Ch.

15 Wavelet-Based Compression

1
Origins and Applications
The Wavelet Transform (WT) is a signal processing tool that is replacing the Fourier
Transform (FT) in many (but not all!) applications.
WT theory has its origins in ideas in three main areas and now is being applied in
countless different areas of application.

Physics Mathematics Signal Processing

Quantum Mech. Harmonic Analysis Signal Decomp.
Group Theory Filter Banks
Operator Theory Image Analysis

Signal Decompositions
Wavelets Filter Banks
Wavelets

Applications
Data Compression Sonar/Radar Turbulence
Computer Vision Biomedical Proc. Geophysical Proc.
Denoising Signals Communications Music Processing
Etc. Etc. Etc.
2
Next
So, Whats Wrong With The FT?
First, recall the FT:
X ( f ) x (t )e j 2 ft dt

x (t ) X ( f )e j 2 ft df

Weight @ f Component @ f

Remember: An integral is like a summation So, the second equation says that
we are decomposing x(t) into a weighted sum of complex exponentials
(sinusoids!) The first equation tells what each weight should be.

Note: These components exist for ALL time!!!

This is not necessarily a good model for real-life signals.

3
Next
 DFT Basis Functions and Time-Freq Tiles
Freq

Time

4
Next
Example: Frequency-Hopping Chirped Pulses
DFT tries to build from
freq components that
last 10 sec

The DFT representation is

correct but does not show
us a joint representation in
time and frequency
This would be a more
desirable time-frequency
representation

It would show at each time

exactly what frequencies
were in existence at that time!
However such a perfect
t-f representation is not
possible Heisenberg
Uncertainty Principle! 5
Next
But What About The Short-Time FT (STFT)?

X ( f , t)

x ( )h( t )e j 2 f d
h( - t) Dummy
Variable!

Spectrogram = |X(f,t)|2

Take the FT of this

Thus, for each t value
we get a FT

Selesnick, Ivan, "Short Time Fourier Transform," Connexions, August 9, 2005.

http://cnx.org/content/m10570/2.4/. 6
Next
STFT T-F Resolution
The window function h(t) sets the characteristic of how the
STFT is able to probe the signal x(t).
The narrower h(t) is, the better you can resolve the time of
occurrence of a feature
However the narrower h(t) is, the wider H(f) is and that means
a reduction in the ability to resolve frequency occurance
Just like windowing of the DFT that youve probably studied!
Each given h(t) has a given time and frequency resolution
t describes the time resolution
f describes the frequency resolution
The Heisenberg Uncertainty Principle states that
1
( t )( f )
4
Improving Time Resolution. Degrades Frequency Resolution
And vice versa

7
Illustration of Time-Frequency Resolution Trade-Off
cos(2 10t ) 0t5
cos(2 25t ) 5 t 10

x (t )
cos(2 50t ) 10 t 15
cos(2 100t ) 15 t 20

Window Width Window Width

25 ms 125 ms

Window Width Window Width

375 ms 1000 ms

http://en.wikipedia.org/wiki/Short-time_Fourier_transform 8
STFT View of Tiling the T-F Plane
Generally only compute the STFT for discrete values of t and f

X ( f m , tn )

x ( )h( nT )e j 2 ( mF ) d

In some applications it is desirable to minimize the number of points in

X(fm,tn) and that means making T and F as large as possible T t and
F f
Then each X(fm,tn) represents the content of the signal in a rectangular cell
of dimension t by f
STFT tiling consists of a
uniform tiling by fixed
rectangles

9
STFT Basis Functions and Time-Freq Tiles
Freq

Time

10
Next
STFT Disadvantages and Advantages
The fact that the STFT tiles the plane with cells having the same
t and f is a disadvantage in many application
Especially in the data compression!
This characteristic leads to the following:
If you try to make the STFT be a non-redundant decomposition (e.g.,
ON like is good for data compression
You necessarily get very poor time-frequency resolution
This is one of the main ways that the WT can help
It can provide ON decompositions while still giving good t-f resolution
However, in applications that do not need a non-redundant
decomposition the STFT is still VERY useful and popular
Good for applications where humans want to view results of t-f
decomposition

11
So What IS the WT???

Recall the STFT: X ( f , t)

x ( )h( t )e j 2 f d

Basis Functions
So X(f,t) is computing by comparing x(t) to each of these basis functions

For the STFT the basis function are created by applying

Time Shift and Frequency Shift to prototype h(t)

This leads to the uniform tiling we saw before

And it also causes the problems with the non-redundant form of the STFT

So we need to find a new way to make T-F basis functions that dont have
these problems!!!

