instance, in the presence of electromagnetic fields.

356

Spatio-temporal wavelets and motion estimation

Of course, this analysis extends in a straightforward way to higher dimensions, just by adding rotations. Details may be found in [Ali00; Section 15.3]. These three types of relativistic wavelets offer additional examples of the general group-based wavelet formalism. They have a definite mathematical interest, but they have not been tested on practical situations, indeed no motion estimation (ME) algorithm based on them has been designed. On the contrary, the kinematical spatio-temporal CWT does lead to an efficient ME algorithm, that we now describe in detail.

10.5

A motion estimation (ME) algorithm We shall now exploit the general CWT formalism developed in the previous section and describe an algorithm for motion estimation. A complete discussion can be found in [Muj99,281,282]. Velocity filtering approaches for motion estimation allow temporal information to be incorporated in the estimation process. These techniques have performance advantages over two-frame-at-a-time based approaches, like block matching and optical flow [Tek95,353] in nonideal environments. The spatio-temporal CWT facilitates adaptive velocity filtering and offers an elegant framework for motion analysis and estimation. It performs a mapping from the Hilbert space H to a parameter space meaningful for motion estimation purposes. It can also be seen as a tool for motionbased filtering, where the filter characteristics are appropriately determined by a set of parameters directly associated with motion features. The wavelet basis matches the motion characteristics of the object of interest rather than its spatial features. General spatial selectivity is taken into account through the scale parameter a. It is assumed that starting conditions, i.e., position and velocity, of the object of interest are known initially. Our ME algorithm deals with the problem of following timevarying motion parameters on a frame-by-frame basis, which allows us to determine object coordinates at any time [Muj99]. In this sense, our ME algorithm can be viewed as an object tracking algorithm after the initial detection has been performed. This section is organized in two parts. First, we formulate the rˆole of the spatio-temporal CWT as a motion parameter estimator and we describe the three partial energy densities used for this purpose. In the second part, we describe how the interaction of these energy densities can be used to track a particular object from incoming video. Our CWT-based tracking algorithm is then introduced.

10.5.1 Partial energy densities The multidimensional nature of the spatio-temporal CWT allows for the definition of a multitude of energy densities either by fixing a subset of the parameter space or, better, by partial integration of the CWT energy,

357

10.5 A motion estimation (ME) algorithm

+ +2 + +2 E(g) ≡ E[s](g) = +ψg |s+ = + Sψ (g)+ ,

τ, θ; a, c}, g = {b,

(10.51)

on subsets of the parameter space. As discussed at length in the spatial case, in Section 2.3.4, this approach has the nice property that it can result in invariant representations with respect to some parameters. Thus, partial integration on subsets of the parameter space results in different energy representations that can be used to extract relevant features. Three energy densities particularly interesting for motion estimation purposes are studied here. (1) Speed-orientation energy density: Here integration is performed over spatial translation, b = (bx , b y ), on a region B : (bxmin < bx < bxmax ) ∩ (b ymin < b y < b ymax ),

(10.52)

while the scale a and the temporal variable τ are fixed. As a result, we obtain the first energy density, +9 :+2 + EaI o ,τo (c, θ) = d 2 b + ψb,τ (10.53) o ,θ;ao ,c | s , b∈B

which can be interpreted as an estimator of local velocity. The boundaries of the spatial region B are updated on a frame-by-frame basis to reflect changes in object location. This allows the algorithm to focus on the object of interest, and reduce interference with other nearby objects. (2) Spatial energy density: In this representation, the speed tuning parameter c, the orientation θ, the scale a, and the temporal translation τ , are fixed, while the spatial translation, b is the variable of interest. The resulting energy density is given by, = EIIτo ,θo ;ao ,co (b)

:+2 1 ++9 + ψb,τ o ,θo ;ao ,co | s . 4 ao

(10.54)

We can think of this energy density as the output energy of a velocity selective filter, where the location of objects moving at a pre-specified velocity, v = co eiθo , can be easily determined. In addition, size selectivity (invariability) can be obtained by appropriately choosing (integrating over) the scale parameter a. (3) Scale energy density: The scale parameter is associated with spatial size. Consequently, there exists an optimum scale aopt that best matches the size of the object of interest. A measure of “scale optimality” can be defined by integrating the global energy density, E, over while fixing the temporal translation τ , and the velocity the spatial variables, b, parameters, c and θ, leading to +9 :+2 1 III + Eco ,θo ,τo (a) = 4 d 2 b + ψb,τ (10.55) o ,θo ;a,co | s . a b∈B

358

Spatio-temporal wavelets and motion estimation

The spatial integration is constrained to the region B to avoid interference with nearby objects. These energy densities can be used to derive local estimates of the motion parameters. It is noteworthy that these energy densities are different from the global energy densities where additional integration is performed over the temporal parameter τ [279]. The global approach can only handle linear motion, while the local approach can deal with accelerated motion and time varying signatures as well. The energy densities presented here are the computational core of the CWT-based tracking algorithm, which is presented in the next section.

