A wavelet tutorial from S. Mallat's book
Academic Press, 1998
A SHORT PRESENTATION BY F.
for those who hate preambles
Hardware and Software Requirements
This site has been tested with Netscape Navigator 2 and 3. Version 2 fails to render characters from the Symbol font, but this does does seriously affect the readability of the presentation. A 256 gray level display is required to view the illustrations. Some
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Wavetour_presentation_US.html (1 of 4) [11/15/2003 9:42:50 PM]
A wavelet tutorial from S. Mallat's book
graphs use color plots for better reading. Most images are interlaced GIF with transparency color set to white.
Warning
This presentation is inspired from S.G. Mallat's book and does not pretend to reflect it exactly. It is concerned with the following topics:
● Fourier analysis (chapter 2)
● timefrequency analysis (chapter 4, except for the quadratic energy distributions))
● frames (chapter 5)
● singularity analysis and reconstruction (chapter 6 except for the multifractals)
● wavelet bases and filter banks (chapter 7)
The following topics from the book are not covered here:
● chapter 3 about discrete signals (except for the FFT and convolutions algorithms, which are briefly described)
● chapter 8 on wavelet packets and local cosine bases
● chapter 9 on approximation
● chapter 10 about estimation (which is being revised by S.G. Mallat for the French edition and the second US edition)
● chapter 11 on compression and coding (hope to do it someday)
Proposed Tours
Four tours are proposed, corresponding to four different topics. These tours are linked to each other sequentially. Many links allow navigation from one to topic to another for a nonlinear browsing.
● Frequency Analysis
● From dyadic wavelets to filters
● From filters to dyadic wavelets
● Regularity Analysis
● Frames
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Wavetour_presentation_US.html (2 of 4) [11/15/2003 9:42:50 PM]
A wavelet tutorial from S. Mallat's book
Index
For direct access, here is a list of links that point to the main topics:
Fourier Transform Instantaneous Frequency of an Analytic Signal TimeFrequency Localization Windowed Fourier Transform and Wavelet Transform Frames and Riesz Bases Windowed Fourier Frames and Wavelet Frames Multiresolution Approximations Wavelet Bases Filter Banks Regularity Analysis of a Signal Detection of Singularities Reconstruction from Dyadic Maxima Edge Detection and Image Reconstruction
Algorithms:
Fast Fourier Transforms and Convolutions Fast Windowed Fourier Transform Fast Wavelet Transform Fast Dyadic Wavelet Transform Decomposition and Reconstruction over Orthonormal Wavelet Bases
Numerical Computations
All numerical figures in the book have been computed using Wavelab, a freeware Matlab Toolbox, available at
Uvi Wave is another freeware Matlab Toolbox.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Wavetour_presentation_US.html (3 of 4) [11/15/2003 9:42:50 PM]
A wavelet tutorial from S. Mallat's book
Updates:
June 4, 1999: Full Strang and Fix conditions added, with proof!
May 18, 2001: added mathematical transition from filter banks to multiresolution analysis
Last update: February 2, 1999
Feedback is welcome.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Wavetour_presentation_US.html (4 of 4) [11/15/2003 9:42:50 PM]
Frequency and Period
Frequency Analysis
As shown in this tour, the notion of instantaneous frequency should be handled with caution. This is why we recall the primary definition of a frequency.
Note: this presentation is proper to the site and does not come from the book.
Period and Frequency
What is a frequency?
A frequency is the inverse of a period.
A one dimensional signal is T>0 periodic if it is unchanged by a
translation of T. Hence its support has an infinite length. Nonetheless, the signal is entirely determined by its values over an interval of length T.
In the representation over an interval of length T, the regularity of the signal implies the equality of the values of the functions and of its derivatives at the left and right ends of the interval.
A musical example
A purely synthetic music note can be represented by a sinusoidal
wave. An instrumental note that is held is a more complex signal.
A music or a speech recording is even more complex; in particular,
the frequencies may vary with time.
As a first approximation, a music piece may be modelized as a sequence of signals defined over intervals whose length is
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Intro/Periodicity_US.html (1 of 2) [11/15/2003 9:42:52 PM]
Frequency and Period
determined by the tempo. Every elementary signal can be periodized.
Fourier Analysis
Classically, the analysis of a signal as a Fourier series or a Fourier integral provides a representation of its frequency contents.
Fourier series
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Intro/Periodicity_US.html (2 of 2) [11/15/2003 9:42:52 PM]
Fourier Series
Fourier Series
An Nperiodic signal can be represented as a series of harmonically related sinusoids.
Fourier has shown that any periodic signal can be decomposed into harmonically related sinusoids:
This analysis seems to fits the needs of music coding by using a Fourier series for every tempo unit. It has several drawbacks:
● 
there may be points where the Fourier series does not converge 

● 
a 
periodic continuous signal with a period different from 2 is 

represented as a discontinuous signal over 

. This causes 

a 
bad convergence of its Fourier series. 