12
The WT comes about from replacing frequency shifting by time scaling
Start with a prototype signal (t) and time scale it:
(t / s ) s ( sf )
Increasing s: Stretches the signal Decreasing s: Scrunches the signal
Scrunches the spectrum Stretches the spectrum
(t)

Small s t Large s
|( f )|

(t/s) (t/s)
f

t t

|(sf )| Small s covers High Large s covers Low |(sf )|

f w/ Broad Width f w/ Narrow Width

f f
13
This shows some typical t-f cells for wavelets

It is generally more convenient

to plot in terms of 1/s due to its
link to frequency
Increasing 1/s

Narrower in Time
Wider in Frequency
Higher Frequency

Wider in Time
Narrower in Frequency
Lower Frequency

1
We still need to satisfy the uncertainty principle: ( t )( f )
4
But now t and f are adjusted depending on what region of frequency
is being probed. 14
Next
All this leads to The Wavelet Transform:

1 t (t) is called the
X ( s, t ) s s d , s 0
x ( )
Mother Wavelet

The Inverse Wavelet Transform (Reconstruction Formula):

2
1 1 t dsd
( )
x (t )
C 0 X ( s , ) s s s 2 ,

C
0

d

Requirements for a Mother Wavelet are:

(t )dt
2
Finite Energy: (t ) L ( )
2 2

( ) d
Parseval
2

( )
Admissibility Condition:
0
| |
d

|()|2 must go to zero fast enough as 0

|()|2 must go to zero fast enough as
(t) must be a bandpass signal 15
The prototype basis function (t) is called the mother wavelet
All the other basis functions come from scaling and shifting the mother wavelet.
There are many choices of mother wavelet:
Each gives rise to a slightly different WT
with slightly different characteristics
suited to different applications.

16
Next
Example of a WT
log2(1/s)

17
Next
 Non-Redundant Form of WT
It is often desirable to use a discrete form of the WT that is non-redundant
that is, we only need X(s,t) on a discrete set of s and t values to reconstruct x(t).
Under some conditions it is possible to do this with only s and t taking these
values: s 2 m t n 2 m for m 3, 2, 1,0,1, 2,3, In practice you
truncate the range
n 3, 2, 1,0,1, 2,3,
of m and n

Lower values of m Smaller values of s Higher Frequency

Incrementing m doubles the scale value and doubles the time spacing
Then the WT becomes a countably infinite collection of numbers (recall the
Fourier series vs. the Fourier transform):

n d , m, n

t


m /2 m
X ( s, t ) x ( ) 1s X m ,n x ( ) 2 2
d
s m , n ( )

Advantages of This Form

The m,n(t) can be an ON basis for L2
Good for Data Compression x (t ) X 2 m /2 2 m t n
Simple, numerically stable inverse m n
m ,n
Leads to efficient discrete-time form 18
This leads to the sampling and the tiling of the t-1/s plane as shown below:

19
Next
 WT Tiling and Basis Functions
Freq

Time

20
Next
Example #1
A synthetic Chirp Signal
Frequency decreases with time
Amplitude increases with time
Notice that
High frequency components dominate early
Low frequency components dominate later
Low frequency components are stronger
Signal Wavelet Transform

log2(1/s)
WT Coefficients are
shown as spikes on an
axis for each 1/s value

time time
21
Next
Example #2
Speech Signal

WT coeffs are displayed as

gray- scale blocks
WT coeffs concentrated
Blocks closely spaced at high f
Blocks widely spaced at low f
log2(1/s)

