P. 1
HPL 92 124 Wavelet Good

HPL 92 124 Wavelet Good

|Views: 29|Likes:
Publicado porthaobinh20

More info:

Published by: thaobinh20 on Oct 07, 2010
Direitos Autorais:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

10/07/2010

pdf

text

original

1rJ3 HEWLETT

a : ~ PACKARD
An Introduction to Wavelets
Lee A. Barford, R. Shane Fazzio, David R. Smith
Instruments and Photonics Laboratory
HPL-92-124
September, 1992
wavelets, wavelet
transform, multi-
resol ution analysis,
nonstationary signal
analysis
The past ten years have seen an explosion of re-
search in the theory of wavelets and their appli-
cations. Theoretical accomplishments include de-
velopment of new bases for many different func-
tion spaces and the characterization of orthonor-
mal wavelets with compact support. Applications
span the fields of signal processing, image proc-
essing and compression, data compression, and
quantum mechanics. At the present time
however, much of the literature remains highly
mathematical, and consequently, a large
investment of time is often necessary to develop a
general understanding of wavelets and their po-
tential uses. This paper thus seeks to provide an
overview of the wavelet transform from an intui-
tive standpoint. Throughout the paper a signal
processing frame of reference is adopted.
© Copyright Hewlett-Packard Company 1992
Internal Accession Date Only
Contents
1 Introduction 1
2 A Motivation for Wavelets 1
3 Wavelets and the Wavelet Transform 6
4 Comparision of the Fourier and Wavelet Transforms 11
5 Examples 15
6 Conclusion 22
ii
1 Introduction
Over the past ten years much has been accomplished in the development of the theory
of wavelets, and people are continuing to find new application domains. Theoretical
accomplishments include specification of new bases for many different function spaces
and characterization of orthogonal wavelets with compact support. Application areas
so far discovered include signal processing, especially for nonstationary signals, image
processing and compression, data compression, and quantum mechanics.
However, at the present time most of the literature remains highly mathematical and
requires a large investment of time to develop an understanding of wavelets and their
potential uses. The purpose of this paper is to provide an overview of wavelet theory by
developing, from an intuitive standpoint, the idea of the wavelet transform. Since a com-
plete study of wavelets would encompass both a lengthy mathematical development and
consideration of many application domains, we adopt a particular viewpoint that lends
itself readily to signal processing applications. Our discussion starts with a comparison
of the wavelet and Fourier transforms of an impulse function. This motivates a discus-
sion of the multiresolution decomposition of a function with finite energy. We then give
the definition of a wavelet and the wavelet transform. Following is a comparison of the
similarities and differences between the wavelet and Fourier transforms. \Ve conclude
with some examples of wavelet transforms of "popular" signals. Other introductions to
wavelets and their applications may be found in [1]' [2], [5], [8], and [10].
2 A Motivation for Wavelets
The short-time Fourier transform is frequently utilized for nonstationary signal analy-
sis. Although a powerful tool, it has some limitations in analyzing time-localized events.
The wavelet transform has similarities with the short-time Fourier transform, but it also
possesses a time-localization property that generally renders it superior for analyzing
nonstationary phenomena. ~ 7 e now review the Fourier and short-time Fourier trans-
forms, discuss some often desirable properties that the short-time Fourier transform
does not possess, and introduce the wavelet transform.
The Fourier and Short-Time Fourier Transforms
For any function f with finite energy, the Fourier transform of f is defined to be the
integral
j(w) =i: f(t)e-iwtdt, (1)
w being the angular rate, equal to 27[" times frequency. A Fourier transform is often
represented by its power spectrum-the square of the modulus of j(w) V5. w. For
example, the power spectrum of an impulse function has a constant value of unity and
is independent of the time at which the impulse occurs. Time of occurrence affects only
the phase of each frequency component.
1
The Fourier transform is best suited to analyze stationary periodic functions-those
that exactly repeat themselves once per period, without modification. It provides a
single spectrum for the whole signal. For nonstationary signals we are interested in the
frequencies that are dominant at any given time. For example, we perceive a musical
melody as a succession of notes, each with its own frequency spectrum, rather than as
one big signal with an overall spectrum. To analyze such signals, we may turn to the
short-time Fourier transform.
The short-time Fourier transform (or STFT) of a function at some time t is the
Fourier transform of that function as examined through some time-limited window cen-
tered on t. A different Fourier transform exists for each position t of the window. These
transforms, produced by sliding the examination window along in time, constitute the
STFT.
If the examination window simply omits the signal outside the window, two problems
are encountered. One is the sudden change in the power spectrum as a discontinuity
enters or leaves the window, compounded by a lack of sensitivity to the position of
the discontinuity within the window. The other problem is spectral leakage: if some
component of the signal has a cycle time which is not an integral divisor of the window
width, the transform exhibits spurious response at many frequencies. These problems
are ameliorated by attenuating samples away from the center of the window, by a
"windowing function," g. An example of a windowing function is the Gaussian, g(t) =
e-
at 2
, for some constant a.
I
Mathematically, the STFT at time T is given by
(2)
The response ofthe STFT, centered at time T =TO, to an impulse function 8(t - to)
occurring at time t =to is given by
jg(W, TO) = 1.: 8(t - to)g(t - To)e-iwtdt
g(to - To)e -iwto . (3)
The power spectrum of the STFT is jg(w, TO) = g2(to - TO). As shown in Figure 1, the
power spectrum is the same for all frequencies. The cross-section of the transform at
constant frequency produces a time-reversed copy of the windowing function. Thus, the
width (standard deviation) of the windowing function limits the accuracy with which
the impulse can be located in time.
Although the STFT windowing function's width is constant, its impact varies with
frequency. At high frequencies the number of waves in a window is high, producing good
accuracy in frequency measurement; yet the window width prevents good localization of
signal discontinuities, which the high frequencies otherwise could provide. Narrowing the
window width to accommodate more precise time-localization of discontinuities causes
other problems. A narrow window width is inappropriate at low frequencies, because
a narrow windowing function spans fewer cycles. It distorts the signal noticeably over
1The STFT using a Gaussian window is known as the Gabor transform.
2

