Advisor : Jian-Jiun Ding, Ph. D.

Presenter : Ke-Jie Liao

NTU,GICE,DISP Lab,MD531
1
Introduction
Continuous Wavelet Transforms
Multiresolution Analysis Backgrounds
Image Pyramids
Subband Coding
MRA
Discrete Wavelet Transforms
The Fast Wavelet Transform
Applications
Image Compression
Edge Detection
Digital Watermarking
Conclusions



2
Why WTs?
F.T. totally lose time-information.
Comparison between F.T., S.T.F.T., and W.T.


f f f
t t t
F.T. S.T.F.T. W.T.
3
Difficulties when CWT DWT?
Continuous WTs Discrete WTs
need infinitely scaled wavelets to represent a given
function Not possible in real world
Another function called scaling functions are used to
span the low frequency parts (approximation parts)of
the given signal.
Sampling
F.T.
,
1
( ) ( )
s
x
x
s
s
t
t


=
0 0
,
0
0
1
( ) ( )
j
s j
j
x k s
x
s
s
t
t


=
Sampling
0 , 0 0
( ) exp ] ( [ 2 ( ) )
j
s
j
x A j s s f x k
t
t t + = +
4
[5]
MRA
To mimic human beings perception characteristic

5
[1]
Definitions
Forward
where

Inverse exists only if admissibility criterion is satisfied.



,
( , ) ( ) ( )
s
W s f x x dx
t
t

=
} ,
1
( ) ( )
s
x
x
s
s
t
t


=
( ) ( )
2
0
1
,
x
f x W s d ds
s
C s s

t
t t

| |
=
|
\ .
} }
2
| ( ) |
| |
f
C df
f

+
= <
}
C

<
6
An example
-Using Mexican hat wavelet
7
[1]
Image Pyramids
Approximation pyramids
Predictive residual pyramids
8 N*N
N/2*N/2
N/4*N/4
N/8*N/8
Image Pyramids
Implementation
9
[1]
Subband coding
Decomposing into a set of bandlimited components
Designing the filter coefficients s.t. perfectly
reconstruction

10
[1]
Subband coding
Cross-modulated condition



Biorthogonality condition
0 1
1
1 0
( ) ( 1) ( )
( ) ( 1) ( )
n
n
g n h n
g n h n
+
=

1
0 1
1 0
( ) ( 1) ( )
( ) ( 1) ( )
n
n
g n h n
g n h n
+
=

(2 ), ( ) ( )
i j
h n k g k i j o =
11
or
[1]
Subband coding
Orthonormality for perfect reconstruction filter


Orthonormal filters
( ), ( 2 ) ( ) ( )
i j
g n g n m i j m o o + =
1 0
( ) ( 1) ( 1 )
n
even
g n g K n =
( ) ( 1 )
i i even
h n g K n =
12
The Haar Transform

1 1
1
1 1
2
| |
=
|

\ .
2
H
| |
0
1
( ) 2 0
2
H k =
| |
1
1
( ) 0 2
2
H k =
DFT
| |
1
1
( ) 1 1
2
h n =
| |
0
1
( ) 1 1
2
h n =
13
[1]
Any square-integrable function can be represented by
Scaling functions approximation part
Wavelet functions - detail part(predictive residual)
Scaling function
Prototype
Expansion functions


/2
,
( ) 2 (2 )
j j
j k
x x k =
2
( ) ( ) x L e R
,
{ ( )}
j j k
V span x =
14
MRA Requirement
[1] The scaling function is orthogonal to its integer
translates.
[2] The subspaces spanned by the scaling function at low
scales are nested within those spanned at higher scales.


1 0 1 2
V V V V V V

c c c c c c c
15
[1]
MRA Requirement
[3] The only function that is common to all is .

[4] Any function can be represented with arbitrary
precision.







j
V
( ) 0 f x =
{0} V

=
2
{ ( )} V L

= R
16
Refinement equation
the expansion function of any subspace can be built
from double-resolution copies of themselves.
1 j j
V V
+
c
( 1)/ 2 1
,
( ) ( )2 (2 )
j j
j k
n
x h n x n

+ +
=

, 1,
( ) ( ) ( )
j k j n
n
x h n x

+
=

1/ 2
( ) ( )2 (2 )
n
x h n x n

Scaling vector/Scaling function coefficients

17
/2
,
( ) 2 (2 )
j j
j k
x x k =
Wavelet function
Fill up the gap of any two adjacent scaling subspaces
Prototype
Expansion functions
( ) x
/2
,
( ) 2 (2 )
j j
j k
x x k =
,
{ ( )}
j j k
W span x =
1 j j j
V V W
+
=
0 0 0
2
1
( )
j j j
L V W W
+
= R
18
[1]
Wavelet function