● 
More generally, it cannot efficiently represent signals which are not synchronized with a tempo, such as speech recordings or images. 
A solution consists in using a transform where the frequency is allowed to have continuous variations. The Fourier series becomes the Fourier transform. The concept of period disappears in the process, and the frequency contents is actually a sinusoid contents.
Fourier transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/Fourier_Series.html [11/15/2003 9:42:54 PM]
Fourier transform
Fourier Transform
The Fourier transform analyses the "frequency contents" of a signal.
Its many properties make it suitable for studying linear time invariant operators, such as differentiation.
It is a global representation of a signal.
Fourier transform
The Fourier transform of f in L 2 is
The inverse Fourier transform represents f as a sum of sinusoids
Properties
The Fourier transform has many algebraic properties. Note that sinusoidal waves are eigenvectors of the differentiation operator.
This makes it possible for the Fourier transform to give indications
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/Fourier_Transform.html (1 of 2) [11/15/2003 9:42:57 PM]
Fourier transform
on the regularity of a signal.
Implementation
To reduce the number of operations, the Fast Fourier Transform separates odd and even frequencies when computing a discrete Fourier transform.
A global representation
The Fourier transform is a global representation of the signal. It cannot analyze it local frequency contents or its local regularity. The convergence condition on the Fourier transform only gives the worst order of regularity. It ignores local regular behaviours.
There exist, however, a definition of the instantaneous frequency of
an analytic signal.
It is useless in practice because it fails to detect the summation of two signals. It is nonetheless a convenient means of defining synthetic signals for numerical experimentations.
Trying to discriminate each of the stacked frequencies leads to a frequency analysis that is localized in frequency as well as in time. This requires some understanding of the timefrequency localization of signal.
TimeFrequency Localization
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/Fourier_Transform.html (2 of 2) [11/15/2003 9:42:57 PM]
Algebraic properties of the Fourier transform
Algebraic Properties of the Fourier Transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/Agebraic_Properties.html [11/15/2003 9:43:00 PM]
Regularity analysis by means of the Fourier transform
Regularity Analysis by Means of the Fourier Transform
A function f is bounded and p times continuously differentiable with bounded derivatives if
This property is extended to Lipschitz regularity:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/Regularity_with_Fourier.html [11/15/2003 9:43:13 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/LipschitzDefUS.GIF
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/LipschitzDefUS.GIF [11/15/2003 9:43:15 PM]
Fast Fourier Transform
Fast Fourier Transform
The discrete Fourier transform of a discrete signal with N samples is
Using this formula to compute the Fourier transform requires N 2 complex additions and multiplications.
A simple computation shows that the even frequency coefficients are the coefficients of the Fourier transform of the N/2 periodic signal
f _{p} [n] = f[n] + f[n+N/2]
and that the odd frequency coefficients are the coefficients of the Fourier transform of
f _{i} [n] = (f[n]f[n+N/2]) e 2i n/N
One verifies by induction that the number of operations required by this method to compute the Fourier transform is of the order of KN log _{2} (N), where K is a constant which does not depend on N.
This is the basic principle of the Fast Fourier Transform. Variants exist that reduce K.
Convolutions and circular convolutions
The circular convolution of two N periodic signals is defined by
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/FFTUS.html (1 of 2) [11/15/2003 9:43:21 PM]
Fast Fourier Transform
The discrete Fourier transform of a circular convolution is the product of the two discrete Fourier transforms.
For two signals f and h with a length M>=32, computing their convolution with an FFT is faster than using the straightforward formula. For that purpose, two M periodic signals are defined:
and one can verify that their circular convolution is equal to the convolution
for 0<=n<2M. The circular convolution is itself computed by performing the FFT of the two signals, computing their product and then its inverse FFT.
Fourier Transform
Fast Windowed Fourier Transform
Fast Wavelet Transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier/FFTUS.html (2 of 2) [11/15/2003 9:43:21 PM]
Instantanous Frequency of an Analytic Signal
Instantaneous Frequency of an Analytic Signal
A signal in L 2 is analytic if its Fourier transform is zero for
negative frequencies. The analytic part f _{a} of a real signal f is given
by its Fourier transform:
and f _{a} can properly be decomposed into a module and a complex phase:
The instantanneous frequency is the nonnegative derivative of the complex phase:
For a sinusoidal wave, this definition coincides with the usual
frequency. Unfortunately, the instantaneous frequency of the sum
of two ordinary sinusoidal waves is the average of their
frequencies, which does not coincide with the result of a Fourier analysis.
Hence the caracterization of the instaneous frequency of a signal in a sense that is consistent with the Fourier analysis in simple cases
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Intro/Instantanous_Frequency.html (1 of 2) [11/15/2003 9:43:29 PM]
Instantanous Frequency of an Analytic Signal
requires other mathematical tools.
However, analytic signals are very useful in the synthesis of signals with a given time varying frequency contents.
Fourier transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Intro/Instantanous_Frequency.html (2 of 2) [11/15/2003 9:43:29 PM]
TimeFrequency Atoms
TimeFrequency Localization
There is no finite energy function which is compactly supported both in the time and frequency domains.
The timefrequency localization is measured in the mean squares sense and is represented as a Heisenberg box.
The Fourier transform can be viewed as a representation of a function as a sum of sinusoidal waves. These sinusoids are very well localized in the frequency, but not in time, since their support has an infinite length. This is a consequence of periodicity.
To represent the frequency behavior of a signal locally in time, the signal should be analyzed by functions which are localized both in time and frequency, for instance, signals that are compactly supported in the time and Fourier domains.
This timefrequency localization is limited by the following two results:
The uncertainty theorem of Heisenberg.
If f is in L 2 , then its time root deviation
deviation
are defined. Then
and its Fourier root
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/atomes/Timefrequency_atoms.html (1 of 3) [11/15/2003 9:43:34 PM]
TimeFrequency Atoms
A balance has to be reached between the time and frequency
resolution. In the limit case of a sinusoid, infinite.
is zero and
is
The previous inequality is an equality if and only if f is a Gabor
chirp .
Compact supports
If f is non zero with a compact support, then its Fourier transform cannot be zero on a whole interval. Similarly, if its Fourier transform is compactly supported, then it cannot be zero on a time interval.
Hence, even if the Heisenberg constraints are verified, it is impossible to have an function in L 2 which is compactly supported both in the time and Fourier domains.
In particular, this means that there is no instantaneous frequency analysis for finite energy signals.
Timefrequency localization is thus achievable only in the mean squares sense.
This localization is represented as a Heisenberg box.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/atomes/Timefrequency_atoms.html (2 of 3) [11/15/2003 9:43:34 PM]
TimeFrequency Atoms
For a family of vectors to be a basis of L 2 , it is reasonable to expect that their Heisenberg boxes pave the time frequency plane.
Two time frequency localization strategies are presented in parallel; the first one leads to the windowed Fourier transform, while the other one leads to the wavelet transform.
Windowed Fourier transforms and wavelet transforms
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/atomes/Timefrequency_atoms.html (3 of 3) [11/15/2003 9:43:34 PM]
Fast Windowed Fourier Transform
Fast Windowed Fourier Transform
The discrete windowed Fourier transform of an N periodic signal is
For a fixed m, the formula is that of a discrete Fourier transform. Hence the computation is performed with N fast Fourier
transforms, which require O(N 2 log _{2} (N)) operations.
Windowed Fourier Transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation
sformees/Fourier_Fenetre/Fast_Windowed_Fourier.html
[11/15/2003 9:43:35 PM]
Windowed Fourier Transform
Windowed Fourier Transform
The windowed Fourier transform replaces the Fourier transform's sinusoidal wave by the product of a sinusoid and a window which is localized in time.
It takes two arguments: time and frequency.
Outline
The windowed Fourier transform is defined by
It uses an atom which is the product of a sinusoidal wave with a finite energy symmetric window g. The windowed Fourier transform family of atoms is obtained by time translations and frequency modulations of the original window:
This atom has a frequency center and is symmetric with respect to u.
The time and frequency spreads of these functions are constant. The family is generated by time and frequency translations of one atom. Here are examples of Heisenberg boxes of windowed Fourier
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
nsformees/Fourier_Fenetre/Windowed_Fourier.html
(1 of 3) [11/15/2003 9:43:40 PM]
Windowed Fourier Transform
atoms:
Properties
The windowed Fourier transform has a constant time frequency resolution. This resolution can be can be changed by rescaling the window g. It is a complete, stable, redundant representation of the signal. Hence it is invertible. The redundancy implies the existence of a reproducing kernel.
Spectrogram
The square modulus of the windowed Fourier transform is the spectrogram of a signal:
Choice of Window
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
nsformees/Fourier_Fenetre/Windowed_Fourier.html
(2 of 3) [11/15/2003 9:43:40 PM]
Windowed Fourier Transform
The properties of the windowed Fourier transform are determined by the window g, or rather its Fourier transform, whose energy should be concentrated around 0. This energy spread is measured by three parameters.
Implementation
The Fast Windowed Fourier Transform is equivalent to a sequence of FFTs.
Specrogram examples
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
nsformees/Fourier_Fenetre/Windowed_Fourier.html
(3 of 3) [11/15/2003 9:43:40 PM]
Fast Wavelet Transform
Fast Wavelet Transform
The discrete wavelet transform uses a discrete sequence of scales a j for j<0 with a=2 1/v , wher V is an integer, called the number of voices in the octave.
The wavelet support is assumed to be [K/2,K/2]. For a signal of size N and 1<=a j <=N/K, a discrete wavelet is defined by sampling the scale at a j and time (for scale 1) at its integer values, that is
The signal and wavelet are Nperiodized. The discrete wavelet transform of f is
which is a circular convolution between f and _{1} [n]= [n]. These circular convolutions are computed with an FFT which requires O(N log _{2} (N)) operations.
The scalogram is computed from the wavelet transform and a parobolic interpolation is performed between three succesive scales a j to better localize the wavelet ridges.
Wavelet transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Fast_Wavelet_Transf.html [11/15/2003 9:43:42 PM]
Wavelet Transform
Wavelet Transform
The wavelet transform replaces the Fourier transform's sinusoidal waves by a family generated by translations and dilations of a window called a wavelet.
It takes two arguments: time and scale.
Outline
The wavelet transform is defined by
where the base atom is a zero average function, centered around zero with a finite energy. The family of vectors is obtained by translations and dilatations of the base atom:
The previous function is centered around u, like the windowed Fourier atom. If the frequency center of is , then the frequency center of the dilated function is /s.
Its time spread is proportional to s. Its frequency spread is proportional to the inverse of s. Here is an example of
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Transform.html (1 of 4) [11/15/2003 9:43:47 PM]
Wavelet Transform
Heisenberg boxes of wavelet atoms:
At the finer scales, more Heisenberg boxes can be placed side to side because there is a better time resolution.
Properties
The wavelet transform has thus a time frequency resolution which depends on the scale s. Under the condition
it is a complete, stable and redundant representation of the signal; in particular, the wavelet transform is left invertible. The redundancy implies the existence of a reproducing kernel.
Scalogram
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Transform.html (2 of 4) [11/15/2003 9:43:47 PM]
Wavelet Transform
If denotes the frequency center of the base wavelet, then the
frequency center of a dilated wavelet is = /s. The scalogram of a signal is defined by
The normalized scalogram is
Choice of Window
As far as the continuous wavelet transform is concerned, a wavelet
is simply a finite energy function with a zero mean. Besides its
Heisenberg box, the most important feature of a wavelet is the number of its vanishing moments:
The vanishing moments property makes it possible to analyse the local regularity of a signal.
A theorem caracterizes fast decaying wavelets with n vanishing
moments as the n th derivatives of a fast decaying function.
Implementation
The wavelet transform is computed with a Fast Wavelet Transform.
It computes a discrete transform with circular convolutions, which
are themselves computed with a FFT.
To speed up computations, dyadic wavelets are often used. The dyadic wavelet transform is implemented by filter banks.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Transform.html (3 of 4) [11/15/2003 9:43:47 PM]
Wavelet Transform
Scalogram examples
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Transform.html (4 of 4) [11/15/2003 9:43:47 PM]
Chirps
Chirps are analytic signals which have a particular instantaneous frequency.
A gaussian chirp is defined by
A linear chirp is defined by
A quadratic chirp is defined by
A hyperbolic chirp is defined by
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Chirps/ChirpsUS.html [11/15/2003 9:43:49 PM]
Gabor Chirps and Wavelets
f is a Gabor chirp if there is
such that
A Gabor wavelet is a particular Gabor chirp
with
For
the wavelet is approximatively analytic.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Gabor/Gabor_Wavelet.html [11/15/2003 9:43:53 PM]
Heisenberg Box
Heisenberg Box
The "time frequency localization" of an atom is represented as a "Heisenberg box" located in the time frequency plane, which is a
rectangle with a time width
frequecy center which coincides with the signal's.
and a frequency heigth
, and time
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/atomes/Heisenberg_box.html [11/15/2003 9:43:55 PM]
Time Frequency Transforms
Windowed Fourier Transform
The windowed Fourier transform replaces the Fourier transform's sinusoidal wave by the product of a sinusoid and a window which is localized in time.
It takes two arguments: time and frequency.
Outline
The windowed Fourier transform is defined by
It uses an atom which is the product of a sinusoidal wave with a finite energy symmetric window g. The windowed Fourier transform family of atoms is obtained by time translations and frequency modulations of the original window:
Wavelet Transform
The wavelet transform replaces the Fourier transform's sinusoidal waves by a family generated by translations and dilations of a window called a wavelet.
It takes two arguments: time and scale.
Outline
The wavelet transform is defined by
where the base atom is a zero average function, centered around zero with a finite energy. The family of vectors is obtained by translations and
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (1 of 6) [11/15/2003 9:43:57 PM]
Time Frequency Transforms
This atom has a frequency center and is symmetric with respect to u.
The time and frequency spreads of these functions are constant. The family is generated by time and frequency translations of one atom. Here are examples of Heisenberg boxes of windowed Fourier atoms:
dilatations of the base atom:
The previous function is centered around u, like the windowed Fourier atom. If the frequency center of is , then the frequency center of the dilated function is /s.
Its time spread is proportional to s. Its frequency spread is proportional to the inverse of s. Here is an example of Heisenberg boxes of wavelet atoms:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (2 of 6) [11/15/2003 9:43:57 PM]
Time Frequency Transforms
Properties
The windowed Fourier transform has a constant time frequency resolution. This resolution can be can be changed by rescaling the window g. It is a complete, stable, redundant representation of the signal. Hence it is invertible. The redundancy implies the existence of a reproducing kernel.
Spectrogram