f
'.
··.···· ••1::· .. ····.·.····· · ·· .. ·
..................::.. :::: .
Figure 1: Power spectrum for the short-time Fourier transform of an impulse function
using a Gaussian window centered at TO =0
one wavelength, degrading accuracy of frequency measurement. Indeed, wavelengths
longer than the window width cannot be measured. From these considerations it seems
advantageous to let the windowing function be broad for analyzing low frequencies and
narrow for high frequencies.
Since human perception of many types of signals has a logarithmic nature, the use
of constant-Q filters in signal processing is not uncommon. A number of studies of the
human senses find that the perceptual "distance" between two stimuli is dependent upon
their ratios (with respect to the appropriate units of measurement for the stimuli). For
example, a musical note one octave above another has twice the frequency. Loudness
is rated in decibels, a logarithmic measurement. Psychophysical experiments conclude
that the contribution of different frequencies to perceived quality of a visual image is
logarithmic in frequency. A filter bank used to process these signals naturally consists of
constant-Q filters-those whose bandwidths are proportional to their center frequencies.
However, interpreting the STFT as a filtering process results in a filter bank whose filters
have constant bandwidth.
Wavelet Transforms
Equation 2 shows that the STFT of a signal is the inner product of the signal with an
element of the set of basis functions g(t - T )e-
iwt
, which vary over frequency w and time
T. As shown by Figure 2a, all basis functions have the same time-amplitude envelope.
The STFT decomposes a signal into a set of frequency bands at any given time.
Wavelet transforms also decompose a signal into a set of "frequency bands" (referred
to as scales) by projecting the signal onto an element of a set of basis functions. Although
the scales do not live in the frequency domain, projection of the signal onto different
scales is equivalent to bandpass filtering with a bank of constant-Q filters. The basis
functions are called wavelets. Wavelets in a basis are all similar to each other, varying
only by dilation and translation, as illustrated in Figure 2b. Wavelet transforms thus
accommodate the two shortcomings of the STFT discussed above.
3
a. STFT b. Wavelet
Figure 2: Comparison of basis functions for the STFT and wavelet transforms. The
STFT places a varying number of waves under the same modulation envelope. The
wavelet basis functions are self-similar: scaled in time to maintain the same number of
oscillations and scaled in amplitude to maintain energy.
Figure 3 displays the square of the modulus of the continuous wavelet transform (in
analogy with the power spectrum) of an impulse function, using a Gaussian wavelet.
The parameter T represents time, and s represents scale. (The s axis in Figure 3 is
actually the base 2 logarithm of the scale.) Location of the impulse is clearly shown at
time T =O. Narrowness of the fine-scale basis functions trades off frequency resolution
in favor of time resolution. And yet, there is no set limit to the broadness of scale
at which analysis might be performed, aside from the length of the signal itself. In
contrast, the STFT is limited in both time and frequency resolution by the fixed width
of its window.
Multiresolution Analysis
In order to analyze a nonstationary signal, we need to determine its behavior at any
individual event. Multiresolution analysis provides one means to do this. A multireso-
lution analysis decomposes a signal into a smoothed version of the original signal and a
set of detail information at different scales. This type of decomposition is most easily
understood by thinking of a picture (which is a two dimensional signal). We remove
from the picture information that distinguishes the sharpest edges, leaving a new picture
that is slightly blurred. This blurred version of the original picture is a rendering at a
slightly coarser scale. We then recursively repeat the procedure. Each time we obtain
some detail information and a more and more blurred (or smoothed) version of the
original image. Removal of the detail information corresponds to a bandpass filtering,
and generation of the smoothed image corresponds to a lowpass filtering. Given the
decomposition, we can reconstruct the original image.
4
Figure 3: Square of the modulus of the continuous wavelet transform of an impulse
function at time T = 0, using a Gaussian wavelet. (The scale axis s is labeled in
exponents of two)
5
Once we have decomposed a signal this way, we may analyze the behavior of the de-
tail information across the different scales. We can extract the regularity of a singularity,
which characterizes the signal's behavior at that point. This provides an effective means
of edge detection. Furthermore, noise has a specific behavior across scales, and hence,
in many cases we can separate the signal from the noise. Reconstruction then yields
a relatively accurate noise free approximation of the original signal. This denoising
technique is developed in [7].
The wavelet transform specifies a multiresolution decomposition, with the wavelet
defining the bandpass filter that determines the detail information. Associated with
the wavelet is a smoothing function, which defines the complementary lowpass filter.
Conditions to be described later ensure that the set consisting of the detail information
at all scales and the smoothed version of the original signal contains no redundant
information.
In lieu of the wavelet transform's ability to localize in time and its ability to specify
a multiresolution analysis, many potential application areas have been identified. These
include edge characterization, noise reduction, data compression, and sub-band coding.
This list is by no means exhaustive-new applications are continually being discovered
both in signal processing and in other domains.
3 Wavelets and the Wavelet Transform
Mathematical Definitions
In this section we focus our attention on the mathematical definition of wavelets and
the wavelet transform. Table 1 provides a summary of the notation used in the rest of
this paper.
As discussed above, a multiresolution analysis of any function f with finite energy
decomposes f into a collection of details at different scales and a smoothed version of the
original function. If we let L2(R) denote the set of all functions with finite energy, then
the details of f at any scale m is simply a projection of f onto a subspace W
m
of L
2(R).
This projection may be formally represented by a projection operator Qm :L2(R)-+W
m
.
Furthermore, there exists another projection operator PM: L2(R) -- VM, VM C L2(R),
such that PMf is the smoothed version of f. Here m takes the values m = 1,2, ...,M,
that is, f is decomposed into the smoothed version PMf and M sets of details at different
scales. As M increases the resolution of the smoothed version of f becomes coarser, and
consequently, the finer detail information is contained in the scales corresponding to low
values of m. For any multiresolution analysis the W
m
are orthogonal both to each other
and to VM. In addition, assuming that Vo = L2(R), we have L2(R) = E f ! ~ = 1 W
m
Ef! VM,
and hence we may write
M
f =PMf + L Qm/.
m=l
(4)
The question now arises as to how to define the Qm. We desire an orthonormal basis
6
L (R)
f
m
n
s
T
TO
'ljJmn
ng
iscretization 0 t e continuous trans ation
s
Table 1: Summary of notation
7
of W
m,
call it ?/Jmn (m fixed), so that
00
Qmf = L (I, ?/Jmn)?/Jmn,
n=-oo
(5)
where (-,.) denotes inner product. As it turns out, wavelets provide the ?/Jmn. We
therefore turn our attention to the definition of wavelets and the wavelet transform and
then show how to define Qm in terms of them.
Before proceeding however a few remarks may provide some clarification. After re-
moving the first set of detail information from f we are left with a slightly smoothed
version of f. Iteratively removing detail information progressively generates more and
more smoothed versions of f. We stop the process once we have enough detail infor-
mation or a smooth enough version to do the analysis we desire. Thus the subspaces
Wand V are intimately related in the following way: given any m, a function f E V
m
is decomposed into details and a smoothed version at the m + 1 scale. That is, f is
decomposed by projecting it onto orthogonal complements in V
m
, one subspace being
Wm+l and the other V
m
+
1
. Formally, we have V
m
= W
m
+1 EB Vm+l'
Wavelets consist of the dilations and translations of a single real valued/ function
?/J E L2(R), called the analyzing wavelet (also known as the basic wavelet or mother
wavelet). By a dilation we mean a scaling of the argument, so that given any function
?/J(t) and a parameter s > 0, j;?/J( ~ ) is a dilation of?/J. Consequently, a dilation of
a function corresponds to a either a spreading out or contraction of the function. We
introduce the factor ~ with the foresight that it yields a normalization necessary to
have an orthonormal wavelet basis. Translation simply means a shift of the argument
along the real axis, that is, given T, the translation of ?/J(t) by Tis ?/J(t - T).
For any analyzing wavelet ?/J we thus define a family of functions ?/Js,r by the dilations
and translations of ?/J,
-1/2 t - T
?/JS,T(t) = s ?/J(-s-)' s, T E R, s > O.
Each ?/JS,T is called a wavelet.
3
We may represent any function f E L
2(R)
by
W f is called the continuous wavelet transform of f.
(6)
(7)
2In general, the analyzing wavelet may be complex valued. For this paper however we assume that
it is real valued.
3 Technically, t/J must also satisfy the admissibility condition
c, = 1
00
1¢(wW dw < 00,
o w
where ¢ denotes the Fourier transform of t/J. See [3, 4, 5].
8
Given So > 1 and TO 1: 0, we may restrict S and T, respectively, to the discrete lattices
S E {so, m «Z} and T E {nsOTO, m, n E Z}. Then 'l/JS,T becomes
"pmn(t) = - nTo), (8)
and Wf becomes
(WJ)(m,n) = (f,,,pmn) = i:f(t)"pmn(t)dt. (9)
(10)
Note that the translation parameter T depends upon the scaling parameter So. When
m is large and positive, "pmn is spread out, and the translation steps become large
accordingly. When m is large (in absolute value) and negative however the "pmn are
very concentrated, and the translation steps are small.
The choice of TO is arbitrary and by convention is taken to be 1. On the other hand,
the choice of So significantly affects the properties of the "pmn. For doing multiresolution
analysis we want the "pmn to be orthonormal. Taking So = 2 allows us to define "p such
that the "pmn are orthonormal [3, 6]. This is also the conventional choice for So. The
"pmn =2-
m/2"p(2- mt
- n) then form an orthonormal basis of L
2(R).
Equation 9, with
this choice of "pmn, is called the dyadic wavelet transform.
As an example, let the analyzing wavelet be defined by
{
I , t E
"p( t) = -1, tEa, 1)
0, otherwise
This wavelet, illustrated in Figure 4, is called the Haar wavelet, and the corresponding
"pmn form a basis of L2(R) called the Haar basis.
Now we return to Qm, the operator that projects a function onto its details at scale
m. We define Qm as in Equation 5 by
00
Qmf = L (j,,,pmn),,pmn =L(WJ)(m,n)"pmn. (11)
n=-oo n
