Documente Academic
Documente Profesional
Documente Cultură
and Examples
L. J. Wang
1
Contents
Introduction
A Short Review of Wavelet Analysis
A Simple Example :
Haar Wavelets
Subband filtering scheme
Conclusions and Further Research
2
I. Introduction
The wavelet transform of a signal f (t ) is the
function of scale (or frequency) and time.
Thus, wavelets provide a tool for time-
frequency localization.
Time-frequency localization
In many applications, given a signal f (t ) , one
is interested in its frequency content locally in
time. This similar to music notation, for
example, each note specified a frequency and a
position in time.
3
The Fourier transform of f (t ) ,
1
j t
F ( ) f ( t ) e dt ,
2
4
The windowed Fourier transform provides a
description of f (t ) in the time-frequency plane.
The wavelet transform of f ( t ) is defined by
1/ 2 t b
F ( a , b) a f ( t ) (
WAV
)dt ,
a
1/ 2 t b
Let (t ) a (
a, b
),
a
where (t ) are called wavelets and
a, b
5
Compression techniques are divided into two
main techniques : transforms (DCT, JPEG, FFT,
Wavelet) and nontransforms (PCM, DPCM).
Compression can be achieved by transforming
the data, projecting it on a basis of functions,
and then encoding the resulted coefficients.
The wavelet transform cuts up the image into a
set of subimages with different resolutions
corresponding to different frequency bands.
One encoding approach is based on quantizing
the coefficients using vector quantization.
6
Because of the nature of the image signal and
the mechanisms of human vision, the transform
used must accept nonstationarity and be well
localized in both the space and frequency
domain.
To avoid redundancy, the transform must be at
least biorthogonal and lastly, in order to save
CPU time, the corresponding algorithm must be
fast. The wavelet transform satisfies each of
these conditions.
7
II. A Short Review of Wavelet Analysis
Scaling functions
The basic constructions of wavelets using
scaling functions is as follows:
1. Define a scaling function
( x) p
k
k (2 x k )
2. Define a subspace V of a vector space U, U is a
collection of elements over the real number R,
then VU.
8
3. Given a nested sequences of subspace V j , V j
is defined as
V j clos L 2 { j, k j, k Z }
where j , k ( 2 j x k ) ,
then we have
V0 V1 V1 V0 V1 V2
Coarser Finer
( containment property )
9
Wavelets
1. In containment property, there exists subspace Wj ,
which are orthogonal complements of V j in V j1 ,
that is, V j 1 V j Wj , j Z ,
and Wj Wj , if j j .
2. Since the subspaces V j are nested, it follows that
J j 1
VJ V j
k 0
Wj k for j < J ,
L2 ( R) Wj
j Z
W1 W0 W1 .
10
3. Given a scaling function in V j , there exists
another function in W0 called the wavelet,
such that { j , k k Z } generates Wj , where
j , k ( 2 j x k ) , j, k Z .
11
Decomposition and Reconstruction
1. Since V1 V0 W0 , (2 x ) and (2 x 1) V1 ,
we have
(2 x ) [ a2 k ( x k ) b 2 k ( x k ) ]
k
(2 x 1) [a
k
1 2 k ( x k ) b12 k ( x k ) ]
2. The decomposition relation can be generalized
as
(2 x l ) [a
k
l 2 k ( x k ) bl 2 k ( x k )], l Z .
12
3. The reconstruction relation can be formulated
as
( x) p
l=-
l ( 2 xl ) ,
( x) q
l=-
l ( 2 xl ) .
f ( x) K
cN , k ( 2 N x k ) f N ,
where cN , k f ( x ), N , k .
13
5. Since V j V j 1 Wj 1 ,
f N has a unique wavelet decomposition :
f N f N 1 g N 1
g N 1 g N 2 g N M f N M
where f j V j and g j Wj for any j .
14
6. To describe decomposition and reconstruction
algorithms, f j and g j can be represented as
follows.
f j ( x) c j , k ( 2 j x k ) ,
k
g j ( x)
k
d j, k ( 2 j x k ) ,
where
c j, k f j ( x ), j ,k ( x ) ,
d j, k f j ( x ), j ,k ( x ) .
15
7. Wavelet decomposition algorithm :
c j 1, k a j 1, l 2 k c j, l ,
l
d j 1, k
b
l
j 1, l- 2 k c j, l .
d N 1 d N 2 dNM
cN cN 1 c N 2 cN M
16
8. Wavelet reconstruction algorithm :
dNM d N M 1 d N 1
cN M cN M 1 cN 1 cN
17
III. A Simple Example : Haar Wavelets
Scaling functions
1. Haar scaling function is defined by
1 for 0 x 1
( x)
0 otherwise
and is shown in Figure 1.
Some examples of its translated and scaled
versions are shown in Figures 2-4.
18
1 1
0 0
-1 -1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Fig.1: Haar scaling function (x). Fig.2: Haar scaling function (x-1).
1 1
0 0
-1 -1
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Fig.3: Haar scaling function (2x). Fig.4: Haar scaling function (2x-1).
19
2. The two-scale relation for Haar scaling function
is
( x) p
k
k (2 x k )
(2 x ) (2 x 1)
Therefore, the two-scale sequence { pk } for Haar
scaling function have non-zero values p0 p1 1
and 0’s for other p j’s .
20
Wavelets
1. The Haar wavelet (x) is given by
1 for 0 x 21
( x) 1 for 21 x 1
0 otherwise
and is shown in Figure 5.
2. The two-scale relation for Haar wavelet is
( x) (2 x) (2 x 1) .
21
1
-1
22
Decomposition relation
1. Both of the two-scale relation together are called
the reconstruction relation.
( x) 1 1 (2 x )
( x ) 1 1 (2 x 1)
2. The decomposition relation can be derived as
follows.
(2 x ) 1 1
( x)
(2 x 1)
2 2
21 ( x)
1
2
23
IV. Subband filtering scheme
ROWS COLUMNS Image
corresponding
h 12 to the low
resolution
level m
h 21
Initial image g 12
corresponding Detail images
to the corresponding
resolution
level m-1
h 12 to the
resolution
level m
g 21
g 12
h Convolve with low-pass filter 21 Keep one column out of two
g Convolve with high-pass filter 12 Keep one row out of two
24
m2 m=2 m=1
Resolution m=2
Low resolution Horizontal
sub-image orientation Resolution m=1
sub-image Horizontal
Resolution m=2 Resolution m=2 orientation
Vertical Diagonal sub-image
orientation orientation
sub-image sub-image
m : resolution level
Figure 7: Image decomposition .
25
COLUMNS ROWS
Image
corresponding ~
to the low 12 h 2
resolution
level m 21 ~
h 2 Reconstructed
image
~
12 g 2 resolution
level m-1
Detail
images
~
resolution
level m
12 h 2
~
21 g 2
~
12 g 2
X Convolve with filter X 12 Put one row of zero between each row
26