# Wavelets

```					   Wavelets (Chapter 7)

CS474/674 – Prof. Bebis
Multiresolution Analysis
• Many signals or images contain features at various
levels of detail (i.e., scales).
– e.g., buildings and people

Small size objects should
be examined at a high
resolution.

Large size objects should
be examined at a low
resolution.
Fourier Analysis

• Fourier analysis expands signals or functions in terms
of sinusoids (or complex exponentials).

• Identifies all spectral components present in a signal or
a function.

(inverse DFT)

where:                                            (forward DFT)
Fourier Analysis – Examples (cont’d)

f1 (t )  cos(2  5  t )

f 2 (t )  cos(2  25  t )

f 3 (t )  cos(2  50  t )
Fourier Analysis – Examples (cont’d)

F1(u)

F2(u)

F3(u)
Fourier Analysis – Examples (cont’d)

f 4 (t )  cos(2  5  t )
 cos(2  25  t )
 cos(2  50  t )

F4(u)
Limitations of Fourier Analysis

1. Fourier analysis cannot not provide simultaneous time
and frequency localization.

2.

3.
Simultaneous time - frequency localization
• Fourier analysis
identifies all spectral
components present
in the signal,
however it does not
provide any
information
regarding the
temporal (time)
localization of these
components.
Simultaneous time - frequency localization (cont’d)

• Frequency domain localization:
– If the signal is a sinusoidal, it is better localized in the
Fourier domain.
– Small changes in frequency will produce changes
everywhere in the spatial domain.
Simultaneous time - frequency localization (cont’d)

• Spatial domain localization:
– If the signal is a simple square pulse, then it is better
localized in the spatial domain.
– Small changes in the spatial domain will produce changes
everywhere in the frequency domain.
Limitations of Fourier Analysis (cont’d)

1. Fourier analysis cannot not provide simultaneous time
and frequency localization.

2. Not very useful for analyzing time-variant, non-
stationary signals/functions.

3.
Stationary vs non-stationary Signals

• Stationary signals:
time-invariant spectra
f 4 (t )

• Non-stationary
signals: time-varying
spectra.
f 5 (t )
Stationary vs non-stationary Signals (cont’d)

F4(u)

Perfect knowledge of what
frequencies exist, but no
these frequencies are
located in time

F5(u)
Limitations of Fourier Analysis

1. Fourier analysis cannot not provide simultaneous time
and frequency localization.

2. Not very useful for analyzing time-variant, non-
stationary signals/functions.

3. Not appropriate for representing discontinuities or
sharp corners (i.e., takes a large number of Fourier
components to represent discontinuities).
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners

FT
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Representing discontinuities or sharp corners
Simultaneous time - frequency
localization: revisited
 Need a local analysis scheme for a time-frequency
representation (TFR).
 Windowed F.T. or Short Time F.T. (STFT)
 Segmenting the signal into narrow time intervals (i.e.,
narrow enough to be considered stationary).
 Take the Fourier transform of each segment.
Short Time Fourier Transform (STFT)
• Steps of STFT:

(1) Choose a window function of finite length
(2) Place the window on top of the signal at t=0
(3) Truncate the signal using this window
(4) Compute the FT of the truncated signal, save.
(5) Incrementally slide the window to the right
(6) Go to step 3, until window reaches the end of the signal
Short Time Fourier Transform (STFT)

• For each time location where the window is centered,
we obtain a different FT
– Each FT provides the spectral information of a separate
time-slice of the signal, providing simultaneous time and
frequency information
Short Time Fourier Transform (STFT)
(cont’d)
Time       Frequency    Signal to                 FT Kernel
parameter    parameter   be analyzed             (basis function)

STFT fu (t , u )    f (t ) W (t  t )  e  j 2 ut dt
t

STFT of signal f(t):          Windowing             Windowing function
Computed for each              function               centered at t=t’
window centered at t=t’
Short Time Fourier Transform (STFT)
(cont’d)
• STFT maps 1-D time domain signals to 2-D time-
frequency signals (i.e., in u and t’ )

STFT (t , u )    f (t ) W (t  t )  e  j 2 ut dt
f
u

t
Example
f(t)

W(t)
Example (cont’d)

STFT (t, u)
u
f
W(t)
Practical Issues

• How to choose the windowing function?
• What shape?
– Rectangular, Gaussian, Elliptic…?
• How wide?
– Wider window require less time steps.
– Window should be narrow enough to make sure that the
portion of the signal falling within the window is stationary.
STFT Window Size
STFT fu (t , u )    f (t ) W (t  t )  e  j 2 ut dt
t

W(t) infinitely long: W (t )  1                  STFT turns into FT,
providing excellent frequency information (good frequency
resolution), but no time information.

