Wavelet Transform

A very brief look

Wavelets vs. Fourier Transform
 In Fourier transform (FT) we represent a
signal in terms of sinusoids
 FT provides a signal which is localized
only in the frequency domain
 It does not give any information of the
signal in the time domain

2
Wavelets vs. Fourier Transform
 Basis functions of the wavelet transform
(WT) are small waves located in different
times
 They are obtained using scaling and
translation of a scaling function and
wavelet function
 Therefore, the WT is localized in both time
and frequency

3
Wavelets vs. Fourier Transform
 In addition, the WT provides a
multiresolution system
 Multiresolution is useful in several
applications
 For instance, image communications and
image data base are such applications

4
Wavelets vs. Fourier Transform
 If a signal has a discontinuity, FT produces
many coefficients with large magnitude
(significant coefficients)
 But WT generates a few significant
coefficients around the discontinuity
 Nonlinear approximation is a method to
benchmark the approximation power of a
transform

5
Wavelets vs. Fourier Transform
 In nonlinear approximation we keep only a few
significant coefficients of a signal and set the
rest to zero
 Then we reconstruct the signal using the
significant coefficients
 WT produces a few significant coefficients for
the signals with discontinuities
 Thus, we obtain better results for WT nonlinear
approximation when compared with the FT

6
Wavelets vs. Fourier Transform
 Most natural signals are smooth with a few
discontinuities (are piece-wise smooth)
 Speech and natural images are such signals
 Hence, WT has better capability for representing
these signal when compared with the FT
 Good nonlinear approximation results in
efficiency in several applications such as
compression and denoising

7
Series Expansion of Discrete-Time Signals

 Suppose that x[n] is a square-summable sequence, that

is x[n]  2 (Z)
 Orthonormal expansion of x[n] is of the form
x[n]   k [l ], x[l ] k [n]   X [k ]k [n] x
2
 X
2

k Z k Z

 Where X [k ]  k [l ], x[l ]  k*[n] x[l ]

l
is the transform of x[n]
 The basis functions  k satisfy the orthonormality
constraint
k [n], l [n]   [k  l ]

8
Haar Basis
 Haar expansion is a two-point avarage
and difference operation
 The basis functions are given as
 1 2 , n  2k
 
2k [n]  1 2 , n  2k , 2k  1 2 k 1[n]  1 2 , n  2k  1
 0, otherwise  0, otherwise

 It follows that
2k [n]  0[n  2k ], 2k 1[n]  1[n  2k ]

9
Haar Basis
 The transform is
1
X [2k ]  2 k ,x   x[2k ]  x[2k  1] ,
2
1
X [2k  1]  2 k 1 ,x   x[2k ]  x[2k  1]
2
 The reconstruction is obtained from
x[n]   X [k ]k [n]
k Z

10
Two-Channel Filter Banks
 Filter bank is the building block of discrete-
time wavelet transform
 For 1-D signals, two-channel filter bank is
depicted below

11
Two-Channel Filter Banks
 For perfect reconstruction filter banks we have
x̂  x
 In order to achieve perfect reconstruction the
filters should satisfy
 g0 [n]  h0 [n]
 g [n]  h [n]
 1 1

 Thus if one filter is lowpass, the other one will be

highpass

12
Two-Channel Filter Banks

13
Two-Channel Filter Banks
 To have orthogonal wavelets, the filter bank
should be orthogonal
 The orthogonal condition for 1-D two-channel
filter banks is
g1[n]  (1)n g0 [n  1]

 Given one of the filters of the orthogonal filter

bank, we can obtain the rest of the filters

14
Haar Filter Bank
 The simplest orthogonal filter bank is Haar
 The lowpass filter is  1
 , n  0, 1
h0 [n]   2

 0, otherwise

 1
 And the highpass filter  2, n0

 1
h1[n]   , n  1
 2
 0, otherwise



15
Haar Filter Bank
 The lowpass output is
1 1
y0 [k ]  h0 [n]* x[n] n  2 k   h0 [l ]x[2k  l ]  x[2k ]  x[2k  1]
l 2 2

 And the highpass output is

1 1
y1[k ]  h1[n]* x[n] n  2 k   h1[l ]x[2k  l ]  x[2k ]  x[2k  1]
l 2 2

16
Haar Filter Bank
 Since y0[k ]  X [2k ] and y1[k ]  X [2k  1] , the filter
bank implements Haar expansion
 Note that the analysis filters are time-reversed
versions of the basis functions

h0[n]  0[n] h1[n]  1[n]

since convolution is an inner product followed by

time-reversal

17
Discrete Wavelet Transform
 We can construct discrete WT via iterated (octave-band) filter banks
 The analysis section is illustrated below

Level 1

Level 2

Level J

18
Discrete Wavelet Transform
 And the synthesis section is illustrated here
 If hi [n] is an orthogonal filter and gi [n]  hi [n] , then we have an
orthogonal wavelet transform
W1
W2 V0
V1
V2

WJ

VJ

19
Multiresolution
 We say that V0 is the space of all square-
summable sequences if V0  2 ( )
 Then a multiresolution analysis consists of
a sequence of embedded closed spaces
VJ   V2  V1  V0  2( )

 It is obvious that J
V j  V0  2( )
j 0

20
Multiresolution
 The orthogonal component of V j 1 in V j will
be denoted by W j 1 :

V j  V j 1  W j 1 V j 1  W j 1

 If we split V0 and repeat on V1 , V2 , ….,VJ 1,

we have
V0  W1 W1  WJ VJ

21
2-D Separable WT
 For images we use separable WT
 First we apply a 1-D filter bank to the rows of the
image
 Then we apply same transform to the columns of
each channel of the result
 Therefore, we obtain 3 highpass channels
corresponding to vertical, horizontal, and
diagonal, and one approximation image
 We can iterate the above procedure on the
lowpass channel

22
2-D Analysis Filter Bank

x h1 h1 diagonal

h0 vertical

h0 h1 horizontal

h0 approximation

23
2-D Synthesis Filter Bank

diagonal g1 g1 x̂

vertical g0

horizontal g1 g0

approximation g0

24
2-D WT Example

Boats image WT in 3 levels

25
WT-Application in Denoising

Boats image Noisy image (additive Gaussian noise)

26
WT-Application in Denoising

Boats image Denoised image using hard thresholding

27
Reference
 Martin Vetterli and Jelena Kovacevic, Wavelets and
Subband Coding. Prentice Hall, 1995.

28