The details of f at each scale m thus consist of the sum of the projections of f onto
the "pmn. Note that these projections are not onto all the "pmn, for at each scale m is
fixed and only n varies. To summarize, given any function f with finite energy and
an analyzing wavelet "p such that the "pmn are orthonormal, we may compute, using
Equations 4 and 11, the details of f (i.e., QmJ) at the scales m = 1,2, ... , M and the
corresponding smoothed version of f at scale M (i.e., PM f = f - QmJ)' thus
obtaining a multiresolution analysis.
Properties and Examples
Wavelets possess some interesting properties in addition to those already discussed. The
admissibility condition (see footnote, p. 8) implies =0, and hence
i:"p(t) dt = = 0,
9
(12)
-1
Figure 4: Haar wavelet
that is, wavelets have zero mean. Note that this indicates the equivalence of a wavelet
and a bandpass filter.
Many wavelets have rapid decay. Y. Meyer [3] constructed a Coo wavelet that decays
faster than any power. P. G. Lemarie and G. Battle [3] independently constructed a
collection of c- wavelets that decay exponentially."
Orthogonal wavelets may be classified as either having compact support or not
having compact support.f I. Daubechies [3] characterized all orthogonal wavelets with
compact support. She showed that the Haar wavelet is the only such wavelet that is
either symmetric or antisymmetric about any point. She also showed that compactly
supported wavelets may be chosen with arbitrary regularity; however, the support width
varies directly with the regularity. The compact support of Daubechies' wavelets and
the rapid decay of the wavelets described by Meyer and Lemarie and Battle help provide
for both the time-localization ability and efficient computation of the wavelet transform.
We finally consider some examples of wavelets. The first example is the Haar wavelet
discussed above. A second example is a wavelet constructed by Lemarie and Battle [6],
illustrated in Figure 5. As an analyzing wavelet it yields an orthonormal basis of L2(R)
but does not have compact support. An example, other than the Haar wavelet, of
a compactly supported wavelet that yields an orthornormal basis of L
2(R)
is shown
in Figure 6. This wavelet was first constructed by Daubechies [3]. Note the lack of
symmetry. Although continuous, this wavelet is non-differentiable at an infinite number
4c
oo
represents the space of all analytic functions, and C
k
represents the space of all k-times con-
tinuously differentiable functions.
!>A function f: R -+ R is said to have compact support if and only if it is zero everywhere except on
a closed, bounded subset of R.
10
of points. It is supported on the interval [0,3].6
4 Comparision of the Fourier and Wavelet Transforms
In this section we discuss some of the similarities and differences between the Fourier
and wavelet transforms. Table 2 gives a summary.
Both the Fourier transform and wavelet transform are given by integral equations in
the form of a correlation. In the Fourier transform the correlation is with dilations of the
function e-
it
. In the wavelet transform the correlation is with dilations and translations
of the analyzing wavelet '1/;, which can be any wavelet.
Both the Fourier transform and wavelet transform may take real or complex func-
tions as their input. The output of the Fourier transform is always complex. However,
there are both real- and complex-valued wavelets. If a complex-valued wavelet is used
as the analyzing wavelet, the wavelet transform is complex-valued. If a real-valued
analyzing wavelet is used, the wavelet transform may be real- or complex-valued (real-
valued if the input function is real-valued and complex-valued if the input function is
complex-valued ).
The Fourier transform maps time into frequency and phase; whereas, the wavelet
transform maps time into scale and time. For each frequency the Fourier transform
6The Battle-Lemarie wavelet is characterized by its Fourier transform [6]:
where
with
and
• W )4 ( W 2 W 4 2 . W )6
N2 =2(sm '2 cos '2) +70(cos '2) + 3(SlR'2 .
The Daubechies wavelet is characterized as follows [10): Given Co = t(1 + V3), Cl = t(3 + V3),
C2 = t(3 - V3), and C3 = t(1 - V3), we define
1
.,(t) =L (-1)kC1_klP(2t - k),
k=-2
where lP(t) is the limit as j -+ 00 of the recursion
3
tPJ(t) =I>ktPi-1(2t - k),
o
with
IP (t) ={I, t E[0,1) .
o 0, otherunse
11
Figure 5: Battle-Lemarie wavelet
Figure 6: Compactly supported Daubechies wavelet
12
Fourier Transform Wavelet Transform
"root" function e
iwt
s-1/2'l/Je-.
T)
Continuous transform j(w) = f(t)e-iwtdt Wcf(s,r) = f(t)s-1/21/Je-.
T)dt
Inverse transform f(t) = j(w)eiwtdJ,;; f(t) = J
o
oo
Wcf(s, r )d:t
T
(up to a proportion-
alityconstant)
Time transformed to amplitude and phase amplitude for each scale and time
for each frequency
Input domain Ror C Ror C
Output range C Ror C
Localization in Yes Yes
frequency
Localization in time No (limited with STFT) Yes
Time for fast O(nlog n) O(n)
discrete transform
Number of nonredundant n n
outputs of discrete
transform
Table 2: Comparison of Fourier and Wavelet Transforms
yields an amplitude and a phase. A signal may then be represented as the sum of sine
waves whose phase and amplitude are given by the Fourier transform. Similarly, the
wavelet transform yields an amplitude for each scale and time. A signal is represented
as the sum across scales of time-centered wavelets whose amplitudes are given by the
wavelet transform.
"Scale" is roughly "minus log frequency", in the following limited sense. A single
scale s contains information from a band of frequencies. The width and the center
frequency of the band are both proportional to -log s. Each scale's band has the same
ratio of bandwidth to center frequency, so scales correspond to a set of constant-Q filters.
Figure 7 shows the logarithmic progression of band widths for the filters corresponding
to different scales.
Both the Fourier transform and wavelet transform are said to "localize in frequency."
That is, both produce an output that is nonzero only in a band when given an input
that contains only frequencies from that band.
The wavelet transform also "localizes in time." A signal that is nonzero only during
a finite span of time has a wavelet transform whose nonzero elements are concentrated
around that time. On the other hand, the Fourier transform does not localize in time.
For example, the Fourier transform of an impulse function contains high amplitudes at
all frequencies. A wavelet transform of an impulse is either contained in a finite band of
time or decreases expontentially with distance from the time of the impulse. STFTs, as
discussed above, can be used to provide some localization in time, although the STFT
13
................................................ -- _ ••••••••••_ ' •••••• _ •• _ ~ . ~ . ~ . : . : .:. :':'-:.·I,·to ·-•••_- ••••• ---- - _ •• _ _.'-- - ----_ ' ••••- --_.. _ ••••
#'#'#' •
.'
\
,
,
"
'"
"
\
,
\
,
\
,
\
,
,
\
,
\
,
\
,
\
,
\
,
\
,
,
,
,
,
,.
0'
" ,\. / \ I
of \/
), ,'\
• I \ I \
I I '\,' \
1
I \ •
J ' , "
I 'Ii , ~
I 1 J' : \
t\ \
'1 I
I I \ \
o / , 0:' \
rI: \: \
'j "
,II: /\ \
rI 1 .: \ \
'1:' \ "
J
"
..
, I '\ \
H, : ,
~ ' . " , -,
f '" ''' ...
Figure 7: Logarithmic progression of frequency band width for filters corresponding to
different scales, using the cubic spline analyzing wavelet. (frequency vs. logarithm of
amplitude)
14
gains its localization in time by trading off both frequency resolution and bandwidth.
No such tradeoff is required when using the wavelet transform.
There are discrete versions of both the Fourier transform and the wavelet trans-
form. The discrete Fourier transform (DFT) is obtained from the Fourier transform
by replacing the integral with a finite sum. The discrete wavelet transform (DWT)
is obtained from the wavelet transform in the same way. Both the DFT and DWT
take D(n
2
) time, where n is the number of input values. A DFT is usually computed
using the Fast Fourier Transform (FFT) algorithm, which takes advantage of certain
symmetries inherent in convolving with e
iwt
when n is a power of two to reduce the
time to D( n log n). Similarly, there is a Fast Wavelet Transform (FWT) algorithm that
takes advantage of symmetries that are inherent in convolving with wavelets when n is
a power of two. A FWT requires D(n) time. The FFT takes n inputs and produces n
outputs ( ~ complex numbers, or ~ frequencies and ~ phases). The FWT also takes n
inputs and produces n outputs: 1 for 8 =log n, 2 for 8 =log n - 1, ... , ~ for 8 =1.
5 Examples
We now show wavelet transforms of some simple signals. All of the examples were
computed with the FWT using 80 =2 (d. Section 3).
Figure 8 shows the wavelet transform of an impulse, using the four coefficient
Daubechies wavelet W
4
as the analyzing wavelet. Note that the maximum amplitude
of the transform decreases at each scale after Scale 2. At first glance each scale of the
transform resembles a dilated W4' However, each scale has more oscillations than the
analyzing wavelet. Just like W
4
, each scale is finitely-supported and has an infinite
number of non-differentiable points. Since W4 is neither symmetric or antisymmetric,
neither is the transform.
Figure 9 shows the wavelet transform of an impulse, using the eight coefficient
Daubechies wavelet W
s
as the analyzing wavelet. With this analyzing wavelet the
maximum amplitude of the transform decreases at every scale. Each scale of the trans-
form has more oscillations than the corresponding scale in Figure 8, because W
s
has
twice as many oscillations as W
4.
Just as Ws is smoother than W4 (for example, Ws is
differentiable), this transform is smoother than the one in Figure 8. Since W
s
is neither
symmetric or antisymmetric, neither is the transform.
Figures 10 and 11 show the wavelet transforms of a step function using W
4
and
W
s
. Many of the properties seen in the previous two figures still hold. For example, the
transform under W
4
is non-differentiable, but the transform under W
s
is. Note, however,
that the maximum amplitudes of the transform at different scales are all roughly equal.
Figure 12 shows the wavelet transform of a triangle wave, using the cubic spline
wavelet. Note that the amplitude of the transform increases from scale to scale. Also,
observe that where the signal is a straight line, the wavelet transform is constant in
every scale.
Figure 13 shows the wavelet transform of two steps and an impulse. The analyzing
wavelet is the so-called "cublic spline wavelet". This wavelet is a twice continuously
15
'Ov
__1'-----_
------'t'------
- - ~ _ . -
Signal
:i
Scalel
:i
Scale 2
:i
ScaleS
:f
Scale 4
:i
Smoothed
: ~
Figure 8: Wavelet transform of an impulse using the four coefficient Daubechies wavelet,
W
4
16
ov
« A
__1_-
- - i ~ -
- - - - - - - - - " ' \ . ~ - - - - - -
Signal
:l
Scale!
:i
Scale 2
:I
Scale 3
:i
Scale 4
:I
Smoothed
:i
Figure 9: Wavelet transform of an impulse using the eight coefficient Daubechies wavelet,
W
s
17
Signal
~ ~
Scalel
~ ~
~
Scale 2
~ I
t
ScaleS
~ i
~ h
~ V
Scale 4
~ I
_ n
.
\j V
Smoothed
~ !
/:
7
Figure 10: Wavelet transform of a step using W
4
18
Signal
~ !
Scalel
~ !
~
Scale 2
~ !
1\ ..
YO
Scale 3
~ !
~ A 0
Scale 4
~ !
.-... -
<>
V
Smoothed
~ I
c:
]
Figure 11: Wavelet transform of a step using W
s
19
...