22
Next
Example #3: Effectiveness of WT T-F Localization Properties
Seismic Trace Signal
Keep 100 largest DCT coefficients
Large Error
especially at start of transient

Same Seismic Trace Signal

Keep 100 largest WT coefficients
SMALL Error
even at start of transient

23
Next
 Summary So Far: What is a Wavelet Transform?
Note that there are many ways to decompose a signal. Some are:
Fourier series: basis functions are harmonic sinusoids;
Fourier transform (FT): basis functions are nonharmonic sinusoids;
Walsh decomposition: basis functions are harmonic square waves;
Karhunen-Loeve decomp: basis functions are eigenfunctions of covariance;
Short-Time FT (STFT): basis functions are windowed, nonharmonic sinusoids;
Provides a time-frequency viewpoint
Wavelet Transform: basis functions are time-shifted and time-scaled versions of
a mother wavelet m ,n (t ) 2 m /2 2 m t n
Provides a time-scale viewpoint

x (t ) X
m n
m ,n t
m ,n X m ,n x( ) m ,n ( )d ,

Wavelet transform also provides time-frequency view: 1/scale relates to f

Decomposes signal in terms of duration-limited, band-pass components
high-frequency components are short-duration, wide-band
low-frequency components are longer-duration, narrow-band
Can provide combo of good time-frequency localization and orthogonality
the STFT cant do this
24
Next
 Fourier Development vs. Wavelet Development
Fourier and others:
expansion functions are chosen, then properties of transform are found
Wavelets
desired properties are mathematically imposed
the needed expansion functions are then derived
Why are there so many different wavelets?
the basic desired property constraints dont use all the degrees of freedom
remaining degrees of freedom are used to achieve secondary properties
these secondary properties are usually application-specific
the primary properties are generally application-nonspecific

25
Next
 Why are Wavelets Effective?
Provide a good basis for a large signal class
wavelet coefficients drop-off rapidly
thus, good for compression, denoising, detection/recognition
goal of any expansion is
have the coefficients provide more info about signal than time-domain
have most of the coefficients be very small (sparse representation)
FT is not sparse for transients WT is sparse for many signals
Accurate local description and separation of signal characteristics
Fourier puts localization info in the phase in a complicated way
STFT cant give localization and orthogonality
Wavelets can be adjusted or adapted to application
remaining degrees of freedom are used to achieve goals
Computation of wavelet coefficient is well-suited to computer
no derivatives or integrals needed
turns out to be a digital filter bank as we will see.
26
Next
The Discrete-Time WT
Recall the formula for the WT coefficients:

X m ,n
x ( )
2 m /2
2

m
n d , m, n ()

m , n ( )

If the signal x(t) is bandlimited to B Hz, we can represent it by its

samples taken every TS = 1/2B seconds: x[k] = x(kTs).
Our Goal: Since the samples x[k] uniquely and completely
describe x(t), they should also uniquely and completely describe
the WT coefficients Xm,n HOW DO WE DO IT??
via ()
x(t) Xm,n
via Sampling
????
x[k]
27
Next
 Mathematically Interpret the DWT equation in terms of Signal
Processing:

X m ,n x ( ) 2 m /2 2 m n d ,

x ( )
2 m /2
2 m

( n 2 m
) d ,

m n 2m

x ( ) m n 2m d ,

x * m n 2
m x(t) m (t ) X m ,n
t n 2m

For fixed m, Xm,n is x(t) convolved Remember: m(t) is a bandpass signal so

with m(t) and sampled at points n2m. this is equivalent to filtering x(t) with a BPF.