The square modulus of the windowed Fourier transform is the spectrogram of a signal:
Choice of Window
The properties of the windowed Fourier transform are determined by the window g, or rather its Fourier transform, whose energy should be concentrated around 0. This energy spread is measured by three parameters.
Implementation
The Fast Windowed Fourier Transform is equivalent to a sequence of FFTs.
At the finer scales, more Heisenberg boxes can be placed side to side because there is a better time resolution.
Properties
The wavelet transform has thus a time frequency resolution which depends on the scale s. Under the condition
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (3 of 6) [11/15/2003 9:43:57 PM]
Time Frequency Transforms
Specrogram examples
it is a complete, stable and redundant representation of the signal; in particular, the wavelet transform is left invertible. The redundancy implies the existence of a reproducing kernel.
Scalogram
If denotes the frequency center of the base wavelet, then the frequency center of a dilated wavelet is = /s. The scalogram of a signal is defined by
The normalized scalogram is
Choice of Window
As far as the continuous wavelet transform is
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (4 of 6) [11/15/2003 9:43:57 PM]
Time Frequency Transforms
concerned, a wavelet is simply a finite energy function with a zero mean. Besides its Heisenberg box, the most important feature of a wavelet is the number of its vanishing moments:
The vanishing moments property makes it possible to analyse the local regularity of a signal.
A theorem caracterizes fast decaying wavelets
with n vanishing moments as the n th derivatives of a fast decaying function.
Implementation
The wavelet transform is computed with a Fast
Wavelet Transform. It computes a discrete transform with circular convolutions, which are themselves computed with a FFT.
To speed up computations, dyadic wavelets are often used. The dyadic wavelet transform is implemented by filter banks.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (5 of 6) [11/15/2003 9:43:57 PM]
Time Frequency Transforms
Scalogram examples
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Transforms.html (6 of 6) [11/15/2003 9:43:57 PM]
Choice of WFT window
Choice of Fourier Window
The properties of the windowed Fourier transform are determined by the window g, or rather its Fourier transform, whose energy should be concentrated around 0. Three important parameters evaluate this energy spread:
● the root mean square bandwidth defined by
If is small, the energy of the window is well concentrated around 0.
● the maximal amplitude A of its first side lobes, measured in decibels:
http://cas.ensmp.fr/~chaplais/Wavetour_present
mees/Fourier_Fenetre/Choice_of_WFT_window.html
(1 of 2) [11/15/2003 9:43:59 PM]
Choice of WFT window
These side lobes can create "shadows" on each side of the frequency center.
● The polynomial exponent which describes the decay of the window's Fourier transform at high frequencies
It sums up the behavior of the Fourier transform beyond the first side lobes.
The following table gives the values of these parameters for classical windows, normalized so that g(0) = 1.
g(t)
1
0,54 + 0,46 cos(2 t)
_{e}_{x}_{p}_{(}_{}_{1}_{8}_{t} 2 _{)}
_{c}_{o}_{s} 2 _{(} _{} _{t}_{)}
Blackman 0,42 + 0,5 cos(2 t) + 0,08 cos(4 t)
http://cas.ensmp.fr/~chaplais/Wavetour_present
mees/Fourier_Fenetre/Choice_of_WFT_window.html
(2 of 2) [11/15/2003 9:43:59 PM]
Spectrogram Examples
Spectrogram Examples
These few examples are made of analytic signals whose intanteous frequencies are known.
They illustrate the windowed Fourier transform's ability to localize instaneous frequencies.
Here is the sum of two parallel linear chirps with its spectrogram.
Now a synthetic signal which is the sum of a linear chirp with an increasing frequency, a quadratic chirp with a decreasing frequency, and two modulated gaussians. Below is its spectrogram and the complex phase of its windowed Fourier transform, computed with a gaussian window.
The components of this synthetic signal have explicit instantaneous frequencies.
http://cas.ensmp.fr/~chaplais/Wavetour_present
mees/Fourier_Fenetre/Spectrogram_Examples.html
(1 of 3) [11/15/2003 9:44:04 PM]
Spectrogram Examples
Here is the sum of two hyperbolic chirps and its spectrogram.
http://cas.ensmp.fr/~chaplais/Wavetour_present
mees/Fourier_Fenetre/Spectrogram_Examples.html
(2 of 3) [11/15/2003 9:44:04 PM]
Spectrogram Examples
The "instantaneous frequencies" are well tracked in the first example. On the contrary, the spectrogram loses the frequency of the hyperbolic chirp when it becomes high too fast. This is due to the fixed time resolution of the windowed Fourier transform.
In these three examples, it seems that the instantaneous frequencies is traced by the spectrogram's mawima, provided these frequencies are not too close.
Windowed Fourier Ridges and Instantaneous Frequencies
http://cas.ensmp.fr/~chaplais/Wavetour_present
mees/Fourier_Fenetre/Spectrogram_Examples.html
(3 of 3) [11/15/2003 9:44:04 PM]
Regularity Analysis
Regularity Analysis
The Fourier transform analyses the global regularity of a function.
The wavelet transform makes it possible to analyze the pointwise regularity of a function.
A signal is regular if it can be locally approximated by a polynomial. The definition of the Lipschitz regularity is
Fourier condition
Naturally, this a global regularity condition.
To get conditions on the local or even pointwise regularity of a signal, it is necessary to use a transform which is localized in time.
Wavelet Transform Condition
Assume that the wavelet has n vanishing moments:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Regularity.html (1 of 3) [11/15/2003 9:44:23 PM]
Regularity Analysis
and has n continuous derivatives with a fast decay.
A
derivative of a fast decaying function.
fast decaying wavelet has n vanishing moments if and only if its is the n th
If f is a function which is a little bit more than n times differentiable at point v,
then it can be approximated by a polynomial of degree n. The wavelet transform of this polynomial is zero; around v, its order is that of the error between the polynomial and the function. If this error can be uniformly estimated on an interval, this yields a tool for regularity analysis on an interval.
This condition relates the pointwise regularity of a signal to the decay of its wavelet transform's modulus.
It can be extended to an interval and, of course, to the whole real axis.
Example
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Regularity.html (2 of 3) [11/15/2003 9:44:23 PM]
Regularity Analysis
A signal and its wavelet transform, computed with the derivative of a Gaussian. Finer scales are at top. Zero coefficients are represented by a medium gray. Hence, the regular parts are medium gray. Notice the cones below the singularities.
Detection of Singularities
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Regularity.html (3 of 3) [11/15/2003 9:44:23 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/momentsUS.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/momentsUS.gif [11/15/2003 9:44:25 PM]
Multiresolution Approximation
Multiresolution Approximations Are The Foundation of Dyadic Wavelets
Dyadic wavelets are wavelets which satisfy an additional scaling property.
This property allows the implementation of a Fast Dyadic Wavelet Transform with filter banks.
The definition of dyadic wavelets comes from the definition of multiresolution approximations.
While browsing these pages, you certainly have downloaded interlaced GIF images. During the download, a progressively detailed image is displayed on screen.
This idea of consecutive approximations at finer and finer resolutions is formalized by the concept of multiresolution approximation (or multiresolution analysis).
Definition
A sequence V _{j} , j in Z, of subspaces of L 2 (R) is a multiresolution approximation if the six following conditions are satisfied:
Condition
Interpretation
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Approx.html (1 of 3) [11/15/2003 9:44:26 PM]
Multiresolution Approximation
V _{j}_{+}_{1} is obtained from V _{j} by
a factor 2 rescaling
For any j, V _{j}_{+}_{1} is a subspace of V _{j}
V _{j} is 2 j translation
invariant.
The intersection of the V _{j} is
0 in L 2 .
The union of the V _{j} is dense
in L 2 .
There is a function such that the integer translations of make a Riesz basis of V _{0} .
There exists an underlying dyadic sequence of time grids, e.g., the intervals satisfy a geometric progression with reason 2.
Any low resolution signal is also a high resolution signal.
There is an underlying time grid with step 2 j . Condition 1 shows that this grid is obtained from the case j=0 by a 2 j rescaling.
A a zero resolution, the only finite energy signal is 0.
At the infinite resolution, all finite energy signals are perfectly reproduced.
The resolution V _{j} is generated by a
basis which is obtained by 2 j translations of a 2 j rescaled . A Riesz basis is a frame of linearly independent vectors.
A less literary definition is available.
The rescaling of does not modify the area of its Heisenberg box, but it changes the proportions of the box, like for non dyadic
wavelets.
Examples of multiresolution
approximations
What of wavelets?
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Approx.html (2 of 3) [11/15/2003 9:44:26 PM]
Multiresolution Approximation
The wavelets are used to build a basis in which are represented the details that are gained between a resolution and the next finer one.
Properties
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Approx.html (3 of 3) [11/15/2003 9:44:26 PM]
Algorithme a trous
Fast Dyadic Transform Algorithme à trous
The fast dyadic wavelet transform is implemented using filter banks. This implementation is very close to the implementation of the fast (bi)orthogonal wavelet transform, except that no subsampling is performed.
For any j>=0, let
and the discrete data is likened to the a _{0} [n]. We also define
For a given filter x with coefficients x[n], x _{j} [n] denotes the filter obtained by
inserting 2 ^{j} 1 zeroes between every x coefficient (hence the French name "algorithme à trous", which means "holes algorithm"), and let
The algorithme à trous computes the fast dyadic wavelet transform in the following way:
(the ~ filters are the dual filters of the biorthogonal system).
Compare this algorithm to the decomposition and reconstruction algorithm over a basis of biorthogonal wavelets. In the decomposition case, the data is convolved with the symmetrized filter, then the output is subsampled. Here the filter is "stretched" to take into account the rescaling and the convolution is performed without any subsampling.
Here is a scheme of the filter bank:
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
ndelettes%20dyadiques/Algorithme_a_trousUS.html
(1 of 2) [11/15/2003 9:44:31 PM]
Algorithme a trous
Dyadic Wavelet Transform
Reconstruction From Dyadic Maxima
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
ndelettes%20dyadiques/Algorithme_a_trousUS.html
(2 of 2) [11/15/2003 9:44:31 PM]
Perfect Reconstruction Filters
Perfect Reconstruction and Conjugate Mirror Filter Banks
A perfect reconstruction filter bank decomposes a signal by filtering and subsampling.
It reconstructs it by inserting zeroes, filtering and summation.
Definition
A (discrete) twochannel multirate filter bank convolves a signal a _{0}
with a lowpass filter h _{1} [n] = h[n] and a highpass filter g _{1} [n] = g[
n] and then subsamples the output:
a _{1} [n] = a _{0} * h _{1} [2n] and d _{1} [n] = a _{0} * g _{1} [2n] .
A reconstructed signal a _{2} is obtained by filtering the zero expanded
signals with a dual lowpass filter h _{2} and a dual highpass filter g _{2} .
If z(x) denotes the signal obtained from x by inserting a zero
between every sample, this can be written as:
a _{2} [n] = z(a _{1} ) * h _{2} [n] + z(d _{1} ) * g _{2} [n] .
The following figure illustrates the decomposition and reconstruction process.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/Perfect_Reconstruction.html (1 of 2) [11/15/2003 9:44:37 PM]
Perfect Reconstruction Filters
The filter bank is said to be a perfect reconstruction filter bank when a _{2} = a _{0} . If, additionally, h = h _{2} and g = g _{2} , the filters are called conjugate mirror filters.
Caracterization
Perfect reconstruction filter banks are caracterized in a theorem by
Vetterli. When the filters have a finite impulse response, the g and
g _{2} filters can easily be derived from the h and h _{2} filters, and the filter synthesis is equivalent to solving
where h and h _{2} are trigonometric polynomials.
From filters to wavelets
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/Perfect_Reconstruction.html (2 of 2) [11/15/2003 9:44:37 PM]
Normalized Saclogram examples
Normalized Scalogram examples
These few examples are based on the same synthetic signals as for the windowed Fourier transform.
They illustrate the wavelet transform's ability to localize instaneous frequencies.
Here is the sum of two parallel linear chirps with its scalogram. When the frequency increases, the frequency resolution of the transform decreases.
Here is the normalized scalogram and the complex phase of the wavelet transform of a synthetic signal which is the sum of a linear chirp with an increasing frequency, a quadratic chirp with a decreasing frequency, and two modulated gaussians. Their computations have been performed with a Gabor wavelet.
The makup of this signal explicitly introduces instantaneous frequencies.
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
transformees/Ondelettes/Scalogram_Examples.html
(1 of 3) [11/15/2003 9:44:52 PM]
Normalized Saclogram examples
Here is the sum of two hyperbolic chirps and its scalogram.
In the first example, the instantaneous frequencies are more blurred in the scalogram than in the spectrogram. On the other hand, the variable time resolution of the wavelet transform makes it possible to track the hyperbolic frequency across time. The decreasing frequency resolution is masked by the vertical asymptotic tendency.
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
transformees/Ondelettes/Scalogram_Examples.html
(2 of 3) [11/15/2003 9:44:52 PM]
Normalized Saclogram examples
Wavelet Ridges and Instantaneous Frequencies
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
transformees/Ondelettes/Scalogram_Examples.html
(3 of 3) [11/15/2003 9:44:52 PM]
Frames
Frames
Frames are a stable, possibly redundant, representation of signals.
It is a generalization of the concept of basis in a linear space.
A frame is a family of vectors which can represent any finite
energy signal by the sequence of its inner products with the vectors
of the family. However, it may be possible that not all sequences of
reals represent an finite energy signal. Oversampling is an example
of a represention of signals in a frame. One can see that not all sequences of values may represent a sequence of samples. In general, frames are a stable and redundant representation of signals.
Definition
A family
there are two constants A>0 and B>0 such that, for any f in H,
of vectors in a Hilbert space H is a frame of H if
If A=B, the frame is said to be tight.
A Riesz basis is a frame whose vectors are linearly independant.
Example: consider a family of three vectors in the plane which are obtained by succesive rotations of a third of turn of one vector. This family is a tight frame of the plane, with A=B=3/2.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Frames_IntroUS.html (1 of 3) [11/15/2003 9:44:57 PM]
Frames
Properties
The frame vectors are supposed to be of unit norm.
If the frame vectors are independent, then A<=1<=B. The frame is
then an orthonormal basis if and only if A=B=1. If A>1, then the frame is redundant. A finite family is always a frame of the linear
space that it generates.
Pseudo Inverse
U denotes the operateur which transforms a signal f into the
sequence of its frame inner products.
U has one or an infinity of left inverses.
The pseudoinverse
on the orthogonal complement of the image of U. It is the minimum norm left inverse.
of U is the left inverse of U which is zero
It is used to build a signal approximation from any sequence of real numbers. The computation of (U * U) 1 f can be performed by a conjugate gradient algorithm.
Dual Frame
The image of the frame the dual frame. For any f in H,
through (U * U) 1 is a frame called
and
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Frames_IntroUS.html (2 of 3) [11/15/2003 9:44:57 PM]
Frames
If the original frame is a Riesz basis, then the two frames form a biorthogonal basis system, that is
Windowed Fourier Frames and Wavelet Frames
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Frames_IntroUS.html (3 of 3) [11/15/2003 9:44:57 PM]
Windowed Frames
Windowed Fourier Frames
The Heisenberg boxes of windowed Fourier atoms have dimensions which do not depend on their time or frequency center. To get a frame after discretization of the windowed Fourier transform, a tiling of the time frequency plane by the discrete family of Heisenberg boxes is desirable. Hence, it only logical to use a regular rectangular grid to place the time frequency centers.
Daubechies gives the following necessary condition on the tiling to yield a frame:
Sufficient conditions also exist.
Two important results should be emphasized:
● there is no compactly supported, continuously differentiable window that generates an orthogonal windowed Fourier basis (Balian Low theorem).