"'7
:1---
:1----




ScaleS
Scale 4
Smoothed
Scale 2
Scalel
Figure 12: Wavelet transform of a triangle wave using a cubic spline wavelet
20
:1
Signal
Scalel
:1
A
t
y
Scale 2
:1
A
,.
V v
Scale 3
:1
/\
.n
C>
V'
Scale 4
:1
L\..
ucc;;;:::;::>""
:1
Smoothed L ~
:7
Figure 13: Wavelet transform of the sum of an impulse and two steps using a cubic
spline wavelet
21
differentiable, piecewise cubic function. Unlike W
4
and W
s,
it is antisymmetric. This
anti symmetry causes the wavelet transform of symmetric signal features to be symmet-
ric. This is visible in Figure 13, where the transform is symmetric about the steps and
impulse.
The cubic spline wavelet does not have finite support. However, it does decay expo-
nentially. In Figure 13, each scale of the transform contains small amplitude oscillations
that extend infintely in both directions. These oscilations are invisible in the figure due
to the rapid decay of the wavelet
Figure 13 also demonstrates the linearity of wavelet transforms. The wavelet trans-
form of the two steps and impulse is the sum of the wavelet transforms of each step and
of the impulse. In the signal the upwards step goes from -0.5 to 0.5, and the downwards
step goes from 0.5 to O. In the wavelet transform, at each scale, the peak resulting from
the downwards step is opposite in sign and half the magnitude of the peak resulting
from the upwards step.
Again, note that the wavelet transform of the impulse decays from scale to scale,
but the wavelet transform of the steps stay roughly constant.
The signal in Figure 14 is the signal of Figure 13 plus noise uniformly distributed
on [-0.3,0.3]. Note that the transform peaks generated by the noise decay like the
wavelet transform of an impulse. By Scale 4 the peaks generated by the step edges are
pronounced, but the noise peaks are quite small. This observation is the basis of several
signal denoising algorithms in the literature. (See, for example, [7]).
6 Conclusion
Due to their capability to localize in time, wavelet transforms readily lend themselves to
nonstationary signal analysis. Detection of short duration events, on the other hand, are
limited in Fourier analysis by the width ofthe windowing function used in the short-time
Fourier transform. Wavelet transforms exist that project a finite energy function onto to
an orthonormal basis of L
2(R).
The corresponding multiresolution analysis decomposes
the function into a set of details at different resolutions and a smoothed version of the
original function. As with the Fourier transform, a "fast wavelet transform" exists.
However, the fast wavelet transform generates a multiresolution analysis in O( n) time;
whereas, a fast Fourier transform takes O( n log n) time.
Our intent in this paper was to present the basic concept of the wavelet transform
from a viewpoint that targets signal analysis applications. Much of the current literature
utilizes a high level of mathematical terminology. Our hope was to provide an brief
introduction to the primary underlying ideas in a relatively intuitive manner.
For those who are interested, we provide an annotated bibliography that includes
some of the key papers in the field. With each listing is a short description of the
contents. Most of the papers require an understanding of Fourier analysis and sometimes
an understanding of more general functional analysis principles. To help specify the
mathematical sophistication of a paper, we adopt a relative rating scale, based upon
our experience of reading the papers, 1 meaning little or no mathematical sophistication
22
Signal
Scalel
Scale 2
'I
.... --...
o
1\ . A. _" A.. u,+ •• . it A
V v \V. fV"" 4pJ eo 'VVv 9V\i
:c::::a "=""-
<> 0
V
......
A
o _
_ e
U
.
L
-----_/



Smoothed
Scale 3
Scale 4
Figure 14: Wavelet transform of the sum of an impulse, a step and uniform noise using
a cubic spline wavelet
23
required and 5 meaning a high level of mathematical sophistication required, relative to
the other papers.
24
References
[1] Cody, Mac A. "The Fast Wavelet Transform," Dr. Dobb's Journal, April 1992.
This paper provides a quick overview of wavelets and the fast wavelet transform.
It also contains and discusses some C code that implements a wavelet transform.
Mathematical level-L
[2] Wavelets: Time-Frequency Methods and Phase Space. Proceedings of the Interna-
tional Conference, Marseille, France, December 14-18, 1987. Combes, J. M., et al.
eds. Springer-Verlag.
This book presents some of the earlier papers in the development of wavelets. The
papers are organized into the five categories:
Part I: Introduction to Wavelet Transforms
Part II: Some Topics in Signal Analysis
Part III: Wavelets and Signal Processing
Part IV: Mathematics and Mathematical Physics
Part V: Implementations
Mathematical level-l through 5
[3] Daubechies, Ingrid. "Orthonormal Bases of Compactly Supported Wavelets," Com-
munications on Pure and Applied Mathematics, Vol. 41, No.7, 1988, pp. 909-996.
In this paper Daubechies reviews multiresolution analysis, the Laplacian pyramid
decomposition scheme, and the fast wavelet transform algorithm for doing a mul-
tiresolution analysis using a wavelet basis. She then demonstrates the classification
of all orthonormal bases of wavelets with compact support. Mathematicallevel-4
[4] Grossman, A. and Morlet, J. "Decomposition of Hardy Functions into Square In-
tegrable Wavelets of Constant Shape," SIAM Journal of Mathematical Analysis,
Vol. 15, No.4, July 1984, pp. 723-736.
This is the paper primarily responsible for the current interest in wavelet trans-
forms. They discuss the definition of the transform, its inverse, the admissibility
condition, and they prove that the wavelet transform is an isometry. Mathematical
level-5
[5] Mallat, S. "Multifrequency Channel Decompositions of Images and Wavelet Mod-
els," IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. 37, No.
12, December 1989, pp. 2091-2110.
We found this paper to be a good introductory paper to the wavelet transform.
It starts by discussing multifrequency channel decompositions in psychophysics,
reviews the short-time Fourier transform, and introduces the wavelet transform.
It then proceeds to discuss pyramidal multiresolution decompositions, two dimen-
sional wavelet transforms, and zero-crossings of multifrequency channels. Mathe-
matical level-3
25
[6] Mallat, S. "A Theory for Multiresolution Signal Decomposition: the Wavelet Rep-
resentation," IEEE Transactions on Pattern Analysis and Machine Intellegence,
Vol. 11, No.7, July 1989, pp. 674-693.
Mallat here presents his O( n) algorithm for computing the fast wavelet transform.
Furthermore, he discusses some of the technical conditions imposed upon the low-
pass and bandpass filters necessary for them to correspond to the projection opera-
tors that generate the multiresolution analysis. (Daubechies further discusses these
conditions in her paper cited above). He concludes by discussing image processing
applications of the wavelet transform and by giving an example of a multiresolution
approximation. Mathematicallevel-3
[7] Mallat, S. and Hwang, W. L. "Singularity Detection and Processing with Wavelets,"
IEEE Transactions on Injormation Theory, Vol. 38, No.2, March 1992, pp. 617-
643.
In this paper Mallat and Hwang discuss a means of using the wavelet transform
to characterize the Lipschitz regularity of a function at a point. They discuss the
detection and measurement of oscillating and nonoscillating singularities and illus-
trate a technique for denoising a signal based upon its wavelet transform modulus
local maxima. Mathematicallevel-4
[8] Rioul, O. and Vetterli, M. "Wavelets and Signal Processing," IEEE Signal Process-
ing Magazine, October 1991, pp. 14-38.
This is an introductory paper to wavelet transforms. It well illustrates the dif-
ference between the short-time Fourier transform and the wavelet transform. The
paper contains a relatively complete set of references that we found very useful.
Mathematicallevel-2
[9] Wavelet and Their Applications. Ruskai, M. B. et al. eds. Jones and Bartlett, 1992.
This book consists of a collection of papers organized into the categories
Signal Analysis
Numerical Analysis
Other Applications
Theoretical Developments
and contains many of the more recent papers. Mathematical level-3 through 5
[10] Strang, G. "Wavelets and Dilation Equations: A Brief Introduction," SIAM Re-
view, Vol. 31, No.4, December 1989, pp. 614-627.
Strang presents a brief mathematical introduction to multiresolution analysis and
the fast wavelet tranform algorithm. However, the paper is rather terse in its ex-
planations. Mathematical level-3
26
[11] IEEE Transactions on Information Theory, Vol. 38, No.2, March 1992, Part II.
Part II of this issue is dedicated to wavelets and their applications. It contains
many recent papers on the applications of wavelets in information theory related
areas. Mathematicallevel-3 through 4
27