10.5.2 Description of the algorithm Computer implementation of the CWT requires discretization of the spatial and temporal variables ( x , t). The energy densities EI , EII , and EIII of (10.53), (10.54), and (10.55) must then be discretized by replacing integrations with appropriate summations. It is assumed the input data is presented as an incoming video signal with one or more objects moving with a constant or accelerated speed. For instance, the l-th object in a given image sequence is denoted sl ( xn,m , ti ), where n and m are spatial indices (horizontal and vertical respectively), and i is the temporal index. The resulting model for the input image sequence is then s( xn,m , ti ) = sl ( xn,m , ti ) + w( xn,m , ti ), (10.56) l

where w is assumed to be zero-mean white noise. The sum in (10.56) must be interpreted carefully when occlusions occurs. In this case, if the point ( xn,m , ti ) is subject to occlusion from two or more objects, the correct expression is s( xn,m , ti ) = s L ( xn,m , ti ) + w( xn,m , ti ),

(10.57)

where L is the index denoting the object closest to the sensor. The CWT-based ME algorithm relies on the three energy densities defined above and consists of a frame-by-frame optimization of the motion parameters associated with a given object in the image sequence. These motion parameters are gathered in a state vector L(ti ) defined as T (10.58) L(ti ) = vt i , xt i , ati , for time t = ti . Note the convention we use for the position variable; xn,m represents the spatial location indexed by the pair (n, m), while xti represents the spatial state at time t = ti . We assume that the starting position, x0 , and velocity, v0 , for the object of interest are known initially. The CWT energy densities are used here as optimality criteria or cost functions for updating the state vector L. This update can be done either by searching for the local

359

10.5 A motion estimation (ME) algorithm

maximum or by a gradient-based approach like the LMS algorithm, on each of the energy densities. In either case, this process is denoted symbolically by EI EII EIII

L(ti ) −→ L(ti+1 ) .

(10.59)

The optimization of equation (10.59) is performed sequentially. That is, when optimizing one component of the state vector (i.e., v, x, or a), the others are kept constant. Indeed, we are searching for an optimal set of motion parameters in the 6-D CWT space. The sequential optimization of the parameters considerably reduces the search space and consequently the computational requirements (with respect to a simultaneous 6-D search). This is possible due to the implicit redundancy of the CWT representation and its ability to isolate motion features. Thus, each of the energy densities corresponds to a 2-D (or 1-D for the scale energy density) “slice” of the CWT parameter space, where one parameter is optimized independently of the others. The CWT can be seen as a spatio-temporal filtering operation where the filter characteristics are controlled by a set of parameters associated with motion features (i.e., velocity and size). More explicitly, manipulating the inner product of equation (10.40) the CWT can be defined as a convolution sum, i.e., t − τ) Sψ (g) = d 2 x dt s( x , t) ψθ,a,c ( x − b, R 2 ×R # d 2 x dt s( x , t) ψθ,a,c (b − x, τ − t) = R 2 ×R

τ ), = s ⊗ ψ θ,a,c (b, where the symbol ⊗ represents the (2+1)-D convolution operator, t − τ ) = ψb,τ,θ;a,c ψθ,a,c ( x − b, ( x , t), and # ( x , t) = ψθ,a,c (− x , −t). ψθ,a,c

The indices of the filtered signal correspond to the spatio-temporal translation parameters of the CWT (i.e., b and τ ). This allows us to take advantage of the Fast Fourier Transform (FFT) to compute the CWT (and its associated energy densities) efficiently in the wavenumber–frequency domain. The separability of the wavelets can be exploited to reduce the required computations to construct the motion and scale selective filters. A block diagram of the state updating process performed by the CWT-based ME algorithm is depicted in Figure 10.2. The N × M × K block of data represents a portion of the incoming video signal with K frames of N columns by M rows. This image sequence is first transformed to the Fourier domain by means of a 3-D DFT. As suggested by Figure 10.2 and equation (10.59) the parameter update is done in a specified order: velocity first, then position, and finally scale. Two arguments support this implementation choice. First, velocity is a motion parameter of higher order than position. For highly maneuvering objects, the relative change in velocity from frame to

360

Spatio-temporal wavelets and motion estimation

K N

3-D FFT

...

M

Spatio-Temporal CWT

Input image sequence

L(n)

Speed-Orientation Energy Density

L(n)’

Space-Time Energy Density

L(n)’’

Scale (size) Energy Density

L(n+1)

Fig. 10.2. CWT-based tracking algorithm.

frame is smaller than the relative change in position. Second, some of the calculations necessary for determining the position energy density are already done at the end of the speed-orientation energy density stage. The three update stages embodied in the tracking algorithm are now described in detail. It is important to point out that the diagram shown in Figure 10.2 represents the operations performed in order to update the state vector L(ti ) for each frame of a given data block.

10.5.2.1

Velocity update stage In this stage the speed-orientation energy density EI is used as the optimality criterion to update the velocity state at time ti+1 . The position and scale states are fixed to their corresponding values at time ti , which are, xti and ati respectively. The discrete version of the speed-orientation energy density EI is

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close