Scaling and wavelet vectors are related by

1 j j
W V
+
c
, 1,
( ) ( ) ( )
j k j n
n
x h n x

+
=

( 1)/ 2 1
,
( ) ( )2 (2 )
j j
j k
n
x h n x n

+ +
=

1/ 2
( ) ( )2 (2 )
n
x h n x n

Wavelet vector/wavelet function coefficients

( ) ( 1) (1 )
n
h n h n
|
=
19
Wavelet series expansion

0 0
0
, ,
( ) ( ) ( )
( ) ( ) ( ) ( ) ( )
a d
j j k j j k
k j j k
f x f x f x
f x c k x d k x

=
= +
= +

0 0 0
2
1
( )
j j j
L V W W
+
= R
( ) f x
( )
a
f x
( )
d
f x
0
j
W
0
j
V
0
1 j
V
+
0
( ) 0
j
d k =
0
j j >
20
Discrete wavelet transforms(1D)
Forward




Inverse
0
0 ,
1
( , ) ( ) ( )
j k
n
W j k f n n
M

=

, 0
1
( , ) ( ) ( ) ,
j k
n
W j k f n n for j j
M

= >

0
0
0 , ,
1 1
( ) ( , ) ( ) ( , ) ( )
j k j k
k j j k
f n W j k n W j k n
M M

=
= +

21
Fast Wavelet Transforms
Exploits a surprising but fortune relationship between
the coefficients of the DWT at adjacent scales.
Derivations for

( ) ( ) 2 (2 )
n
p h n p n
|
| | =

( , ) W j k

( )
(2 ) ( ) 2 2(2 )
j j
n
p k h n p k n
|
| | =

( )
1
( 2 ) 2 2
j
m
h m k p m
|
|
+
=

2 m k n = +
22
Fast Wavelet Transforms
Derivations for
( , ) W j k

/ 2
/ 2 1
( 1)/ 2 1
1
( , ) ( )2 (2 )
1
( )2 ( 2 ) 2 (2 )
1
( 2 ) ( )2 (2 )
( 2 ) ( 1, )
j j
n
j j
n m
j j
m n
m
W j k f n n k
M
f n h m k n m
M
h m k f n n m
M
h m k W j k

|
|
+
+ +
=
(
=
(

(
=
(

= +

,
1
( , ) ( ) ( )
j k
n
W j k f n n
M

=

( )
1
(2 ) ( 2 ) 2 2
j j
m
n k h m k n m
|
| |
+
=

2 , 0
( , ) ( ) ( 1, ) |
n k k
W j k h n W j n
= >
= - + 23
Fast Wavelet Transforms
With a similar derivation for


An FWT analysis filter bank
( , ) W j k

2 , 0
( , ) ( ) ( 1, ) |
n k k
W j k h n W j n
= >
= - +
24
[1]
FWT
25
[1]
Inverse of FWT
Applying subband coding theory to implement.
acts like a low pass filter.
acts like a high pass filter.
ex. Haar wavelet and scaling vector
( ) h n

( ) h n
|

DFT
| |
1
( ) 1 1
2
h n

=
| |
1
( ) 1 1
2
h n

=
| |
1
( ) 2 0
2
H k

=
| |
1
( ) 0 2
2
H k

=
26
[1]
2D discrete wavelet transforms
One separable scaling function


Three separable directionally sensitive wavelets

( , ) ( ) ( ) x y x y | | | =
( , ) ( ) ( )
H
x y x y | =
( , ) ( ) ( )
V
x y y x | =
( , ) ( ) ( )
D
x y x y =
x
y
27
2D fast wavelet transforms
Due to the separable properties, we can apply 1D FWT
to do 2D DWTs.

28
[1]
2D FWTs
An example
LL LH
HL HH
29
[1]
2D FWTs
Splitting frequency characteristic
30
[1]
Image Compression
have many near-zero coefficients
JPEG : DCT-based
JPEG2000 : FWT-based
, ,
H V D
W W W

DCT-based FWT-based
31
[3]
Edge detection
32
[1]
Digital watermarking
Robustness
Nonperceptible(Transparency)
Nonremovable

Digital watermarking Watermark extracting
Channel/
Signal
processin
g
Watermark
Original and/or
Watermarked data
Secret/Public key
Secret/Public key
H
o
s
t
d
a
t
a
Watermark
or
Confidence
measure
33
Digital watermarking
An embedding process

34
Wavelet transforms has been successfully applied to
many applications.
Traditional 2D DWTs are only capable of detecting
horizontal, vertical, or diagonal details.
Bandlet?, curvelet?, contourlet?
35
[1] R. C. Gonzalez, R. E. Woods, "Digital Image
Processing third edition", Prentice Hall, 2008.
[2] J. J. Ding and N. C. Shen, Sectioned Convolution
for Discrete Wavelet Transform, June, 2008.
[3] J. J. Ding and J. D. Huang, The Discrete Wavelet
Transform for Image Compression,,2007.
[4] J. J. Ding and Y. S. Zhang, Multiresolution Analysis
for Image by Generalized 2-D Wavelets, June, 2008.
[5] C. Valens, A Really Friendly Guide to Wavelets,
available in http://pagesperso-
orange.fr/polyvalens/clemens/wavelets/wavelets.html

36