Contents
1 Introduction 1 1
6

2 A Motivation for Wavelets
3
4

Wavelets and the Wavelet Transform Comparision of the Fourier and Wavelet Transforms

11

5 Examples
6

15

Conclusion

22

ii

w. The wavelet transform has similarities with the short-time Fourier transform. [5]. we adopt a particular viewpoint that lends itself readily to signal processing applications. We then give the definition of a wavelet and the wavelet transform. the Fourier transform of f is defined to be the integral j(w) w = i: f(t)e-iwtdt. The purpose of this paper is to provide an overview of wavelet theory by developing. the power spectrum of an impulse function has a constant value of unity and is independent of the time at which the impulse occurs. Theoretical accomplishments include specification of new bases for many different function spaces and characterization of orthogonal wavelets with compact support. Since a complete study of wavelets would encompass both a lengthy mathematical development and consideration of many application domains. [8]. The Fourier and Short-Time Fourier Transforms For any function f with finite energy. 1 . Other introductions to wavelets and their applications may be found in [1]' [2]. ~7e now review the Fourier and short-time Fourier transforms. especially for nonstationary signals. from an intuitive standpoint. Following is a comparison of the similarities and differences between the wavelet and Fourier transforms. A Fourier transform is often represented by its power spectrum-the square of the modulus of j(w) V5. it has some limitations in analyzing time-localized events. and people are continuing to find new application domains.1 Introduction Over the past ten years much has been accomplished in the development of the theory of wavelets. but it also possesses a time-localization property that generally renders it superior for analyzing nonstationary phenomena. and [10]. equal to 27[" times frequency. \Ve conclude with some examples of wavelet transforms of "popular" signals. 2 A Motivation for Wavelets The short-time Fourier transform is frequently utilized for nonstationary signal analysis. Our discussion starts with a comparison of the wavelet and Fourier transforms of an impulse function. and quantum mechanics. (1) being the angular rate. Time of occurrence affects only the phase of each frequency component. at the present time most of the literature remains highly mathematical and requires a large investment of time to develop an understanding of wavelets and their potential uses. image processing and compression. the idea of the wavelet transform. data compression. and introduce the wavelet transform. For example. However. discuss some often desirable properties that the short-time Fourier transform does not possess. Although a powerful tool. Application areas so far discovered include signal processing. This motivates a discussion of the multiresolution decomposition of a function with finite energy.

the width (standard deviation) of the windowing function limits the accuracy with which the impulse can be located in time. The short-time Fourier transform (or STFT) of a function at some time t is the Fourier transform of that function as examined through some time-limited window centered on t. to an impulse function 8(t .To)e -iwto . It distorts the signal noticeably over 1 The STFT using a Gaussian window is known as the Gabor transform.TO). The power spectrum of the STFT is jg(w. These problems are ameliorated by attenuating samples away from the center of the window.To)e-iwtdt (3) g(to . two problems are encountered. It provides a single spectrum for the whole signal. the transform exhibits spurious response at many frequencies. These transforms. rather than as one big signal with an overall spectrum.I Mathematically. An example of a windowing function is the Gaussian.to)g(t . we perceive a musical melody as a succession of notes. A narrow window width is inappropriate at low frequencies. Narrowing the window width to accommodate more precise time-localization of discontinuities causes other problems. As shown in Figure 1. centered at time T = TO. constitute the STFT. For nonstationary signals we are interested in the frequencies that are dominant at any given time. which the high frequencies otherwise could provide. we may turn to the short-time Fourier transform. producing good accuracy in frequency measurement. To analyze such signals. the STFT at time T is given by (2) The response ofthe STFT. each with its own frequency spectrum. without modification. yet the window width prevents good localization of signal discontinuities.at 2 . the power spectrum is the same for all frequencies. produced by sliding the examination window along in time. because a narrow windowing function spans fewer cycles.The Fourier transform is best suited to analyze stationary periodic functions-those that exactly repeat themselves once per period. If the examination window simply omits the signal outside the window. TO) = 1. One is the sudden change in the power spectrum as a discontinuity enters or leaves the window. The cross-section of the transform at constant frequency produces a time-reversed copy of the windowing function. by a "windowing function.to) occurring at time t = to is given by jg(W. TO) = g2(to . A different Fourier transform exists for each position t of the window. for some constant a. compounded by a lack of sensitivity to the position of the discontinuity within the window.: 8(t . its impact varies with frequency. Although the STFT windowing function's width is constant. g(t) = e." g. The other problem is spectral leakage: if some component of the signal has a cycle time which is not an integral divisor of the window width. 2 . For example. At high frequencies the number of waves in a window is high. Thus.

. interpreting the STFT as a filtering process results in a filter bank whose filters have constant bandwidth. wavelengths longer than the window width cannot be measured. a musical note one octave above another has twice the frequency.·. all basis functions have the same time-amplitude envelope.. Wavelet transforms also decompose a signal into a set of "frequency bands" (referred to as scales) by projecting the signal onto an element of a set of basis functions. ····.. a logarithmic measurement. For example. Wavelet Transforms Equation 2 shows that the STFT of a signal is the inner product of the signal with an element of the set of basis functions g(t .... Wavelet transforms thus accommodate the two shortcomings of the STFT discussed above. From these considerations it seems advantageous to let the windowing function be broad for analyzing low frequencies and narrow for high frequencies. Although the scales do not live in the frequency domain. As shown by Figure 2a.. However. Wavelets in a basis are all similar to each other. the use of constant-Q filters in signal processing is not uncommon. The STFT decomposes a signal into a set of frequency bands at any given time... varying only by dilation and translation. Psychophysical experiments conclude that the contribution of different frequencies to perceived quality of a visual image is logarithmic in frequency.. 3 ...····· '. The basis functions are called wavelets. Indeed. Loudness is rated in decibels.iwt . projection of the signal onto different scales is equivalent to bandpass filtering with a bank of constant-Q filters.... A number of studies of the human senses find that the perceptual "distance" between two stimuli is dependent upon their ratios (with respect to the appropriate units of measurement for the stimuli). • Figure 1: Power spectrum for the short-time Fourier transform of an impulse function using a Gaussian window centered at TO = 0 one wavelength.. . · .···· ••1::·. as illustrated in Figure 2b. which vary over frequency w and time T. :: :::: f ··... · ·· .. A filter bank used to process these signals naturally consists of constant-Q filters-those whose bandwidths are proportional to their center frequencies... degrading accuracy of frequency measurement. Since human perception of many types of signals has a logarithmic nature.T )e.

