You are on page 1of 28

A Theory for Multiresolution

Signal Decomposition: The


Wavelet Representation
Stephane G. Mallat,
IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 11, No. 7, July, 1989,
pp. 674-693
Outline
Introduction
Multiresolution transform
Multiresolution approximation
Implementation of multiresolution transform
The wavelet representation
The detail signal
Implementation of orthogonal wavelet
representation
Conclusion
Introduction
Multiresolution approximation is wanted
It is natural to analyze first the image at a coarse
resolution then gradually increase the resolution


Introduction
Low pass filter can do this job
Ex: subsampling
Methodology:
Laplacian pyramid
Wavelet pyramid
and so on (maybe)

Outline
Introduction
Multiresolution transform
Multiresolution approximation
Implementation of multiresolution transform
The wavelet representation
The detail signal
Implementation of orthogonal wavelet
representation
Conclusion
Multiresolution transform
Refresh some definitions
Vector space
Basis
Inner product
Notation: <a,b>
Orthogonal
Orthonormal



Multiresolution transform
Definitions:



( )
( ) R L V x f
R L
V f A
V
x f f A
A
x f
j
j j
j
j
j
2
2
2
2 2
2
j
2
2
), (
functions integrable - square , measurable of space vector :
: Ex
space ion vector approximat :
resolution 2 at ) ( of ion approximat the is : Ex
operator ion approximat :
signal original : ) (
c
e
Multiresolution transform
-approximation
There exist scaling function (x)




Is an orthonormal basis of j
V
2
( ) ( )
Z n
j
j
j
n x
e

2 2
2
|
Multiresolution transform
-approximation
Examples of (x)
Multiresolution transform
-approximation
n u
j j
u u f

=
- =
2 2
| ) ( ) ( |
( )
Z n
j d
j
n
n
j j
n u u f f A
n x n u u f x f A
j j
j j j
e

+ =
=


- =

) 2 ( ), (
) 2 ( ) 2 ( ), ( 2 ) (
2 2
2 2 2
|
| |
Multiresolution transform
-approximation
can be interpreted as a low-pass
filtering of f(x) followed by a uniform
sampling at the rate
Original discrete signal is defined as
This filter is special because of its
orthogonal family
f A
d
j
2
j
2
f A
d
0
2
Multiresolution transform
-approximation
1 j j
1 j
1
2 2
2 2
2 2
basis with represent can We
of basis l orthonorna an also is of family
+
+
+

c
| |
|
j
j j
V
V V
Multiresolution transform
-Implementation
) ( ) (
~
) ( ), ( ) (
) 2 (
~
) 2 ( ), ( ) 2 (
~
) 2 ( ,
sides both with f(x) of product inner
) 2 ( ) 2 ( ), 2 ( 2
) 2 (
1
1
1
1 1
2
2 2
) 1 (
2 2
) 1 (
2
) 1 (
2 2
) 1 (
2
n h n h
n u u n h
f A k n h f A
k u u f k n h n u f (u)
k u k u n u
n x
Z n
d
k
d
k
j j
j
k
j j j
j
j j
j j
j j j
j
=
=
=
=
-
=
e

=
+
+

=
+ +

+
+
+ +

| |
| |
| | |
|
Multiresolution transform
-Implementation





This operation is called a pyramid transform

computed be can | all
, start with
and h
~
by computed be can
0
2
1
2
2 2
0
1
<
=
+
j
d
d d
d d
f A
A A
f A f A
j
j j
Multiresolution transform
-Implementation
(x) h(n) approximation method
Prefer (x) which is continuous differentiable
and asymptotic decay
Good localization properties in both freq. and
spatial domain
It is possible to choose H() to obtain a (x)

Outline
Introduction
Multiresolution transform
Multiresolution approximation
Implementation of multiresolution transform
The wavelet representation
The detail signal
Implementation of orthogonal wavelet
representation
Conclusion
The wavelet representation
Definitions
Detail signal: difference between f A f A
j j 1
2 2
and
+
j
2 2 2 2
2 resolution at f(x) of signal detail is ) (
O : space vector detail is O
operator detail is
2
1 j j
2
x f P
V V
P
j
j j
j
O
O
+
=
The wavelet representation
( )
( )
wavelet orthogonal an called is ) (
) ( L of basis l orthonorma an is ) 2 ( 2
O of basis l orthonorma an is ) 2 ( 2
) (
2
2

2
) (

2
) , (
2
2 2
2
j
x
R n x
n x
H e G with
G
Z j n
j j
Z n
j j
i
j
j

t e
e
e
|
e
e
e
e

e

+ =
|
.
|

\
|
|
.
|

\
|
|
.
|

\
|
=
The wavelet representation
) 2 ( ) 2 ( ), ( 2 ) (
2 2
2
n x n u u f x f P
j
n
n
j j
O
j j
j

+ =
=

- =


( )
Z n
j
n u u f f D
j j
e
) 2 ( ), (
2 2

n u
j j
u u f

=
- =
2 2
| ) ( ) (
The wavelet representation




Called orthogonal wavelet representation
( ) ( )
1
2 2
1
, by d represente be can
s s j J
d d
f D f A f A
j J
The wavelet representation
Examples of (x)

The wavelet representation
is a band pass filter,
which decompose the
signals into a set of
independent frequency
channels
Because overlap of
frequency channels,
can only provide
intuitive approach to
the model.
The wavelet representation
- implementation
1 j j
1 j
1
2 2
2 2
2 2
basis with represent can We
O of basis l orthonorna an also is of family
+
+
+

c
|
|
j
j j
V O
The wavelet representation
- implementation
) 2 ( ) 2 ( ), 2 ( 2
) 2 (
) 1 (
2
) 1 (
2 2
) 1 (
2
1 1
k u k u n u
n x
j
k
j j j
j
j j j
j
+

=
+ +

-
=
+ +

| |

) ( ) (
~
) 1 ( ) 1 ( ... ... ) ( ), ( ) (
) 2 (
~
) 2 ( ), ( ) 2 (
~
) 2 ( ,
1
2
2 2
) 1 (
2 2
1
1
1
n g n g
n h proof n u u n g
f A k n g f D
k u u f k n g n u f (u)
n
Z n
d
k
k
j j
j j
j j
=
= = =
=
=

=
+

+
+


|
|
The wavelet representation
- implementation
The wavelet representation
1 j
2
2 resolution at the signal the
of ty irregulari the of measure a gives D of energy The
j
+
f
( ) ( )
1
2 2
1
, by d represente be can
s s j J
d d
f D f A f A
j J
Reconstruction
Conclusion
Theory of orthogonal wavelet representation
has been showed
Didnt mention the extension : 2D transform
Application:
Texture discrimination
fractal analysis

You might also like