Wavelet Frames
To cover the time frequency plane with wavelet Heisenberg boxes, a regular grid is not used; time steps which are inverse proportional to the frequecy step are used instead, the latter being itself proportional to the scale.
The wavelet is assumed to satisfy the reconstruction condition
which garantees the invertibility of the wavelet transform. Daubechies gives necessary conditions for the previous tiling to yield a frame:
Sufficient conditions also exist.
The following differences with the windowed Fourier frames should be emphasized:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Windowed_Frames.html (1 of 3) [11/15/2003 9:45:01 PM]
Windowed Frames
● the dual frame of a windowed Fourier frame is also a window Fourier frame.
● there are continuously differentable wavelets that generate frames (look at the construction of wavelet bases for more)
● in the general case, the dual frame of a wavelet frame is not a wavelet frame. However, in the cases of bases, a dual wavelet basis can be built by other means (look at the wavelet bases for more, especially the biorthogonal ones)
Translation invariance
In both cases (Fourier or wavelets), the frame representation has the drawback of not being translation invariant with respect to time or frequency. Now, most interesting signal patterns are not naturally synchronized with frame intervals. In particular, the structure of a signal may be degraded at the lower resolutions.
This motivates the study of the dyadic wavelet transform, which is discrete in scale but not in time (in practice, this means that signals are oversampled when switching to coarser resolutions).
Another time invariant represenation is the representation by dyadic wavelet maxima. It is less redundant, but is not complete.
Why wavelet bases are studied nonetheless
In practice, the dyadic wavelet transform is implemented by perfect reconstruction filter banks. These fast filter banks correspond to wavelet bases which are built from multiresolution approximations.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Windowed_Frames.html (2 of 3) [11/15/2003 9:45:01 PM]
Windowed Frames
Multiresolution Approximations and Wavelet Bases
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/frames/Windowed_Frames.html (3 of 3) [11/15/2003 9:45:01 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier_Fenetre/windowdesign.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Fourier_Fenetre/windowdesign.gif [11/15/2003 9:45:03 PM]
Windowed Fourier Rigdes
Windowed Fourier Ridges
The windowed Fourier ridges are the maxima points of the spectrogram.
They indicate the instantaneous frequencies within the limits of the transform's resolution.
The latter is determined by the Heisenberg boxes which tile the time frequency plane.
Windows Used
The windows g used here are symmetric with respect to 0 and have a support within [1/2,1/2], as in the previous table.
The windowed Fourier ridges are the maxima points of the spectrogram. If the amplitude and frequency have a small variation within the Fourier window, and if the instantaneous frequency is higher than the window's passing band, then the frequencies which maximize the spectrogram approximate the instantaneous frequencies. At these points, the complex phase of the transform is almost constant.
Time Frequency Resolution
The windowed Fourier ridges of the sum of two analytic signals can discriminate their two instantaneous frequencies if their difference is greater than the scaled window's bandwidth:
http://cas.ensmp.fr/~chaplais/Wavetour_presentat
transformees/Fourier_Fenetre/Fourier_Ridges.html
(1 of 2) [11/15/2003 9:45:09 PM]
Windowed Fourier Rigdes
where s is the scaling which has been applied to the Fourier window, and is the bandwidth of the unscaled window g.
This is a condition on the absolute frequency difference. It is related to the structure of the time frequency tiling.
Hence, the windowed Fourier ridges can detect instantaneous frequencies provided they are not too close.
Examples of Windowed Fourier Ridges
http://cas.ensmp.fr/~chaplais/Wavetour_presentat
transformees/Fourier_Fenetre/Fourier_Ridges.html
(2 of 2) [11/15/2003 9:45:09 PM]
Wavelets
Orthogonal Wavelets
Let us recall that a multiresolution approximation is a nested sequence of linear spaces. The orthogonal complement W _{j} of V _{j} in V _{j}_{}_{1} can be thus defined:
Then there is a function such that the family
n), n in Z, is an orthonormal basis of W _{j} . The family _{j}_{,}_{n} , j in Z
and n in Z, is an orthonormal basis of L 2 and
_{j}_{,}_{n} (t) = 2 j/2 (2 j t
is an orthogonal wavelet associated to the multiresolution approximation. A signal f in L 2 can be decomposed as
where is an orthogonal scaling function of the multiresolution.
Biorthogonal Wavelets
Biorthogonal wavelets are defined similarly to orthogonal wavelets, except that the starting point is biorthogonal multiresolution approximations. The following decompositions are performed:
Like in the orthogonal case, a signal in L 2 can be written as
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Wavelets.html (1 of 4) [11/15/2003 9:45:17 PM]
Wavelets
A theorem by Mallat and Meyer builds an orthogonal wavelet from a scaling function.
Example
Here is a cubic spline scaling function and the corresponding cubic spline BattleLemarié wavelet, and their Fourier transform. The wavelet is a cubic spline because it is a linear combination of cubic splines.
The wavelet is not compactly supported.
From the real axis to the interval
Example
Below is a biorthogonal system which includes a cubic Bspline. Dropping the orthogonality constraint makes possible to have both regularity and symmetry.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Wavelets.html (2 of 4) [11/15/2003 9:45:17 PM]
Wavelets
Wavelet bases are bases of L 2 (R). There are several methods to transform them into wavelet bases over an interval. Once discretized, they are used to process finite signals.
From dimension 1 to dimension 2
There are several methods to build wavelet bases on functional spaces in dimension greater than 1. The simplest ones uses separable wavelets.
Properties
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Wavelets.html (3 of 4) [11/15/2003 9:45:17 PM]
Biorthogonal cubic Bspline scaling function
Biorthogonal spline wavelet
Dual scaling function
^{D}^{u}^{a}^{l} ^{W}^{a}^{v}^{e}^{l}^{e}^{t}
The construction of biorthogonal wavelets over the interval or in dimension 2 will not be presented here. It follows the same lines as in the orthogonal case.
Wavelets
Properties
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Wavelets.html (4 of 4) [11/15/2003 9:45:17 PM]
Dyadic Wavelet Transform
Dyadic Wavelet Transform
Dyadic wavelet transforms are scale samples of wavelet transforms following a geometric
sequence of ratio 2. Time is not sampled.
This transform uses dyadic wavelets.
It is implemented by perfect reconstruction filter
banks.
Definition
The dyadic wavelet transform of f is defined by
It defines a stable complete representation if its Heisenberg boxes cover all of the frequency axis, that is, if there exist A et B such that
The family of dyadic wavelets is a frame of L 2 (R).
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Dyadic_Transform.html (1 of 2) [11/15/2003 9:45:21 PM]
Dyadic Wavelet Transform
Wavelet synthesis
To build dyadic wavelets, it is sufficient to satisfy the previous condition. To do so, it is possible to proceed as for the construction of orthogonal and biorthogonal wavelet bases, using conjugate
mirror or perfect reconstruction filter banks.
The wavelets satisfy then scaling equations and the fast dyadic
wavelet transform is implemented using filter banks.
Implementation
The fast dyadic wavelet transform uses the same filters as for the computation of the fast wavelet transform of a discrete signal, except that no subsampling is performed.
Back to top
Next path
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Dyadic_Transform.html (2 of 2) [11/15/2003 9:45:21 PM]
Reconstruction From Dyadic Maxima
Reconstruction From Dyadic Wavelet Modulus Maxima
A signal is representated by its low pass approximation and the modulus maxima of its dyadic wavelet transform.
This representation allows an almost perfect reconstruction of a signal.
Outline
The continuous wavelet transform detects isolated singularities with their order of singularity. The regular part of the signal is coded in its coarsest approximation. It is sensible to try to reconstruct a signal from this coarse resolution and from its wavelet modulus maxima.
In practice, only the dyadic wavelet transform is considered to take advantage of the fast algorithme à trous which implemented by filter banks.
From a theoretical point of view, Meyer and Berman have proved that the representation by dyadic maxima is not complete because several signals may exhibit the same wavelet maxima.
In practice, numerical experiments have shown that it is possible to reconstruct usual signals with a relative mean sqaure error smaller than 10 2 . On images, the difference is not visible.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Dyadic_Maxima.html (1 of 2) [11/15/2003 9:45:23 PM]
Reconstruction From Dyadic Maxima
Implementation
A signal is to be reconstructed from the values and locations u _{j}_{,}_{p} of
its wavelet modulus maxima, j being the scale and p the time localization. This difficult problem is replaced in practice by a simpler one which consists in finding a minimum norm signal among those which have the assigned wavelet coefficients at the maxima locations. Solving this problem tends to create signal with modulus maxima at the right locations with the correct values.
Since this problem actually bears on discrete signals, this simplified probleme is an inverse frame problem, which can be solved using a conjugate gradient algorithm. To this reconstruction a previously stored low frequency component defined by the sample averages is added.
An example in PDF format (32 Kb) is available. Here is a preview
of it:
Images, Edge Dectection and Reconstruction
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Dyadic_Maxima.html (2 of 2) [11/15/2003 9:45:23 PM]
Detection of singularities
Detection of singularities
Wavelet transform modulus maxima are related to the singularities of the signal.
More precisely, the following theorem proves that there cannot be a singularity without a local maximum of the wavelet transform at the finer scales.
This theorem indicates the presence of a maximum at the finer scales where a singularity occurs. In the general case, is sequence of modulus maxima is detected which converges to the singularity. Below are the modulus maxima of the previous example:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Detection_of_singularities.html (1 of 3) [11/15/2003 9:45:51 PM]
Detection of singularities
Modulus maxima are in yellow (light grey on greyscale monitor). A PDF file shows the connection between wavelet modulus maxima and the signal singularities. Since log _{2} (s) >=0 because of the discretization, the detection on the wavelet transform is restricted to log _{2} (s)>=1 to preserve the continuous case approximation.
Warning: these are the modulus maxima of the wavelet transform. Instantanuous frequencies are detected from the maxima of the normalized scalogram:
which differs in two ways: normalization, and the fact that the
variable is homogeneous to a frequency, and not to a scale.
When the wavelet is the n th derivative of a gaussian, the maxima curves are connected and go through all of the finer scales.
The decay rate of the maxima along the curves indicate the order of the isolated singularities (this a consequence of theorems 6.4 et 6.6 when extended to an interval):
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Detection_of_singularities.html (2 of 3) [11/15/2003 9:45:51 PM]
Detection of singularities
The modulus maxima are displayed as a function of the scale in loglog axes, and the slope gives the estimated singularity order. Below is such a curve for two singularities: the solid line corresponds to the singularity at t=14 and the dotted line to the singularity at t=108. Fine scales are on the left.
For t=14, the slope is 1/2, and the signal is 0Lipschitz here, that is, it has a discontinuity. For t=108, the slope is close to 1, which indicates that the signal is 1/2 Lipschitz here.
Reconstruction From Dyadic Maxima
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Regularite/Detection_of_singularities.html (3 of 3) [11/15/2003 9:45:51 PM]
Biorthogonal Wavelets and Discrete Filters
Biorthogonal Wavelets and Discrete Filters
The scaling equations on the scaling functions and wavelets show that the decomposition and reconstruction of a signal from a resolution to the next one is implemented by perfect reconstruction
filter banks.
The scaling equations imply the coefficients a _{1} [n] and d _{1} [n] of a signal in V _{j} and W _{j} are computed from its coefficients a _{0} [n] in V _{j}_{}_{1} by applying the filters h and g and subsampling the output:
a _{1} [n] = a _{0} * h _{1} [2n] and d _{1} [n] = a _{0} * g _{1} [2n] .
with h _{1} [n] = h[n] and g _{1} [n] = g[n].
In practice this recursion is initialized by considering that the discrete signal samples are some fine resolution coefficients.
The coefficients of h and g are defined by the scaling equations
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Biortho_Wave_and_filt.html (1 of 3) [11/15/2003 9:45:56 PM]
Biorthogonal Wavelets and Discrete Filters
or, in the Fourier domain:
Conversely, a _{0} [n] is reconstructed from a _{1} [n] and d _{1} [n] by inserting zeroes between two consecutive samples and summing their convolutions with the dual filters h _{2} et g _{2} which define the dual scaling equations:
a _{0} [n] = z(a _{1} ) * h _{2} [n] + z(d _{1} ) * g _{2} [n]
where the z operator represents the insertion of zeroes.
The coefficients of h _{2} and g _{2} are defined by the scaling equations
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Biortho_Wave_and_filt.html (2 of 3) [11/15/2003 9:45:56 PM]
Biorthogonal Wavelets and Discrete Filters
or, in the Fourier domain:
This algorithm is used to evaluate the scaling functions and wavelets. Indeed the coefficients of a scaling function (resp. wavelet) are all zero but for one within their on resolution (resp. detail) space basis. The reconstruction algorithm provides the coefficients in the finer resolutions. For high resolutions, the scaling coefficients are considered to be samples of the function.
Hence the construction of biorthogonal wavelets is equivalent to the synthesis
of perfect reconstruction filters having a
stability property.
Filtering
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Biortho_Wave_and_filt.html (3 of 3) [11/15/2003 9:45:56 PM]
Vetterli
g and g _{2} can be eliminated from the previous equations, which leads to the necessary condition:
For filite impulse response filters, the Fourier transforms of such filters are trigonometric polynomials, and conditions (7.121) and (7.129) can be interpreted as Bezout identities in the ring of trigonometric polynomials. In this ring, units are trigonometric monomials. Equations (7.121)  (7.122) form a linear system with respect to h _{2} and g _{2} , and it can be shown that the associated matrix is unimodular, i.e., its determinant is a trigonometric monomial. then there exists a real number a and an integer l such that
Back to perfect reconstruction filters
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/VitterliUS.html [11/15/2003 9:45:59 PM]
From filters to wavelets
From filters to wavelets
Wavelets and scaling functions
Biorthogonal wavelets and scaling functions are caracterized by a perfect reconstruction filter bank; orthogonal wavelets and scaling functions are caracterized by a pair of conjugate mirror filters. Nonetheless, a perfect
reconstruction filter bank (or any pair of conjugate mirror filters) does not necessarily generate a wavelet system. Indeed, some attention has to be paid
to the stability of the decomposition and reconstruction schemes as the
number of scales increases, that is when the number of filter bank cascades goes to the infinity. This is expressed by an additional condition (7.37) on the
conjugate mirror filter h for it to define a scaling function.
Perfect reconstruction filter banks and algorithme à trous
The decomposition can be performed on the signal a _{1} to generate a signal a _{2}
and a signal d _{2} ; repeating this construction produces a low resolution signal
a _{j} and a sequence of detail signals d _{1}
d _{j} .
A recursive decomposition which similar to the previous one can be
performed by the algorithme à trous to generate low resolution signal A _{j} and
a sequence of detail signals D _{1} the following equations:
D _{j} . The two decompositions are related by
a _{j} [n] = A _{j} [2 ^{j} n]
d _{j} [n] = D _{j} [2 ^{j} n]
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
ndelettes%20dyadiques/FromFilters2Wavelets.html
(1 of 2) [11/15/2003 9:46:03 PM]
From filters to wavelets
From algorithme à trous to scaling functions
In the Fourier domain, the transfer between a _{0} and A _{j} is
Let us operate a time rescaling T = 2 ^{}^{j} t so that the interval between the non zero coefficients of the slower filter is always one. Then the interval between the non zero coefficients of the tightest filter is 2 ^{}^{j} . The transfer becomes
Let j go to the infinity. If the previous transfer converges in L ^{2} , then its limit is the Fourier transform of a finite energy signal which necessarily satisfies a scaling equation:
Such functions are at the core of multiresolution analysis, which is itself the sarting point for the definition of dyadic wavelets.
Multiresolution analysis
Filter Synthesis
http://cas.ensmp.fr/~chaplais/Wavetour_presenta
ndelettes%20dyadiques/FromFilters2Wavelets.html
(2 of 2) [11/15/2003 9:46:03 PM]
Wavelet Ridges
Wavelet Ridges
The wavelet ridges are the maxima points of the normalized scalogram.
They indicate the instantaneous frequencies within the limits of the transform's resolution.
The latter is determined by the Heisenberg boxes which tile the time frequency plane.
Wavelets used
Approximatively analytic wavelets are used:
like Gabor wavelets. The atoms are similar to a windowed Fourier transform's, but, after rescaling, the window width is proportional to the "frequency" = /s.
Hence, similar windows are used, but with a different time frequency tiling.
The wavelet ridges are the maxima points of the normalized
scalogram. Under conditions which are similar to the spectrogram's, the frequencies which maximize the normalized scalogram approximate the instantaneous frequencies. The difference is that the time frequency resolution structure is different.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Ridges.html (1 of 2) [11/15/2003 9:46:09 PM]
Wavelet Ridges
Time frequency resolution
The wavelet ridges of the sum of two analytic signals can discriminate their two instantaneous frequencies if their relative differences are greater than the relative wavelet bandwidth:
and
where is the wavelet bandwidth and its frequency center.
These conditions bear on the relative frequency differences. They are related to the structure of the time frequency tiling.
Hence, the wavelet ridges can detect instantaneous frequencies provided their relative distances are not too small.
Examples of Wavelet Ridges
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/transformees/Ondelettes/Wavelet_Ridges.html (2 of 2) [11/15/2003 9:46:09 PM]
Orthogonal Multiresolution Approximations
Orthogonal Multiresolution Approximations
Orthogonalization
The Riesz basis can be transformed into an orthogonal basis generated by integer translations of an elementary function, which is called a scaling function. It is a particular case of a biorthogonal system where both bases are equal.
Scaling function built from cubic spline approximations and its Fourier transform. Observe the time frequency localization. The scaling function is a cubic spline because it is generated by cubic splines
Scaling Equation
One can verify that the other resolutions are generated by a suitable dilatation of these bases of translated atoms. Since the resolutions are embedded, there is necessarily a sequence of real numbers h[n] such that
or, in the Fourier domain
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Orthogonal_Multires.html (1 of 3) [11/15/2003 9:46:16 PM]
Orthogonal Multiresolution Approximations
assuming the Fourier transform is continuous.
It is proved that h is a conjugate mirror filter. It entirely determines the scaling function and most of its properties. In particular, the scaling function is compactly supported if and only if h has a finite number of non zero coefficients. h is sais to be a Finite Impulse Response (FIR) filter. For more info, see the page on the properties of of orthogonal
wavelets and how they are related the the filter h.
As an example, the filter coefficients that correspond to the cubic spline scaling function are given.
The study of the filters h which generate multiresolution approximations has produced many important theorems. The following one gives necessary and sufficient conditions for h to generate a scaling function:
Condition (7.35) means that h is a conjugate mirror filter. Condition (7.36) is simply a normalization. Condition (7.37) garantees that the function defined by (7.38) has a finte energy.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Orthogonal_Multires.html (2 of 3) [11/15/2003 9:46:16 PM]
Orthogonal Multiresolution Approximations
The scaling equations shows that the scaling coefficients a _{1} [n] = < a , (t/2n) > of a in V _{1} are obtained from the scaling coefficients a _{0} [n] = < a , (tn) > in V _{0} by a convolution with the conjugate mirror filter h and a subsampling:
a _{1} [n] = a _{0} * h _{1} [2n]
Back to Multiresolution Approximations
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Orthogonal_Multires.html (3 of 3) [11/15/2003 9:46:16 PM]
Properties of Orthogonal Wavelets
Properties
Dilated wavelets are related by a scaling equation.
Rescaling can be interpreted as discrete filtering.
Vanishing moments, support, regularity and symmetry of the wavelet and scaling function are determined by the scaling filter.
Scaling equation
(t/2) and (t) are related by a scaling equation, similar to
equation which relates (t/2) and (t). It is a consequence of
(7.60):
the
In the Fourier domain, this equation becomes
The h and g filters are a conjugate mirror filter bank.
Vanishing moments
A wavelet has m vanishing moments if and only if its scaling
function can generate polynomials of degree smaller than or equal
to m. While this property is used to describe the approximating
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Orth_Wave.html (1 of 4) [11/15/2003 9:46:21 PM]
Properties of Orthogonal Wavelets
power of scaling functions, in the wavelet case it has a "dual" usage, e.g. the possibility to caracterize the order of isolated singularities.
The number of vanishing moments is entirely determined by the coefficients h[n] of the filter h which is featured in the scaling
equation.
If the Fourier transform of the wavelet is p continuously differentiable, then the three following conditions are equivalent:
● the wavelet has p vanishing moments
● the scaling function can generate polynomials of degree smaller than or equal to
p
● the transfer function of the filter h and its p1
first derivatives vanish at = .
Compact support
Compactly supported wavelets and scaling functions exist.
The scaling function is compactly supported if and only if the filter
h has a finite support, and their supports are the same. If the
support of the scaling function is [N _{1} ,N _{2} ], then the wavelet support
is [(N _{1} N _{2} +1)/2,(N _{2} N _{1} +1)2].
Atoms are thus compactly supported if and only if the filter h is.
Daubechies has proved that, to generate an orthogonal wavelet with
p vanishing moment, a filter h with minimum length 2p had to be
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Orth_Wave.html (2 of 4) [11/15/2003 9:46:21 PM]
Properties of Orthogonal Wavelets
used. Daubechies filters, which generate Daubechies wavelets, have a length of 2p. The Daubechies filter coefficients are available as ASCII text files which can be used in a spreadsheet, for instance.
Regularity
Wavelet regularity is much less important than their vanishing moments. Il is studied in a theorem by Tchamitchian
The following two properties are important:
● there is no compactly supported orthogonal wavelet which indefinitely differentiable
● for Daubechies wavelets with a large p, the scaling function and wavelet are lLipschitz, where l is of the order of 0.2 p. For large classes of orthogonal wavelets, more regularity implies more vanishing moments.
Meyer wavelets are indefinitely differentiable orthogonal wavelets, with an infinite support. They are generally implemented in the Fourier domain.
Symmetry
Symmetric scaling functions and wavelets are important because they are used to build bases of regular wavelets over an interval, rather than the real axis. Daubechies has proved that, for a wavelet to be symmetric or antisymmetric, its filter must have a linear complex phase, and the only symmetric compactly supported conjugate mirror filter is the Haar filter, which corresponds to a discontinuous wavelet with one vanishing moment. Besides the Haar wavelet, there is no symmetric compactly supported orthogonal wavelet.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Orth_Wave.html (3 of 4) [11/15/2003 9:46:21 PM]
Properties of Orthogonal Wavelets
Orthogonal Wavelets and Discrete Filters
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Orth_Wave.html (4 of 4) [11/15/2003 9:46:21 PM]
Coefficients du filtre spline cubique
h[n]
0,766130398
0,433923147
0,050201753
0,110036987
0,032080869
0,042068328
0,017176331
0,017982291
0,008685294
0,008201477
0,004353840
0,003882426
0,002186714
0,001882120
0,001103748
0,000927187
0,000559952
0,000462093
0,000285414
0,000232304
0,000146098
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Coefficients_du_filtre_spl.html [11/15/2003 9:46:23 PM]
Properties of Multiresolution Approximations
Properties
Orhogonality and biorthogonality
When the Riesz basis is an orthogonal basis, the multiresolution approximation is orthogonal, and the base atom is called a scaling function. It is always possible to orthogonalize a multiresolution approximation.
However, orthogonalities imposes some constraints that may not be desirable. One of the most important is that a compactly supported (orthogonal) scaling function cannot symmetric and continuous. The symmetry is useful in the analysis of finite signals.
Some of these restrictions (notably the absence of symmetry) can be avoided by using biorthogonal multiresolution approximations.
Approximation
The "approximation" denomination means that an orthogonal or biorthogonal multiresolution analysis (or approximation) can be related to a sequence of respectively orthogonal or oblique projectors, which efficiently approximate regular signals. The order of approximation is determined by the degree of the polynomials that can be reconstructed in the resolutions.
Construction and digital filters
Multiresolution approximations are determined by one or two atoms which generate respectively the orthogonal or biorthogonal multiresolutions.
By definition of a multiresolution, (t/2) is a linear combination of the (tn). This relation is called a scaling equation:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Properties.html (1 of 3) [11/15/2003 9:46:27 PM]
Properties of Multiresolution Approximations
In the Fourier domain, this becomes
The coefficients h[n] in the scaling equation entirely determine , and finding them is equivalent to the design of a filter bank, plus
some stability conditions to be able to generate L 2 .
Wavelets
As the scale j gets finer the approximations becomes more accurate (see Lena). Switching from the resolution j to j1 adds details to the signal. The same way that approximations can be decomposed on resolution bases, these extra details can be decomposed in detail bases.
Details bases, like resolution bases, ore obtained by translating a single resealed atom. This atom is called a wavelet. The order of approximation of the multiresolution is equal to the number of vanishing moments of the wavelet. It also represents the wavelet's ability to detect the isolated singularities of a signal.
Wavelet Bases
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Properties.html (2 of 3) [11/15/2003 9:46:27 PM]
Properties of Multiresolution Approximations
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Multiresolution_Properties.html (3 of 3) [11/15/2003 9:46:27 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/MallatMeyer2US.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/MallatMeyer2US.gif [11/15/2003 9:46:32 PM]
Wavelets Over an Interval
Wavelets Over an Interval
Until now, only wavelets over the real axis have been considered, e.g. wavelets that are suited to the analysis of signals defined over the whole axis. In most cases, signals are compactly supported; images, in particular, are explicitely defined over a rectangle of pixels.
The wavelets considered here are compactly supported.
A [0,N] supported signal can be represented as the product of a
general signal with the caracteristic function of [0,N]. The discontinuities of this function require special attention. Three
methods are known to handle them, the last one being the most efficient.
Wavelet periodization
The wavelets are periodized by the following transformation:
with j<=log _{2} N. This is equivalent to a signal periodization.
Wavelets which are completely inside the interval are not changed. Wavelets that ovelap the boudaries are cut into two pieces loacated
at the left and right edges of the interval. Separately, each of the
pieces have no vanishing moment. This creates large wavelet coefficients when the periodized signal is not itself continuous.
Wavelet folding
To bypass this problem, the signal is symmetrically folded around
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Interval_Wavelets.html (1 of 3) [11/15/2003 9:46:40 PM]
Wavelets Over an Interval
the right edge of the interval and periodized over the double sized interval.
This yields a continuous periodic signal.
Porting the signal transformation to the wavelet basis shows that the vector family is a basis of L ^{2} [0,N] if the wavelet is symmetric or antisymmetric. This puts orthogonal bases asides.
In fact, the continuity problem reappears at the next derivative. The following approach takes the problem at the root, which is how to make wavelets over an interval with vanishing moments.
Edge wavelets
Boundary effects are explicitely handled. Consider an Daubechies orthogonal basis with p vanishing moments.
From the Strang et Fix conditions, it appears that there exists a polynomial _{k} of degree k such that:
for k<p.
This equation is multiplied by the caracteristic function of [0,N]. Assuming that the support of is [p+1,p], scaling functions with indices p<=k<Np are not changed by this restriction. To recover the Strang and Fix condition on the interval, p "left" edge scaling function and p "right" edge scaling functions are to be found such
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Interval_Wavelets.html (2 of 3) [11/15/2003 9:46:40 PM]
Wavelets Over an Interval
that
If this equation is satisfied, it reamains valid after rescaling since the n ^{k} , up to a power of 2, are the scaling coefficients of _{k} at all resolutions. There remains to find the filters h and H which satisfy the scaling equation:
where
translation at the resolution j, and to verify the orthogonality condition.
denotes the whole set of scaling functions obtained by
The coefficients of these filters are available in Wavelab, using the function MakeCDJVFilter.
Back to Wavelet Bases
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Interval_Wavelets.html (3 of 3) [11/15/2003 9:46:40 PM]
Strang and Fix conditions
What approximations?
An orthogonal multiresolution approximation defines an orthogonal projector on each of the resolution spaces. In the biorthogonal case, the decomposition
defines a (non necessarily orthogonal) projector on V _{0} , and, after rescaling, a projector on each resolution V _{j} . The projection of a signal f is:
This projection is an approximation of f under the following conditions:
The sufficient condition can be interpreted as follows: the projection on V _{j} is able to "catch" Taylor expansions of f up to
degree p at intervals of length 2 j .
The general Stang and Fix conditions are available with proof (PDF v.3, 107 K).
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Strang_and_Fix.html (1 of 2) [11/15/2003 9:46:53 PM]
Strang and Fix conditions
Remark
The definition of multiresolution approximations specifies
Is this compatible with the previous theorem? (Answer)
Back to multiresolution approximations
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/Strang_and_Fix.html (2 of 2) [11/15/2003 9:46:53 PM]
MakeCDJVFilter
MakeCDJVFilter  Set up filters for CDJV Wavelet Transform
Usage
[a,b,c] = MakeCDJVFilter(request,degree)
Inputs
request
string: 'HighPass', 'LowPass', 'Precondition', 'Postcondition'
degree
integer: 2 or 3 (number of vanishing moments)
Outputs
a,b,c filter, left edge filter, right edge filter
('HighPass', 'LowPass')
a conditioning matrix ('Precondition', 'Postcondition')
Description
CDJV have developed an algorithm for wavelets on the interval which preserves the orthogonality, vanishing moments, smoothness, and compact support of Daubechies wavelets on the line.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/MakeCDJVFilter.html (1 of 2) [11/15/2003 9:46:57 PM]
MakeCDJVFilter
The algorithm for wavelets on the interval of CDJV involves four objects not present in the usual periodized algorithm: right edge filters, left edge filters, and pre and post conditioning operators.
These objects are supplied by appropriate requests to MakeCDJVFilter.
See Also
IWT_CDJV, FWT_CDJV, CDJVDyadDown
References
Cohen, Daubechies, Jawerth and Vial, 1992.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/MakeCDJVFilter.html (2 of 2) [11/15/2003 9:46:57 PM]
Two Dimensional Wavelets
Two Dimensional Wavelets
The simplest way to build two dimensional wavelet bases is to use separable products on a one dimensional wavelet and scaling function. This yields the following scaling function
and there are three wavelets:
Back to Wavelets Bases
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Wavelets_in_dim_2.html [11/15/2003 9:47:00 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/ScalingEq.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Multiresolution/ScalingEq.gif [11/15/2003 9:47:01 PM]
Index of /~chaplais/Wavetour_presentation/Tables/Daubechies
Index of /~chaplais/Wavetour_presentation/Tables/Daubechies
Name
Last modified
Size Description
Parent Directory