28
Next
This leads to a filterbank:

m (t )
o
X mo ,n
t n 2mo

x(t) m 1 (t )
o
X ( mo 1),n
( mo 1)
t n2

M (t ) X M ,n
t n2M

m ( f ) M mo+2 mo+1 mo

f
It IS possible to implement this filterbank in DT
but in general it is not possible to efficiently implement it 29
Consider a Special Case, Though
Let x(t) be bandlimited to B Hz. Choose the mother wavelet to be a modulated
sinc function and mo such that the spectrum is as below

( f )

Signal
Spectrum is
zero above B

30
Next
Now consider the relationship
between the filter spectrum and X( f )
spectra of the scaled versions of
the mother wavelet

( f )

(2 f )

Dont have to go to
finer scale (i.e.,
smaller s more
negative m) because (4 f )
the BL signal has no
power there!
Finest scale we need
is on the order of 1/FS
31
Next
 Note that for these ideal filters
we can do this X( f )

Running the signal through

(2f ) is the same as
running it first through ( f ) (f) (f)
( f ) and then through
(2f ) etc.
Sampled
nT Slower
(t) (2 f )
(2f ) (2f )
x(t)
n2T
(t) (t/2)

n4T
(4 f ) (4f )
(t/2) (t/4)
(4f )

(t/4) Etc.
32
Next
Note that we can do this in
terms of DT processing S()

h1[n ] 2
x[n]
H0() H1()
h0 [n ] h1[n ] 4

h0 [n ]
h1[n ] 8 /2
Etc. H0(2) H1(2)
???
The tildes are used to indicate filter s
that correspond to a time-scaled version
of the filter w/o the tilde /4 /2

h1[n] and h0[n] implement the DT versions of (t) and (t), respectively.
The change in notation to h is to allow for the general non-sinc case
where the DT filters used are not simply DT versions of the CT filters (like
they are here for the sinc case). 33
Next
Modification of DT approach:
S()

h1[n ] 2
x[n] H() H0() H1()

h0 [n ] 2 h1[n ] 2
/2
h0 [n ] 2 H0() H1()

Etc.
/2
This development is valid for the sinc wavelet
But the general development is not as straight- After
forward as this example might indicate!!! Dec by 2

34
Next
 WT
Coefficients

X1n

X2n

X3n

X4n

Y4n
This structure is the final result of this
development although the way we did the
development is easily applicable only to the sinc What are
wavelet case the result holds true but is better these??
developed using the so-called multi-resolution
viewpoint which we will develop next.
35
Next
 X1n

X2n

X3n

X4n

This structure provides a way to compute Y4n

the inverse WT using a cascaded filter Sum
bank structure that is essentially a reverse
version of the filter bank used to compute
the WT.

36
Next
Multi-Resolution Viewpoint
Provides solid development of
DT Filter Bank Implementation

37
Next
 Multi-Resolution Approach
Stems from image processing field
consider finer and finer approximations to an image
Define a nested set of signal spaces

V 2 V1 V0 V1 V2 L2
We build these spaces as follows:
Let V0 be the space spanned by the integer translations of a fundamental
signal (t), called the scaling function: spanned by (t k)

So far we can use just about any function (t), but well see that to get the
nesting only certain scaling functions can be used.

38
Next
 Multiresolution Analysis (MRA) Equation
Now that we have V0 how do we make the others and ensure that they
are nested?
If we let V1 be the space spanned by integer translates of (2t) we get
the desired property that V1 is indeed a space of functions having
higher resolution.
Now how do we get the nesting?
We need that any function in V0 also be in V1; in particular we need
that the scaling function (which is in V0 ) be in V1, which then requires
that
(t ) h0 [n ] 2 (2t n )
n

where the expansion coefficient is h0[n]2

This is the requirement on the scaling function to ensure nesting: it
must satisfy this equation
called the multiresolution analysis (MRA) equation
this is like a differential equation for which the scaling function is the
solution
39
Next
 The h0[n] Specify the Scaling Function
Thus, the coefficients h0[n] determine the scaling function
for a given set of h0[n], (t)
may or may not exist (t ) h0 [n ] 2 (2t n )
n
may or may not be unique
Want to find conditions on h0[n] for (t) to exist and be unique, and
also:
to be orthogonal (because that leads to an ON wavelet expansion)
to give wavelets that have desirable properties

h0[n] (t)

