Davelets and Signal essa ors IEE « Processing OLIVIER RIOUL and MARTIN VETTERLI Tove theory procs WV isc fr 8 number af teehngues whe a boon decoy le endenty for variowe signal Processing apcatione, or ex Mite multiescution sigh processing. used in compet Trotny subband coding developed ‘Erapech andinage compress nd woveet snes expansions evel apie materate Have been rent recopzed ms fire views ota singe they. trac wavelet theory covers guia arg arate oth he continuous and inederee ne cases proves ery ner techqes that eae apple to many take froin processing, ‘and thetere has muerous potential sppesone that, the Wvse Teasfonm (Wot inter: cetforth nnalin tno stationary gan bee frowdee an aerate to ihe casa Shoetine Fours Transion (SF) or Gabor temo oADO AULT? POWSD, The tase ference nae flows Cokaat tothe STFT which ance sng analysis ‘iow tne WT uses short windows at igh equenes Along window at ow quencies Tene spit of sovealled “constant or constant sels Stott regucrcy ana. The WT also reat {etme tequeny analyte bast on the Won ie Sisteti FA. FLAQO, FOB Tor some apleon ts dentable tose the WT as a sgl decomposition onset a at anions inet basa funenseate nclteatways ute Covet analy, They are obained trom sige Brettye walt by atone and contains (al essa ca{ngs} as well as shill The prototype wavelet can be thought of as a bandpass fer. and the constant-Q property ofthe other bandpass fters (wavelets) follows because they are sealed versions ofthe prototype ‘Therefore, in a WT, the notion of scale Is introduced fs an allemative to frequency. leading to a so-called time-scale representation. This means that 8 signal ‘mapped into a time-scale plane (the equivalent of the {ime frequency plane used in the STFT) “There are several types of wavelet transforms. and, depending on the applieaion, one may be preferred to the others. For continuous input sighal the time and scale parameters ean be continuous [GROS®), leading tothe Continuons Wavelet Transform (CWT). They may a wellbe discrete [DAUSS, MALS3b, MEYS9, DAUSOal, leading to a Wavelet Series expansion, Finally. the wavelet transform ean be defined for discrete-time sig nals [DAUSE, RIO9OD, VETOOD), leading to a Diserete Wavelet Transform (DWN. In the latter case iC uses ‘multirate signal processing techniques [CRO83] and 1s. ‘elated fo subband coding sehemes used in speech and image compression. Notice the analogy with the (Con tunuous) Fourier Transform, Fourer Series, and the Diserete Fourier Transform ‘Wavelet theory has been developed as a unifying framework only recently. although similar ideas and constructions took place a carly as the beginning of the century [HAALO, FRA2S, L1T37, CALB4). The idea of looking at a signal at various seales and analyzing it with various resolutions has in fact emerged inde pendently in many different fields of mathematics, Physics and engineering. In the mid-eighties. re Searchers ofthe “French school.” lea by a geophysieist. f theoretical physicist and a mathematician (namely. Moriet, Grossmann, and Meyer) bullt song mathe ‘matical foundations around the subject and named thetr work “Ondelettes" (Wavelets). They also interacted ‘considerably with ather feds, ‘The attention of the signal processing community ‘was soon caught when Daubechies and Mallat. in ad ‘ition to their contribution to the theory of wavelets, tstablished connections to discrete signal processing results [DAUS®], [MALS] Since then, a number of ‘heoretial, as well as practical contributions have been made on various aspects of WTS, and the subject Is rowing rapiely (WAV), {1102} The present paper fs meant both as a review and as «a tutorial, It covers the main definitions and properties ‘of wavelet transforms, shows connections among the ‘arious fells where results have been developed, and focuses on signal processing applications. Its purpose fs to presenta simple, synthete view of wavelet theory twit an easy-to-read, non-rigorous favor. An extensive bibllograpny is provided for the reader who wants to go Into more detail on a particular subject. coer NON-STATIONARY SIGNAL ANALYSIS, ‘The aim of signal analysis is to extract relevant Information trom a sjgnal by transforming it. Some methods make « prior assumptions on the signal to be Analyzed: this ma yield sharp results if these assump: tions are valid, but Is obviously not of general ap plicabilly. In this paper we focus on