Daubechies_coefficients 04Jun1999 12:56 2.6K
Daubechies_coefficie 
> 
04Jun1999 12:56 2.7K 

Daubechies_coefficie 
> 
04Jun1999 12:56 2.6K 
Apache/2.0.43 Server at cas.ensmp.fr Port 80
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Tables/Daubechies/ [11/15/2003 9:47:02 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/TchamitchianUS.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/TchamitchianUS.gif [11/15/2003 9:47:04 PM]
Orthogonal Wavelets and Discrete Filters
Orthogonal Wavelets and Discrete Filters
g and h are conjugate mirror filters.
Conjugate mirror filters are a particular instance of perfect
reconstruction filter banks. The dyadic nature of multiresolution approximations are closely related to the possibility of implementing elementary signal subsampling by erasing one sample every two, and elementary oversampling by zero insertion between two consecutive samples.
The coefficients _{1} [n] and d _{1} [n] of a signal in V _{j} and W _{j} are computed from its coefficients a _{0} [n] in V _{j}_{}_{1} by applying conjugate mirror filters and subsampling the output:
a _{1} [n] = a _{0} * h _{1} [2n] and d _{1} [n] = a _{0} * g _{1} [2n] .
with h _{1} [n] = h[n] and g _{1} [n] = g[n].
In practice this recursion is initialized by considering that the discrete signal samples are some fine resolution coefficients.
The coefficients of h are defined by the scaling equation
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Orth_Wavelets_and_Filters.html (1 of 3) [11/15/2003 9:47:08 PM]
Orthogonal Wavelets and Discrete Filters
or, in the Fourier domain:
and the coefficients of g are defined by the wavelet scaling equation
or, in the Fourier domain:
Conversely, a _{0} [n] is reconstructed from a _{1} [n] and d _{1} [n] by inserting zeroes between two consecutive samples and summing their convolutions with h and g:
a _{0} [n] = z(a _{1} ) * h [n] + z(d _{1} ) * g [n]
where the z operator represents the insertion of zeroes.
Wavelets and scaling functions are evaluated as in the orthogonal
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Orth_Wavelets_and_Filters.html (2 of 3) [11/15/2003 9:47:08 PM]
Orthogonal Wavelets and Discrete Filters
case.
Hence the construction of orthogonal wavelets is equivalent to the synthesis
of conjugate mirror filters having a
stability property.
We will concentrate on Finite Impulse Response filters, e.g., to compactly supported wavelets.
Filtering
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Orth_Wavelets_and_Filters.html (3 of 3) [11/15/2003 9:47:08 PM]
Properties of Biorthogonal Wavelets
Properties
Properties of biorthogonal wavelets are to be compared to the orthogonal case.
Scaling equation
As in the orthogonal case, (t) and (t/2) are related by a scaling equation which is a consequence of the inclusions of the resolution spaces from coarse to fine:
Similar equations exist for the dual functions which determine the filters h _{2} and g _{2} .
Vanishing moments
A biorthogonal wavelet has m vanishing moments if and only if its dual scaling function generates polynomials up to degree m. This can be verified by looking at the biorthogonal decomposition
formulas.
Hence there is an equivalence theorem between vanishing moments and the number of zeroes of the filter's transfer, provided that duality has to be taken into account. Thus the following three properties are equivalent:
● the wavelet has p vanishing moments
● the dual scaling function _{2} generates
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Biorth_Wave.html (1 of 4) [11/15/2003 9:47:11 PM]
Properties of Biorthogonal Wavelets
polynomials up to degree p ● the transfer function of the dual filter h _{2} and it p1 first derivatives vanish at =
and the dual result is also valid. Duality appears naturally, because the filters determine the degree of the polynomials which can be generated by the scaling function, and this degree is equal to the number of vanishing moments of the dual wavelet.
Compact support
If the filters h et h _{2} have a finite support, then the scaling functions
have the same support, and the wavelets are compactly supported.
If the supports of the scaling functions are respectively [N _{1} ,N _{2} ]
and [M _{1} ,M _{2} ], then the corresponding wavelets have support [(N _{1} 
M _{2} +1)/2,(N _{2} M _{1} +1)/2] and [(M _{1} N _{2} +1)/2,(M _{2} N _{1} +1)]
respectively.
The atoms are thus compactly supported if and only if the filters h et h _{2} are.
Regularity
Tchamitchian's theorem provides again a sufficient regularity condition. Remember that this condition bears on the filter h which determines the scaling equation. Hence the regularity of the primal atoms are related to the primal filters.
Wavelet balancing
Consider the following decomposition of f:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Biorth_Wave.html (2 of 4) [11/15/2003 9:47:11 PM]
Properties of Biorthogonal Wavelets
The number of vanishing moments of a wavelet is determined by its dual filter. It corresponds to the approximating power of the dual multiresolution sequence. This is why it is preferred to synthesize a decomposition filter h with many vanishing moments, and possibly with a small support.
On the other hand, this same filter h determines the regularity of , and hence of . This regularity increases with the number of vanishing moments, that is, with the number of zeroes of h.
Symmetry
Unlike the orthogonal case, it is possible to synthesize biorthogonal wavelets and scaling functions which are symmetric or antisymmetric and compactly supported. This makes it possible to use the folding technique to build wavelets on an interval.
If the filters h and h _{2} have and odd length and are symmetric with respect to 0, then the scaling functions have an even length and are symmetric, and the wavelets are also symmetric. If the filters have an even length and are symmetric with respect to n=1/2, then the scaling functions are symmetric with respect to n=1/2, while the wavelets are antisymmetric.
Example
Spline wavelets and scaling functions are an interesting example of biorthogonal systems. One of the scaling functions is a Bspline. A coefficient table is available. There is a general closed form
formula for these filters.
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Biorth_Wave.html (3 of 4) [11/15/2003 9:47:11 PM]
Properties of Biorthogonal Wavelets
Biorthogonal Wavelets and Discrete Filters
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Properties_of_Biorth_Wave.html (4 of 4) [11/15/2003 9:47:11 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/decompjBiortho.gif
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/decompjBiortho.gif [11/15/2003 9:47:14 PM]
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Tables/Splines_biortho/splines
n 
p ,\tilde p 
h[n] {\tilde h}[n] 