STFT b. Given the decomposition. we need to determine its behavior at any individual event. and s represents scale. Multiresolution Analysis In order to analyze a nonstationary signal. We then recursively repeat the procedure.a. A multiresolution analysis decomposes a signal into a smoothed version of the original signal and a set of detail information at different scales. This type of decomposition is most easily understood by thinking of a picture (which is a two dimensional signal). This blurred version of the original picture is a rendering at a slightly coarser scale. Narrowness of the fine-scale basis functions trades off frequency resolution in favor of time resolution. aside from the length of the signal itself. and generation of the smoothed image corresponds to a lowpass filtering. Each time we obtain some detail information and a more and more blurred (or smoothed) version of the original image. there is no set limit to the broadness of scale at which analysis might be performed. Figure 3 displays the square of the modulus of the continuous wavelet transform (in analogy with the power spectrum) of an impulse function. the STFT is limited in both time and frequency resolution by the fixed width of its window. we can reconstruct the original image.) Location of the impulse is clearly shown at time T = O. leaving a new picture that is slightly blurred. We remove from the picture information that distinguishes the sharpest edges. In contrast. Removal of the detail information corresponds to a bandpass filtering. Multiresolution analysis provides one means to do this. (The s axis in Figure 3 is actually the base 2 logarithm of the scale. The wavelet basis functions are self-similar: scaled in time to maintain the same number of oscillations and scaled in amplitude to maintain energy. The STFT places a varying number of waves under the same modulation envelope. Wavelet Figure 2: Comparison of basis functions for the STFT and wavelet transforms. And yet. 4 . using a Gaussian wavelet. The parameter T represents time.

using a Gaussian wavelet. (The scale axis s is labeled in exponents of two) 5 .Figure 3: Square of the modulus of the continuous wavelet transform of an impulse function at time T = 0.

and sub-band coding. noise has a specific behavior across scales. This list is by no means exhaustive-new applications are continually being discovered both in signal processing and in other domains. VM C L2(R).VM. If we let L2(R) denote the set of all functions with finite energy. which characterizes the signal's behavior at that point. that is. Here m takes the values m = 1. and hence we may write f = PM f + L m=l M Qm/. In addition. We desire an orthonormal basis 6 . we may analyze the behavior of the detail information across the different scales. This denoising technique is developed in [7]. Associated with the wavelet is a smoothing function. (4) The question now arises as to how to define the Qm. Furthermore. This projection may be formally represented by a projection operator Qm :L2(R)-+ W m . assuming that Vo = L2(R). such that PMf is the smoothed version of f. and consequently. with the wavelet defining the bandpass filter that determines the detail information. a multiresolution analysis of any function f with finite energy decomposes f into a collection of details at different scales and a smoothed version of the original function.. Furthermore.Once we have decomposed a signal this way. We can extract the regularity of a singularity. As M increases the resolution of the smoothed version of f becomes coarser. we have L2(R) = Ef!~=1 W m Ef! VM. in many cases we can separate the signal from the noise. . Table 1 provides a summary of the notation used in the rest of this paper. then the details of f at any scale m is simply a projection of f onto a subspace Wm of L 2(R). For any multiresolution analysis the W m are orthogonal both to each other and to VM. many potential application areas have been identified. 3 Wavelets and the Wavelet Transform Mathematical Definitions In this section we focus our attention on the mathematical definition of wavelets and the wavelet transform. This provides an effective means of edge detection. f is decomposed into the smoothed version PM f and M sets of details at different scales. data compression.. and hence. noise reduction. These include edge characterization. Reconstruction then yields a relatively accurate noise free approximation of the original signal. there exists another projection operator PM: L2(R) -. As discussed above. the finer detail information is contained in the scales corresponding to low values of m. .M. Conditions to be described later ensure that the set consisting of the detail information at all scales and the smoothed version of the original signal contains no redundant information. The wavelet transform specifies a multiresolution decomposition.2. In lieu of the wavelet transform's ability to localize in time and its ability to specify a multiresolution analysis. which defines the complementary lowpass filter.

L (R) f m n s T ng TO iscretization 0 t e continuous trans ation 'ljJmn s Table 1: Summary of notation 7 .

We therefore turn our attention to the definition of wavelets and the wavelet transform and then show how to define Qm in terms of them.r by the dilations and translations of ?/J..) denotes inner product. That is. Formally. Translation simply means a shift of the argument along the real axis. the analyzing wavelet may be complex valued. 4. t/J must also satisfy the admissibility condition c.of W m. Thus the subspaces Wand V are intimately related in the following way: given any m. ¢ denotes the Fourier transform of t/J. See [3. we have Vm = W m+1 EB Vm+l' Wavelets consist of the dilations and translations of a single real valued/ function ?/J E L2(R). Iteratively removing detail information progressively generates more and more smoothed versions of f. = where 1 o 00 1¢(wW dw < w 00. that is. Consequently. Before proceeding however a few remarks may provide some clarification. After removing the first set of detail information from f we are left with a slightly smoothed version of f. For this paper however we assume that it is real valued. As it turns out. ?/Jmn)?/Jmn.?/J( ~) is a dilation of?/J. called the analyzing wavelet (also known as the basic wavelet or mother wavelet). 8 .T).3 We may represent any function f E L 2(R) by (7) W f is called the continuous wavelet transform of f. given T.T(t) = s -1/2 ?/J(-s-)' s. ?/JS. We introduce the factor ~ with the foresight that it yields a normalization necessary to have an orthonormal wavelet basis.T is called a wavelet. s > O. a dilation of a function corresponds to a either a spreading out or contraction of the function. one subspace being Wm+l and the other Vm+1. T E R. We stop the process once we have enough detail information or a smooth enough version to do the analysis we desire. 3 Technically. j. a function f E Vm is decomposed into details and a smoothed version at the m + 1 scale. t- T (6) Each ?/JS. so that given any function ?/J(t) and a parameter s > 0. wavelets provide the ?/Jmn. call it ?/Jmn (m fixed). so that Qmf = L n=-oo 00 (I. 2In general. For any analyzing wavelet ?/J we thus define a family of functions ?/Js. (5) where (-. By a dilation we mean a scaling of the argument. the translation of ?/J(t) by Tis ?/J(t . f is decomposed by projecting it onto orthogonal complements in Vm . 5].

8) implies ~(o) = 0. and the translation steps become large accordingly. 1) (10) 0.n) then form an orthonormal basis of L 2(R). When m is large and positive. n E Z}. To summarize.n) = (f. When m is large (in absolute value) and negative however the "pmn are very concentrated.Given So > 1 and TO 1: 0. the choice of So significantly affects the properties of the "pmn. using Equations 4 and 11.2. with this choice of "pmn. given any function f with finite energy and an analyzing wavelet "p such that the "pmn are orthonormal. and hence i: "p(t) dt = 9 ~(o) = 0. respectively.. 6]. The "pmn = 2. m. we may compute.e.L:~=1 QmJ)' thus obtaining a multiresolution analysis. As an example. (12) . Then 'l/JS. Equation 9.. otherwise This wavelet.T becomes "pmn(t) = s~m/2"p(somt . for at each scale m is fixed and only n varies. Note that these projections are not onto all the "pmn. Taking So = 2 allows us to define "p such that the "pmn are orthonormal [3. is called the dyadic wavelet transform. let the analyzing wavelet be defined by "p( t) = { I . M and the corresponding smoothed version of f at scale M (i. "pmn is spread out. Now we return to Qm.. The choice of TO is arbitrary and by convention is taken to be 1. We define Qm as in Equation 5 by 00 Qmf = L n=-oo (j. QmJ) at the scales m = 1. is called the Haar wavelet. PM f = f . and the translation steps are small.e.n)"pmn. tEa. and W f becomes (8) (WJ)(m. On the other hand.pmn) = i: f(t)"pmn(t)dt. illustrated in Figure 4. m « Z} and T E {nsOTO.pmn). p. the details of f (i.pmn = L(WJ)(m.~) -1. . Properties and Examples Wavelets possess some interesting properties in addition to those already discussed.mt . to the discrete lattices S E {so. n (11) The details of f at each scale m thus consist of the sum of the projections of f onto the "pmn... and the corresponding "pmn form a basis of L2(R) called the Haar basis. we may restrict S and T.. The admissibility condition (see footnote. (9) Note that the translation parameter T depends upon the scaling parameter So.nTo)... For doing multiresolution analysis we want the "pmn to be orthonormal.m/2"p(2. t E [o.. This is also the conventional choice for So.. the operator that projects a function onto its details at scale m.

wavelets have zero mean. of a compactly supported wavelet that yields an orthornormal basis of L 2(R) is shown in Figure 6. Many wavelets have rapid decay. Daubechies [3] characterized all orthogonal wavelets with compact support. The first example is the Haar wavelet discussed above. The compact support of Daubechies' wavelets and the rapid decay of the wavelets described by Meyer and Lemarie and Battle help provide for both the time-localization ability and efficient computation of the wavelet transform. !>A function f: R -+ R is said to have compact support if and only if it is zero everywhere except on a closed. An example.wavelets that decay exponentially. illustrated in Figure 5.-1 Figure 4: Haar wavelet that is. Note that this indicates the equivalence of a wavelet and a bandpass filter. other than the Haar wavelet. 4coo 10 . and C k represents the space of all k-times continuously differentiable functions.f I. P. Although continuous. As an analyzing wavelet it yields an orthonormal basis of L2(R) but does not have compact support. Note the lack of symmetry. She showed that the Haar wavelet is the only such wavelet that is either symmetric or antisymmetric about any point. We finally consider some examples of wavelets. this wavelet is non-differentiable at an infinite number represents the space of all analytic functions. the support width varies directly with the regularity. G. bounded subset of R. however. Y. This wavelet was first constructed by Daubechies [3]. Meyer [3] constructed a Coo wavelet that decays faster than any power." Orthogonal wavelets may be classified as either having compact support or not having compact support. Lemarie and G. Battle [3] independently constructed a collection of c. A second example is a wavelet constructed by Lemarie and Battle [6]. She also showed that compactly supported wavelets may be chosen with arbitrary regularity.