n t
MRA
Equation

h0[n] must
satisfy some
conditions
40
Next
 Whence the Wavelets?
The spaces Vj represent increasingly higher resolution spaces
To go from Vj to higher resolution Vj+1 requires the addition of
details
These details are the part of Vj+1 not able to be represented in Vj
This can be captured through the orthogonal complement of Vj w.r.t
Vj+1
Call this orthogonal complement space Wj
all functions in Wj are orthogonal to all functions in Vj
That is:

j , k (t ), j ,l (t ) j , k (t ) j ,l (t ) dt 0 j , k , l Z

Consider that V0 is the lowest resolution of interest

How do we characterize the space W0 ?
we need to find an ON basis for W0, say {0,k(t)} where the basis
functions arise from translating a single function (well worry about the
scaling part later):
0, k (t ) (t k )
41
Next
 Finding the Wavelets
The wavelets are the basis functions for the Wj spaces
thus, they lie in Vj+1
In particular, the function (t) lies in the space V1 so it can be
expanded as
(t ) h1[n ] 2 (2t n ), n Z
n

This is a fundamental result linking the scaling function and the

wavelet
the h1[n] specify the wavelet, via the specified scaling function

(t)
h1[n]

n t
Wavelet
Equation
(WE)
h1(n) must
satisfy some
conditions
42
Next
 Wavelet-Scaling Function Connection
There is a fundamental connection between the scaling function and its
coefficients h0[n] , the wavelet function and its coefficients h1[n]:

h0[n] (t)

n MR t
Equation
(MRE)
How are h1[n] and
h0[n] related?

(t)
h1[n]

n t
Wavelet
Equation
(WE)
43
Next
 Relationship Between h1[n] and h0[n]
We state here the conditions for the important special case of
finite number N+1 of nonzero h0[n]
ON within V0: (t ) (t k )dt (k )
ON between V0 and W0 :
(t ) (t k )dt (k )
Given the h0[n] that define the desired scaling function, then the h1[n]
that define the wavelet function are given by

h1[n ] ( 1) n h0 [ N n ] Well see soon that the h

coefficients are really DT
where N is the order of the filter filter coefficients

Much of wavelet theory addresses the origin, characteristics, and

ramifications of this relationship between h1[n] and h0[n]
requirements on h0[n] and h1[n] to achieve ON expansions
how the MRE and WE lead to a filter bank structure
requirements on h0[n] and h1[n] to achieve other desired properties
extensions beyond the ON case

44
Next
 The Resulting Expansions
Suppose we have found a scaling function (t) that satisfies the MRE
Then (t - k) is an ON basis for V0


More generally, an ON basis for V jo is 2 (2 t k )
jo /2 jo
k

Since Vjo is a subspace of L2(R) we can find the best approximation

to x(t)L2(R) as follows
Not the most useful
x jo (t ) c jo ,k 2 jo /2
(2 t k )
jo expansion in practice
k

with c jo ,k x (t ), jo ,k (t )

x (t )2 jo /2 (2 jo t k )dt

xjo(t) is a low-resolution approximation to x(t)

Increasing jo gives a better (i.e., higher resolution) approximation

V 2 V1 V0 V1 V2 L2
45
Next
 The Resulting Expansions (cont.)
Weve set things up so that for some jo and its space Vjo we have that
L2 V j0 W j0 W j0 1 W j0 2

2 (2 t k )k
jo /2
We know that an ON basis for V jo is jo

We also know than on ON basis for W is 2 (2 t k )k

j /2 j
j

Thus we another form of the expansion: A VERY

useful

x (t ) c j0 ,k 2 j0 /2
(2 t k )
j0
j ,k
d 2 j /2
(2 j
t k) expansion in
k k j j0
practice

Low-Resolution Wavelet Details

Approximation

c jo ,k x (t ), jo ,k (t ) (2 t k )dt d j ,k x (t ), j ,k (t ) x (t )2 j /2 (2 j t k )dt
jo /2
x (t )2 jo

Same as the Xj,k WT coefficients

in the earlier notes 46
Next
 The Resulting Expansions (cont.)
If we let jo go to then