0 
p 
= 2 
, \tilde p =4 0,70710678119 
0,99436891104 

1,1 
0,35355339059 0,41984465133 

2,2 
0,1767766953 

3,3 
0,06629126074 

4,4 
0,03314563037 

0,1 
p= 
3 
, 
\tilde p=7 
0,53033008589 0,9516421219 

1,2 
0,1767766953 
0,02649924095 

2,3 
0,30115912592 

3,4 
0,03133297871 

4,5 
0,07466398507 

5,6 
0,01683176542 

6,7 
0,0090632583 

7,8 
0,0030210861 
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/Tables/Splines_biortho/splines [11/15/2003 9:47:17 PM]
Synthesis of Biorthogonal Wavelets
Synthesis of Compactly Supported Biorthogonal Wavelets
Synthesis of perfect reconstruction filter banks
The construction of perfect reconstruction filter banks is simpler than the construction of conjugate mirror filters because the quadrature condition is replaced by a Bezout identity:
In particular, spectral factorization is no longer required.
Biorthogonal wavelet synthesis
A theorem by Cohen, Daubechies and Fauveau gives sufficient
conditions for building biorthogonal wavelets.
One quite interesting example is given by biorthogonal spline wavelets. It is iteresting because it has symmetric scaling functions, and because there existe a closed form formula for the filters.
The spline example
The h filter is taken to be
with =0 if p is even and =1 if p is odd. The scaling function is a
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/Synthesis_of_Biortho.html (1 of 3) [11/15/2003 9:47:21 PM]
Synthesis of Biorthogonal Wavelets
Bspline of degree p1 (this can verified by using the recursion which relates Bsplines of different degrees). It is a symmetric function with respect to 0 if p is odd, and symmetric with respect to 1/2 if p is odd. The corresponding wavelet is respectively symmetric or antisymmetric. The dual wavelet has p vanishing moments.
The only constraint on the number of vanishing moments of the primal wavelet is that it should have the same parity as p. Hence the symmetries are the same as in the previous case. For q=(p+p _{2} )/2, the biorthogonal filter h _{2} of minimum length is given by
Here is an example for p=3 and p _{2} =7
Biorthogonal cubic Bspline scaling function
^{D}^{u}^{a}^{l} ^{s}^{c}^{a}^{l}^{i}^{n}^{g} ^{f}^{u}^{n}^{c}^{t}^{i}^{o}^{n}
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/Synthesis_of_Biortho.html (2 of 3) [11/15/2003 9:47:21 PM]
Synthesis of Biorthogonal Wavelets
Biorthogonal spline wavelet
^{D}^{u}^{a}^{l} ^{W}^{a}^{v}^{e}^{l}^{e}^{t}
The same filters are used to implement the dyadic wavelet transform
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/filtres/Synthesis_of_Biortho.html (3 of 3) [11/15/2003 9:47:21 PM]
Construction of Wavelet Bases
Properties
Dilated wavelets are related by a scaling equation.
Rescaling can be interpreted as discrete filtering.
Vanishing moments, support, regularity and symmetry of the wavelet and scaling function are determined by the scaling filter.
Scaling equation
(t/2) and (t) are related by a scaling equation, similar to the
equation which relates (t/2) and
(t). It is a consequence of (7.60):
In the Fourier domain, this equation becomes
Properties
Properties of biorthogonal wavelets are to be compared to the orthogonal case.
Scaling equation
As in the orthogonal case, (t) and (t/2) are related by a scaling equation which is a consequence of the inclusions of the resolution spaces from coarse to fine:
Similar equations exist for the dual
functions which determine the filters
h _{2} and g _{2} .
Vanishing moments
A biorthogonal wavelet has m vanishing moments if and only if its dual scaling function generates polynomials up to degree m. This can be verified by looking at the biorthogonal decomposition
formulas.
Hence there is an equivalence
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Construction_of_wave_bas.html (1 of 5) [11/15/2003 9:47:32 PM]
Construction of Wavelet Bases
The h and g filters are a conjugate 

mirror filter bank. 