The Fourier transform maps time into frequency and phase. '2 characterized follows [10): Given = t(1 + V3). If a real-valued analyzing wavelet is used. otherunse E 11 . Both the Fourier transform and wavelet transform may take real or complex functions as their input. as Co Cl - = t(3 + V3). However. we define .of points. Both the Fourier transform and wavelet transform are given by integral equations in the form of a correlation. In the Fourier transform the correlation is with dilations of the function e. In the wavelet transform the correlation is with dilations and translations of the analyzing wavelet '1/. If a complex-valued wavelet is used as the analyzing wavelet..or complex-valued (realvalued if the input function is real-valued and complex-valued if the input function is complex-valued ). Table 2 gives a summary.it .and complex-valued wavelets. there are both real. The output of the Fourier transform is always complex. j -+ 00 of the recursion 3 tPJ(t) with = I>ktPi-1(2t o k).V3).(t) where lP(t) is the limit as =L k=-2 1 (-1)k C 1_klP(2t . whereas. the wavelet transform may be real. the wavelet transform maps time into scale and time. For each frequency the Fourier transform 6The Battle-Lemarie wavelet is characterized by its Fourier transform [6]: where with and N2 The Daubechies wavelet is C2 = t(3 .1) . It is supported on the interval [0. IP (t) o = {I. V3). t [0.3]. = 2(sm '2 )4 ( cos '2) 2 + 70(cos '2) 4 + 3(SlR W )6 .k). the wavelet transform is complex-valued. 0. which can be any wavelet.. and C3 = t(1 • W W W 2 .6 4 Comparision of the Fourier and Wavelet Transforms In this section we discuss some of the similarities and differences between the Fourier and wavelet transforms.

Figure 5: Battle.Lemarie wavelet Figure 6: Compactly supported Daubechies wavelet 12 .

r )s-1/21/Je~T )d:tT amplitude for each scale and time Ror C Ror C Yes Yes O(n) n Yes No (limited with STFT) O(nlog n) n Table 2: Comparison of Fourier and Wavelet Transforms yields an amplitude and a phase. the Fourier transform of an impulse function contains high amplitudes at all frequencies. The width and the center frequency of the band are both proportional to -log s.r) f(t) = f~oo f(t)e-iwtdt f(t) = J~oo j(w)eiwtdJ. The wavelet transform also "localizes in time.Fourier Transform "root" function Continuous transform Inverse transform (up to a proportionality constant) Time transformed to Input domain Output range Localization in frequency Localization in time Time for fast discrete transform Number of nonredundant outputs of discrete transform Wavelet Transform eiwt j(w) s-1/2'l/Je-. Figure 7 shows the logarithmic progression of band widths for the filters corresponding to different scales. amplitude and phase for each frequency Ror C C = f~oo f(t)s-1/21/Je-. A single scale s contains information from a band of frequencies.. the wavelet transform yields an amplitude for each scale and time. A signal is represented as the sum across scales of time-centered wavelets whose amplitudes are given by the wavelet transform. For example. Each scale's band has the same ratio of bandwidth to center frequency. although the STFT 13 . STFTs. Similarly." A signal that is nonzero only during a finite span of time has a wavelet transform whose nonzero elements are concentrated around that time. "Scale" is roughly "minus log frequency". can be used to provide some localization in time. On the other hand.. the Fourier transform does not localize in time. A wavelet transform of an impulse is either contained in a finite band of time or decreases expontentially with distance from the time of the impulse. both produce an output that is nonzero only in a band when given an input that contains only frequencies from that band. so scales correspond to a set of constant-Q filters. as discussed above." That is. T)dt oo = J~oo Jo Wcf(s. T) Wcf(s. in the following limited sense. A signal may then be represented as the sum of sine waves whose phase and amplitude are given by the Fourier transform. Both the Fourier transform and wavelet transform are said to "localize in frequency.

logarithm of amplitude) 14 .. I rI .'\ \ ... ..~.... H.. " . ..~.. ' .·to · ... \ " ...- . using the cubic spline analyzing wavelet. \ . '1:' . \ .. / \ I 0' of ).. :':'-:...._ •• _ _... \ I o • I\ I I I '\..~ .......: ... (frequency vs....' I J 1 ' ..--_. " '" " f ''' . \ \ \ • ..... _ •••• .. Figure 7: Logarithmic progression of frequency band width for filters corresponding to different scales...II: 1 .... \ . \ J ..'-.:.. : " ~'. I 'Ii I 1 J' t\ '1 I I / . .. \ : \ \ \ .\. \ \ r I: 'j 0:' \: " \ \ \ \ .. '" ..... \/ \ .•••_...:..·I.••••• ...-_ ' ••••.: /\ \ ... -... -- _ ••••••••••_ ' •••••• _ •• _ #'#'#' • ~ ... \ \ " \ ." ' \ .

1. Note that the maximum amplitude of the transform decreases at each scale after Scale 2. Figure 9 shows the wavelet transform of an impulse. The discrete wavelet transform (DWT) is obtained from the wavelet transform in the same way. A DFT is usually computed using the Fast Fourier Transform (FFT) algorithm. Ws is differentiable). At first glance each scale of the transform resembles a dilated W4' However. 5 Examples We now show wavelet transforms of some simple signals. the transform under W4 is non-differentiable. 2 for 8 = log n . Since W4 is neither symmetric or antisymmetric. There are discrete versions of both the Fourier transform and the wavelet transform. Also. however. using the cubic spline wavelet. where n is the number of input values. Figures 10 and 11 show the wavelet transforms of a step function using W 4 and W s . or ~ frequencies and ~ phases). The analyzing wavelet is the so-called "cublic spline wavelet". neither is the transform. No such tradeoff is required when using the wavelet transform. neither is the transform. Figure 12 shows the wavelet transform of a triangle wave. Section 3). but the transform under W s is. Note. because W s has twice as many oscillations as W 4. Since Ws is neither symmetric or antisymmetric. each scale is finitely-supported and has an infinite number of non-differentiable points. The discrete Fourier transform (DFT) is obtained from the Fourier transform by replacing the integral with a finite sum. Each scale of the transform has more oscillations than the corresponding scale in Figure 8. ~ for 8 = 1. This wavelet is a twice continuously 15 . . Many of the properties seen in the previous two figures still hold. the wavelet transform is constant in every scale. A FWT requires D(n) time. each scale has more oscillations than the analyzing wavelet. which takes advantage of certain symmetries inherent in convolving with e iwt when n is a power of two to reduce the time to D( n log n). that the maximum amplitudes of the transform at different scales are all roughly equal. For example. using the four coefficient Daubechies wavelet W4 as the analyzing wavelet. With this analyzing wavelet the maximum amplitude of the transform decreases at every scale. The FFT takes n inputs and produces n outputs (~ complex numbers.gains its localization in time by trading off both frequency resolution and bandwidth. Figure 13 shows the wavelet transform of two steps and an impulse. this transform is smoother than the one in Figure 8. The FWT also takes n inputs and produces n outputs: 1 for 8 = log n. Figure 8 shows the wavelet transform of an impulse. Both the DFT and DWT take D(n 2 ) time. observe that where the signal is a straight line... there is a Fast Wavelet Transform (FWT) algorithm that takes advantage of symmetries that are inherent in convolving with wavelets when n is a power of two. Note that the amplitude of the transform increases from scale to scale. All of the examples were computed with the FWT using 80 = 2 (d. Just like W 4 . Just as Ws is smoother than W4 (for example. Similarly. using the eight coefficient Daubechies wavelet W s as the analyzing wavelet. .

W4 16 .Signal Scalel Scale 2 :i_ _1'-----_ :i------'t'-----:i --~_. 'Ov ScaleS Scale 4 :i :~ :f Smoothed Figure 8: Wavelet transform of an impulse using the four coefficient Daubechies wavelet.

~------ « A ov Figure 9: Wavelet transform of an impulse using the eight coefficient Daubechies wavelet.Signal _ _1 _ :l Scale! Scale 2 Scale 3 Scale 4 Smoothed :i :I :i :I :i --i~- ---------"'\. Ws 17 .

_ n \j V 7 /: Figure 10: Wavelet transform of a step using W 4 18 .Signal Scalel Scale 2 ScaleS Scale 4 Smoothed ~~ ~~ ~I ~i ~I ~! ~ t ~h ~ V .