methods that are Applicable {0 any general signal. In addition, we con Sider invertible translormations. The analysis (hus wn ambiguously represents the signal. and more involved ‘operations such as parameter estimation. coding and pattern recognition can be performed on the "transform Ede.” where relevant properties ma be more evident ‘Such transforms have been applied to stationary signals, that i, signals whose properties do not evolve {i tume (the notion of stationarity is formalized precisely fn the statistical signal processing iterature). For such signals x, the natural “stationary transform” is the \well known Pourter transform [FOUSS} xe (Cay ea a ‘The analysis coefcients XU define the notion of bal frequency fin a signal. As shown to (1), they are ‘computed as inner products ofthe signal with sinewave bass functions of infinite duration. Asa result, Fourier analysis works well f (is composed ofa few stationary components (eg. sinewaves). However, any abrupt ‘change in time in a non-stationary signal a is spread ‘ut over the whole Requeney axis in XY}. Therefore. an ‘analysis adapted to nanstadonary signals requires more than the Fourer Transform “The ustal approach is 0 introduce time dependency Jn the Fourier analysts while preserving linearity. The {dea ts to introduce a Toca fequeney” parameter Gocal {in time) so that the “local” Fourier Transform looks at the signal through a window over which the signal is approximately stationary. Another, equivalent way ts 0 ‘modify the sinewave basis functions used inthe Fourier ‘Transform to basis functions which are more con- contrated in time (hut less concentrated in frequency) ‘SCALE VERSUS FREQUENCY ‘The Short-Time Fourier Transform: Analysis with Fixed Resolution. ‘The “instantaneous frequency” [FLASS] has often been considered as 1 way to Inigoduce frequency de cuca sa1 Tefen pane corp he er Tine Fourer ranajerms ofa uandowed segment othe signal er {eet snes) ras mode analyte ier bank rion spe pendence on time. If the signal is not narrow-band however, the instantaneous Trequency averages dif {erent spectral components in time. To become accurate in time, we therefore need a two-dimensional time-fe ‘quency representation Sts) of the signal (composed ‘of spectral characteristics depending on time. the local frequency being defined through an appropriate definition of Sif. Such a representation ts similar to the notation used is musieal score, which also shows frequencies” played in time. ‘The Fourier Transform) was first adapted by Gabor [GAB46] to define Si, f1as follows. Consider a signal 0, ‘and assume itis stationary when seen through a win {dow gif of limited extent, centered at time location & ‘The Fourier Transform (1) of the windowed signals ty git yield the Shr Tne Fourier Tranaorm Sh ste fis) ao dt-net ar @ Which maps the signa into a two-dimensional function fa time-frequency plane (tJ). Gabor originally only efined a synthesis formula, but the analysis given i (2) follows easly. ‘The parameter fin (2) 1s simllar to the Fourier frequency and many properties ofthe Fourier transform any over to the SIFT. However, the analysts here {epenisertically on the choice of the window Figure I shows vertical stripes inthe time-frequency plane, illustrating this “windowing af the signal” view of the SIFT, Given a version of the signal windowed ‘round time {one computes all “irequencies” of the Stvr ‘An alternative view Is based on a filter bank inter pretation of the same process. Ata given frequency J. (2) amounts to filtering the signal “at all umes” with a Dbandpass filter having as impulse response the window function modulated to that frequency. This is shoven as the horizontal stripes in Fig. 1. Thus, the STET may be seen as a modulated Ster bank (ALLT7) [POR8O]. From this dual interpretation, a possible drawback related to the time and frequency resolution can be Shown, Consider the ability of the STF to diseriminate Detween ro pure sinusoids, Given a window function al and its Fourier «ransform Gif. define the vandvwith” af ofthe Blter as a ruey ° » FR. Basis farctons and ta Func rekon othe Short Tine Farr Tanajorn (STF and the Wane Transform {The tes represent the essential concentration ne ne equency plane of a given bass function. (a) Coverage af the tne foqueney plane or the STET. hy Jr Use WT. fe) Corresponding bats futons or the STF. for the WT Puatelets 6 esr crn