Vanishing moments 

A 
wavelet has m vanishing moments 
if 
and only if its scaling function can 
generate polynomials of degree smaller than or equal to m. While this property is used to describe the approximating power of scaling functions, in the wavelet case it has a "dual" usage, e.g. the possibility to caracterize the order of isolated singularities. 

The number of vanishing moments is entirely determined by the 

coefficients h[n] of the filter h which 

is 
featured in the scaling equation. 
If 
the Fourier transform of the 
wavelet is p continuously differentiable, then the three following conditions are equivalent: 

● the wavelet has p vanishing moments 

● the scaling function can generate 

polynomials of degree smaller than or equal to 

p 
theorem between vanishing moments and the number of zeroes of the filter's transfer, provided that duality has to be taken into account. Thus the following three properties are equivalent:
● the wavelet has p vanishing moments
● the dual scaling function _{2} generates polynomials up to degree p
● the transfer function of the dual filter h _{2} and it p1 first derivatives vanish at =
and the dual result is also valid. Duality appears naturally, because the filters determine the degree of the polynomials which can be generated by the scaling function, and this degree is equal to the number of vanishing moments of the dual wavelet.
Compact support
If the filters h et h _{2} have a finite support, then the scaling functions have the same support, and the
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Construction_of_wave_bas.html (2 of 5) [11/15/2003 9:47:32 PM]
Construction of Wavelet Bases
● the transfer function of the filter h and its p1 first derivatives vanish at = .
Compact support
Compactly supported wavelets and scaling functions exist.
The scaling function is compactly supported if and only if the filter h has a finite support, and their supports are the same. If the support of the scaling function is [N _{1} ,N _{2} ], then the wavelet support is [(N _{1} 
N _{2} +1)/2,(N _{2} N _{1} +1)2].
Atoms are thus compactly supported if and only if the filter h is.
Daubechies has proved that, to
generate an orthogonal wavelet with
p vanishing moment, a filter h with
minimum length 2p had to be used. Daubechies filters, which generate Daubechies wavelets, have a length of 2p. The Daubechies filter coefficients are available as ASCII
text files which can be used in a spreadsheet, for instance.
wavelets are compactly supported. If the supports of the scaling functions are respectively [N _{1} ,N _{2} ] and
[M _{1} ,M _{2} ], then the corresponding
wavelets have support [(N _{1} 
M _{2} +1)/2,(N _{2} M _{1} +1)/2] and [(M _{1} 
N _{2} +1)/2,(M _{2} N _{1} +1)] respectively.
The atoms are thus compactly supported if and only if the filters h et h _{2} are.
Regularity
Tchamitchian's theorem provides again a sufficient regularity condition. Remember that this condition bears on the filter h which determines the scaling equation. Hence the regularity of the primal atoms are related to the primal filters.
Wavelet balancing
Consider the following decomposition of f:
The number of vanishing moments of a wavelet is determined by its dual filter. It corresponds to the approximating power of the dual
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Construction_of_wave_bas.html (3 of 5) [11/15/2003 9:47:32 PM]
Construction of Wavelet Bases
Regularity
Wavelet regularity is much less important than their vanishing moments. Il is studied in a theorem
by Tchamitchian
The following two properties are important:
● there is no compactly supported orthogonal wavelet which indefinitely differentiable
● for Daubechies wavelets with a large p, the scaling function and wavelet are lLipschitz, where l is of the order of 0.2 p. For large classes of orthogonal wavelets, more regularity implies more vanishing moments.
Meyer wavelets are indefinitely differentiable orthogonal wavelets, with an infinite support. They are generally implemented in the Fourier domain.
Symmetry
Symmetric scaling functions and wavelets are important because they are used to build bases of regular wavelets over an interval, rather than the real axis. Daubechies has proved that, for a wavelet to be symmetric or antisymmetric, its filter must have a
multiresolution sequence. This is why it is preferred to synthesize a decomposition filter h with many vanishing moments, and possibly with a small support.
On the other hand, this same filter h determines the regularity of , and hence of . This regularity increases with the number of vanishing moments, that is, with the number of zeroes of h.
Symmetry
Unlike the orthogonal case, it is possible to synthesize biorthogonal wavelets and scaling functions which are symmetric or antisymmetric and compactly supported. This makes it possible to use the folding technique
to build wavelets on an interval.
If the filters h and h _{2} have and odd length and are symmetric with respect to 0, then the scaling functions have an even length and are symmetric, and the wavelets are also symmetric. If the filters have an even length and are symmetric with respect to n=1/2, then the scaling functions are symmetric with respect to n=1/2, while the wavelets are antisymmetric.
Example
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Construction_of_wave_bas.html (4 of 5) [11/15/2003 9:47:32 PM]
Construction of Wavelet Bases
linear complex phase, and the only symmetric compactly supported conjugate mirror filter is the Haar filter, which corresponds to a discontinuous wavelet with one vanishing moment. Besides the Haar wavelet, there is no symmetric compactly supported orthogonal wavelet.
Orthogonal Wavelets and Discrete
Filters
Spline wavelets and scaling functions are an interesting example of biorthogonal systems. One of the scaling functions is a Bspline. A coefficient table is available. There is a general closed form formula for these filters.
Biorthogonal Wavelets and Discrete
Filters
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes/Construction_of_wave_bas.html (5 of 5) [11/15/2003 9:47:32 PM]
250
250
250
200
200
200
150
150
150
100
100
100
50
50
50
0 0 0
Edge Detection
Multiscale Edge Detection and Reconstruction
As in the one dimensional case, dyadic modulus maxima are used to dectect edges.
Provided that the two dimensional geometry is taken into account, these edges can be interpreted as contours.
A similar algorithm to the one dimensional case reconstructs a good approximation of an image from its edges.
Multiscale edges
In images, what is most often perceived as an edge is a curve across which there is a sharp variation of brightness. To make things simpler, the image will be assumed to be monochrome. While the actual concept of an edge is more involved and depends in particular on a priori knowledge about the featured objects, this presentation has the advantage of leading to a precise mathematical definition of an "edge point".
To do so, consider a two dimensioanl wavelet defined by partial differentiation of a kernel:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Edge_Detection.html (1 of 4) [11/15/2003 9:47:52 PM]
Edge Detection
The dyadic wavelet transform is defined by
with, for k=1,2,
The two coordinates of the dyadic wavelet transform are that of the gradient of the convolution of the signal with the dilated kernel:
The multiscale edge points are the points where the dyadic transform modulus is locally maximum along this direction. This corresponds to a locally sharpest variation of image intensity orthogonally to the lines of constant brightness.
Examples
A synthetic example analyzes the edges of a circle.
Another example analyses a classical wavelet picture.
Remark
It is rare that an image line has no hole in it. The brain compensate these defaults using more elaborate image analysis. Notice that the use of color is useful. As illustration, here is an optical illusion where joining edges is far from being obvious:
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Edge_Detection.html (2 of 4) [11/15/2003 9:47:52 PM]
Edge Detection
Reconstruction
As in the one dimensional case, the frame inverse operator can be used to reconstruct a minimum norm image with prescribed values at the maxima locations. Mean square relative errors of l0 2 can be obtained.
On an example, one can see that the reconstruction error is visually neglectible.
Implementation
The computations are performed with separable wavelets whose Fourier transforms are
where g is a finite difference filter; the two wavelets then approximate the partial derivatives of
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Edge_Detection.html (3 of 4) [11/15/2003 9:47:52 PM]
Edge Detection
where is a scaling function defined by a finite impulse response filter h. The dyadic wavelet transform is computed by two dimensional extension of the algorithme à trous.
Back to top
Next Path
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Edge_Detection.html (4 of 4) [11/15/2003 9:47:52 PM]
Cercle
The original image is on top.
Wavelet transform 
Wavelet 

Horizontal 
Vertical 
Wavelet 
angle for a 
transform 
wavelet 
wavelet 
transform 
non zero 
modulus 
transform 
transform 
modulus 
modulus 
maxima 
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Circle.html [11/15/2003 9:48:20 PM]
Lena Edges
A high resolution version of the example is available in PDF format (482K).
Original image:
Wavelet 

Wavelet 
transform 

transform 
Wavelet 
modulus 

Horizontal 
Vertical 
Wavelet 
angle for a 
transform 
maxima above 
wavelet 
wavelet 
transform 
non zero 
modulus 
a given 
transform 
transform 
modulus 
modulus 
maxima 
threshold 
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Lena_Edges.html [11/15/2003 9:50:28 PM]
Reconstruction of Lena
_{O}_{r}_{i}_{g}_{i}_{n}_{a}_{l} _{L}_{e}_{n}_{a} _{i}_{m}_{a}_{g}_{e}
Reconstruction from thresholded modulus maxima and coarse approximation
Reconstruction from modulus maxima and coarse approximation
http://cas.ensmp.fr/~chaplais/Wavetour_presentation/ondelettes%20dyadiques/Reconstruction_of_Lena.html [11/15/2003 9:51:49 PM]
Wavelets and Filters
Orthogonal Wavelets and Discrete Filters
g and h are conjugate mirror filters
Much more than documents.
Discover everything Scribd has to offer, including books and audiobooks from major publishers.
Cancel anytime.