.Signal Scalel Scale 2 Scale 3 Scale 4 Smoothed ~! ~! ~! ~! ~! ~I ~ YO 1\ . ~A 0 <> ..V - c: ] Figure 11: Wavelet transform of a step using W s 19 .-..

. :I~~~ O~"'7 1~ ~ "'7 ScaleS Scale 4 Smoothed 1~ ~ o~~ ~ Figure 12: Wavelet transform of a triangle wave using a cubic spline wavelet 20 ..Scalel Scale 2 :1--:1---:I==---==-~ .

n C> y A V /\ V' L\.:::. v .Signal Scalel Scale 2 Scale 3 Scale 4 Smoothed :1 :1 :1 :1 :1 :1 A t .. ucc.::>"" :7 L ~ Figure 13: Wavelet transform of the sum of an impulse and two steps using a cubic spline wavelet 21 ....

The cubic spline wavelet does not have finite support. based upon our experience of reading the papers. (See. With each listing is a short description of the contents. Note that the transform peaks generated by the noise decay like the wavelet transform of an impulse. the peak resulting from the downwards step is opposite in sign and half the magnitude of the peak resulting from the upwards step. Much of the current literature utilizes a high level of mathematical terminology.5 to O. However. Detection of short duration events. This is visible in Figure 13. the fast wavelet transform generates a multiresolution analysis in O( n) time. In the wavelet transform.0. Wavelet transforms exist that project a finite energy function onto to an orthonormal basis of L 2(R). [7]). Most of the papers require an understanding of Fourier analysis and sometimes an understanding of more general functional analysis principles. on the other hand. The signal in Figure 14 is the signal of Figure 13 plus noise uniformly distributed on [-0. In the signal the upwards step goes from -0. As with the Fourier transform. we adopt a relative rating scale. note that the wavelet transform of the impulse decays from scale to scale.5 to 0. we provide an annotated bibliography that includes some of the key papers in the field. In Figure 13. wavelet transforms readily lend themselves to nonstationary signal analysis. The wavelet transform of the two steps and impulse is the sum of the wavelet transforms of each step and of the impulse. These oscilations are invisible in the figure due to the rapid decay of the wavelet Figure 13 also demonstrates the linearity of wavelet transforms. Unlike W 4 and W s. However. Our intent in this paper was to present the basic concept of the wavelet transform from a viewpoint that targets signal analysis applications.differentiable. at each scale. piecewise cubic function. it is antisymmetric. To help specify the mathematical sophistication of a paper. where the transform is symmetric about the steps and impulse. Again. a fast Fourier transform takes O( n log n) time. Our hope was to provide an brief introduction to the primary underlying ideas in a relatively intuitive manner. it does decay exponentially. each scale of the transform contains small amplitude oscillations that extend infintely in both directions. For those who are interested. The corresponding multiresolution analysis decomposes the function into a set of details at different resolutions and a smoothed version of the original function. 6 Conclusion Due to their capability to localize in time. 1 meaning little or no mathematical sophistication 22 . a "fast wavelet transform" exists. are limited in Fourier analysis by the width ofthe windowing function used in the short-time Fourier transform.3]. This observation is the basis of several signal denoising algorithms in the literature. but the noise peaks are quite small.3.5. for example. whereas. and the downwards step goes from 0. but the wavelet transform of the steps stay roughly constant. By Scale 4 the peaks generated by the step edges are pronounced. This antisymmetry causes the wavelet transform of symmetric signal features to be symmetric.

u.v 4pJ it A 9V\i V 0 Scale 4 :I------. --..~ V v 1\ .. a step and uniform noise using a cubic spline wavelet 23 .-------_/ Figure 14: Wavelet transform of the sum of an impulse.L--~::o... _" A.. \V. f V " " A..Signal Scalel Scale 2 o Scale 3 ...+ eo ••'VV.. A :I-........ ..~~ .. :c::::a "=""- Smoothed L :I---.--.::~~ _ e o _ U <> 'I ~~rOVV4 ...

relative to the other papers.required and 5 meaning a high level of mathematical sophistication required. 24 .

the Laplacian pyramid decomposition scheme. its inverse. and introduces the wavelet transform.4. 15." SIAM Journal of Mathematical Analysis. "The Fast Wavelet Transform. reviews the short-time Fourier transform. "Orthonormal Bases of Compactly Supported Wavelets. A." Communications on Pure and Applied Mathematics. and zero-crossings of multifrequency channels." Dr. pp. and Morlet. M. It starts by discussing multifrequency channel decompositions in psychophysics. She then demonstrates the classification of all orthonormal bases of wavelets with compact support. Speech. pp. Mathematical level-L [2] Wavelets: Time-Frequency Methods and Phase Space. Dobb's Journal. The papers are organized into the five categories: Part Part Part Part Part I: Introduction to Wavelet Transforms II: Some Topics in Signal Analysis III: Wavelets and Signal Processing IV: Mathematics and Mathematical Physics V: Implementations Mathematical level-l through 5 [3] Daubechies. et al. Ingrid. and the fast wavelet transform algorithm for doing a multiresolution analysis using a wavelet basis. December 14-18. the admissibility condition. France. 723-736. Mathematical level-3 25 . and Signal Processing. Vol. J. This book presents some of the earlier papers in the development of wavelets. Mac A. Vol. 1988. 2091-2110. "Multifrequency Channel Decompositions of Images and Wavelet Models. July 1984.. Springer-Verlag. They discuss the definition of the transform. In this paper Daubechies reviews multiresolution analysis. "Decomposition of Hardy Functions into Square Integrable Wavelets of Constant Shape. pp. S. Mathematicallevel-4 [4] Grossman. Proceedings of the International Conference. December 1989.References [1] Cody. We found this paper to be a good introductory paper to the wavelet transform. 909-996. This paper provides a quick overview of wavelets and the fast wavelet transform. two dimensional wavelet transforms." IEEE Transactions on Acoustics. eds. It then proceeds to discuss pyramidal multiresolution decompositions. No. 12. Combes. April 1992. Mathematical level-5 [5] Mallat.7. J. No. 41. No. Marseille. 1987. and they prove that the wavelet transform is an isometry. It also contains and discusses some C code that implements a wavelet transform. Vol. 37. This is the paper primarily responsible for the current interest in wavelet transforms.

This is an introductory paper to wavelet transforms.[6] Mallat. July 1989. pp. Mallat here presents his O( n) algorithm for computing the fast wavelet transform. Mathematical level-3 through 5 [10] Strang. 617643. 1992. Furthermore. W. Jones and Bartlett. "A Theory for Multiresolution Signal Decomposition: the Wavelet Representation. They discuss the detection and measurement of oscillating and nonoscillating singularities and illustrate a technique for denoising a signal based upon its wavelet transform modulus local maxima. O. In this paper Mallat and Hwang discuss a means of using the wavelet transform to characterize the Lipschitz regularity of a function at a point. S. "Wavelets and Dilation Equations: A Brief Introduction. 38. March 1992. December 1989. 674-693. pp. eds. Vol. M. pp. G. "Singularity Detection and Processing with Wavelets. Mathematicallevel-4 [8] Rioul. Vol. No. 614-627. and Vetterli.7." IEEE Transactions on Injormation Theory. S. he discusses some of the technical conditions imposed upon the lowpass and bandpass filters necessary for them to correspond to the projection operators that generate the multiresolution analysis. He concludes by discussing image processing applications of the wavelet transform and by giving an example of a multiresolution approximation. Ruskai. 31. No. This book consists of a collection of papers organized into the categories Signal Analysis Numerical Analysis Other Applications Theoretical Developments and contains many of the more recent papers. 11. Strang presents a brief mathematical introduction to multiresolution analysis and the fast wavelet tranform algorithm. October 1991. No. B. L. Vol. et al.2. It well illustrates the difference between the short-time Fourier transform and the wavelet transform." SIAM Review. 14-38. the paper is rather terse in its explanations. pp. Mathematicallevel-2 [9] Wavelet and Their Applications. (Daubechies further discusses these conditions in her paper cited above). However. The paper contains a relatively complete set of references that we found very useful. Mathematical level-3 26 .4." IEEE Transactions on Pattern Analysis and Machine Intellegence. "Wavelets and Signal Processing. Mathematicallevel-3 [7] Mallat. M." IEEE Signal Processing Magazine. and Hwang.

Vol.[11] IEEE Transactions on Information Theory. Mathematicallevel-3 through 4 27 . No. March 1992. It contains many recent papers on the applications of wavelets in information theory related areas. Part II of this issue is dedicated to wavelets and their applications. Part II.2. 38.

You're Reading a Free Preview

Descarregar
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->