L2 W 2 W1 W0 W1 W2

And the above expansion becomes

Not the most useful
x (t ) d j ,k 2 (2 t k )
j /2 j
expansion in practice
k j

d j ,k x (t ), j ,k (t )

x (t )2 j /2 (2 jo t k )dt

This is most similar to the true wavelet decomposition as it was

originally developed

47
Next
 The Expansion Coefficients cj0(k) and dj(k)

We consider here only the simple, but important, case of ON

expansion
i.e., the s are ON, the s are ON, and the s are ON to the s
Then we can use standard ON expansion theory:

c j0 ,k x (t ), j0 ,k (t ) x (t ) j0 ,k (t ) dt

d j ,k x (t ), j ,k (t ) x (t ) j ,k (t ) dt

We will see how to compute these without resorting to computing

inner products
we will use the coefficients h1[n] and h0[n] instead of the wavelet and
scaling function, respectively
we look at a relationship between the expansion coefficients at one level
and those at the next level of resolution

48
Next
Filter Banks and DWT

49
Next
 Generalizing the MRE and WE
Here again are the MRE and the WE:

(t ) h0 [n ] 2 (2t n ) (t ) h1[n ] 2 (2t n )

n n

scale & translate: replace t 2 j t k

We get:
MRE WE

(2 j t k ) h0 [m 2k ] 2 (2 j 1 t m ) (2 j t k ) h1[m 2k ] 2 (2 j 1 t m )
m m

Connects Vj to Vj+1 Connects Wj to Vj+1

50
Next
 Linking Expansion Coefficients Between Scales
Start with the Generalized MRA and WE:

(2 j t k ) h0 [m 2k ] 2 (2 j 1 t m ) (2 j t k ) h1[m 2k ] 2 (2 j 1 t m )
m m

c j ,k x (t ), j ,k (t ) d j ,k x (t ), j ,k (t )

c j ,k h0 [m 2k ] x (t ), 2( j 1)/2 (2 j 1 t m ) d j ,k h1 [m 2k ] x (t ), 2( j 1)/2 (2 j 1 t m )
m m

c j1 (m)

c j ,k h [m 2k ] c
0 j 1, m
d j ,k h1[m 2k ] c j 1,m
m m

51
Next
 Convolution-Decimation Structure
c j ,k h [m 2 k ] c
0 j 1, m
d j ,k h1[m 2k ] c j 1,m
m m

Convolution
y0 [n ] c j 1[n ] h0 [ n ] y1[n ] c j 1[n ] h1[ n ]
h0 [m n ] c j 1[m] h1[m n ] c j 1[m ]
m m

Decimation

n=0 1 2 3 4 5 6 7 8 9

k=0 1 2 3 4

n = 2k = 0 2 4 6 8

52
Next
 Computing The Expansion Coefficients
The above structure can be cascaded:
given the scaling function coefficients at a specified level all the lower
resolution cs and ds can be computed using the filter structure

HPF dj,k
2
h1(-n) Wj
cj+1,k
Vj+1
LPF Cj,k HPF dj-1,k
2 2
h0(-n) Vj h1(-n) Wj-1

LPF cj-1,k HPF dj-2,k

2 2
h0(-n) Vj-1 h1(-n) Wj-2

LPF cj-2,k
2
h0(-n) Vj-2

53
Next
 Filter Bank Generation of the Spaces
Vj-1
Vj
Vj+1
Vj-2 Wj-2 Wj-1 Wj

/8 /4 /2

HPF dj,k
2
h1(-n) Wj
cj+1,k
Vj+1
LPF Cj,k HPF dj-1,k
2 2
h0(-n) Vj h1(-n) Wj-1

LPF cj-1,k HPF dj-2,k

2 2
h0(-n) Vj-1 h1(-n) Wj-2

LPF cj-2,k
2
h0(-n) Vj-2
54
Next
 WT Coefficients
How do we get HPF
h1(-n)
2
d0,k

c1,k
these???!!!
LPF c0,k HPF d-1,k
2 2
h0(-n) h1(-n)
c1,k x(t )2 (2t k )dt
1/2

LPF c-1,k HPF d-2,k

2 2
h0(-n) h1(-n)

k
Conjecture: c j 1,k x LPF
2
c-2,k

2B Samples of BL Signal h0(-n)

Consider case where (2t k ) sinc(2 Bt k ) sinc 2 B(t k / 2 B )

X( f )
( f/2 )
kT
(t)
x(t)
k2T
(t) (t/2) ( f ) (f) (f)

55
Next

(2t k ) sinc(2 Bt k )
n
First, prove this: x x(t )sinc(2 Bt k )dt
2 B

x(t )sinc(2 Bt k )dt x(t )sinc 2 B(t k / 2 B) dt

Generalized
FT of sinc = rectangle Parsevals
So for the sinc wavelet B Theorem
j 2 f
X ( f ) 1 e
k

case the conjecture is true 2B

df
with perfect equality B Pure
Cleverness!!
B j 2 ft j 2 f 2 B
X ( f )e
k
e df
For other cases there is B t 0
some high enough scale where
this result holds B j 2 f ( t 2kB )
approximately! X ( f )e df
B t 0 Inverse FT

x (t 2kB )t 0

x ( 2kB ) 56
Next
So now we can change this
WT Coefficients d0,k
HPF
2
h1(-n)
c1,k

LPF c0,k HPF d-1,k

2 2
h0(-n) h1(-n)

LPF c-1,k HPF d-2,k

2 2
h0(-n) h1(-n)

LPF c-2,k
2
h0(-n)

Into this
WT Coefficients d0,k
HPF
2
h1(-n)
x[n]

LPF c0,k HPF d-1,k

2 2
h0(-n) h1(-n)

LPF c-1,k HPF d-2,k

This is the DWT and the 2 2
h0(-n) h1(-n)
most widespread way of
computing the wavelet c-2,k
LPF
transform in practice! h0(-n)
2
57
Next
 Connection between Notations of WT Notes


1 t 1 t dsd
X ( s, ) x ( t ) dt x (t ) X ( s, ) 2
s s 0 s s s

X mn x (t ) 2 2 t n dt
m /2 m
x (t ) X mn 2 m /2 2 m t n
m n

Compute dj(k) & cj(k)
x (t ) d j ( k )
2 j /2
2 j
t k

using filter bank k j

x (t ) c j0 ,k 2 j0 /2
(2 t k )
j0
d j ,k 2 j /2 (2 j t k )
k k j j0

End 58
 59
Next
 60
Next
 61
Next
Computational Complexity of DWT
For a signal of length N: # of Multiplies & Adds is O(N)
Lower order than the FFT which is O(Nlog2N)
But watch out for the multiplicative constant!
Each Filter has length L << N

For 1st Stage:

Each of two filters: computes N/2 outputs, each requiring L multiples
# Multiplies for 1st stage = NL/2 + NL/2 = NL
For 2nd Stage:
Each of two filters: computes N/4 outputs, each requiring L multiples
# Multiplies for 1st stage = NL/4 + NL/4 = NL/2
For 3rd Stage:
Each of two filters: computes N/8 outputs, each requiring L multiples
# Multiplies for 1st stage = NL/8 + NL/8 = NL/4

# of Mult. < NL 2 m 2 NL
m 0
62
Next
 63
Next
 WT-BASED COMPRESSION EXAMPLE
Side Info

h1(1) to h1(8) h1q(1) to h1q(8)

Q1

h1(9) to h1(16) h1q(9) to h1q(16)

s(t) 8-Level Q2
Wavelet
Transform
l8(1) to l8(8) l8q(1) to l8q(8)
Q256

Bits allocated to quantizers to minimize MSE

Then allocations less than Bmin are set to zero
Frequency

Eliminates negligible cells

Side info sent to describe allocations

Time
64
Next
Original Fingerprint Image Decoded Fingerprint Image after 26:1 CR

From http://www.amara.com/IEEEwave/IW_fbi.html
See also http://www.c3.lanl.gov/~brislawn/FBI/FBI.html 65
Next
 66
Next
 67
Next
End

68