W(t) infinitely short: W (t )   (t )  gives the time signal
back, with a phase factor, excellent time localization but no
frequency information.

STFT fu (t , u )    f (t )   (t  t )  e j 2 ut dt  f (t )  e  jut 
t
STFT Window Size (cont’d)

•   Wide analysis window  good frequency resolution,
poor time resolution.

•   Narrow analysis window  good time resolution,
poor frequency resolution.

•   Once the window is chosen, the resolution is set for
both time and frequency.
Example
different size windows

(four frequencies, non-stationary)
Example (cont’d)

STFTfu (t, u)

STFTfu (t, u)
Example (cont’d)

STFTfu (t, u)

STFTfu (t, u)
Heisenberg Principle

1
t  f 
4

Time resolution: How well      Frequency resolution: How
two spikes in time can be      well two spectral components
separated from each other in   can be separated from each
the transform domain.          other in the transform domain

t and f cannot be made arbitrarily small !
Heisenberg Principle (cont’d)

• One cannot know the exact time-frequency
representation of a signal.
– We cannot precisely know at what time instance a
frequency component is located.
– We can only know what interval of frequencies are present
in which time intervals.
The Wavelet Transform

• Overcomes the preset resolution problem of the STFT by
using a variable length window (i.e., multi-resolution):

– Analysis of high frequencies Use narrower windows for
better time resolution.

– Analysis of low frequencies  Use wider windows for
better frequency resolution.

• The function used to window the signal is called the
wavelet
The Wavelet Transform (cont’d)

Wide windows do not provide good localization
at high frequencies.
The Wavelet Transform (cont’d)

Use narrower windows at high frequencies.
The Wavelet Transform (cont’d)

Narrow windows do not provide good localization
at low frequencies.
The Wavelet Transform (cont’d)

Use wider windows at low frequencies.
The Wavelet Transform (cont’d)
• Using narrow windows for analyzing high frequencies
and wide windows for analyzing low frequencies works
quite well for signals having high frequency components
for short durations and low frequency components for
long durations.
What is a Wavelet?
• A wavelet is a "small wave“ of varying frequency and
limited duration.
• This is in contrast to sinusoids, used by FT, which have
infinite energy.

Sinusoid               Wavelet
What is a Wavelet? (cont’d)
• There are many different wavelets …

Haar wavelet      Morlet wavelet   Daubechies wavelet

• Basis functions, satisfying certain properties (e.g.,
orthogonal)
Basis Functions
(quick review – Section 7.2.1)
• Given a vector space of functions, S, then if any f(t) ϵ S can
be expressed as
f (t )   ak k (t )
k

the set of functions φk(t) are called the expansion set of S.

• If the expansion is unique, the set is a basis.
Basis Functions (cont’d)

• Starting a basis set φk(t), the vector space S can be
defined as the set of all functions f(t) that can be
expressed as:
f (t )   ak k (t )
k

• This is called the span of the basis set.
What is a Wavelet? (cont’d)
• The position and duration (scale) of wavelets vary.
Continuous Wavelet Transform
Translation parameter,   Scale parameter,          Normalization
measure of time          measure of frequency        constant

1                     t          (forward

CWTf ( , s)             f  t  

 dt   CWT)
s      t             s      

Continuous wavelet transform            The mother wavelet. All kernels are
of the signal f(t)                      obtained by translating (shifting) and/or
scaling the mother wavelet .

Scale = 1/frequency
Continuous Wavelet Transform (cont’d)

• Inverse CWT:

1                        t 
  CWT f ( , s) ( s )d ds

f (t ) 
s    s
CWT: Main Steps

1. Take a wavelet and compare it to a section at the start
of the original signal.

2. Calculate wavelet coefficient (i.e., convolution).
CWT: Main Steps (cont’d)

3. Shift the wavelet to the right and repeat steps 1 and 2 until
you've covered the whole signal.
CWT: Main Steps (cont’d)

4. Scale (stretch) the wavelet and repeat steps 1 through 3.

5. Repeat steps 1 through 4 for all scales.
Coefficients of Wavelet Transform
1                  t    

CWTf ( , s)          f  t  

 dt
s   t             s      
Properties of Wavelets

• Simultaneous localization in time and scale (i.e,
frequency)
- The location of the wavelet allows to explicitly represent the
location of events in time.
- The shape of the wavelet allows to represent different detail or
resolution.
Properties of Wavelets (cont’d)

• Sparsity: for functions typically found in practice,
many of the coefficients in a wavelet representation
are either zero or very small.

• Linear-time complexity: transforming to and from a
wavelet representation can generally be accomplished
in O(N) time.
– For some wavelets, it is still O(NlogN)
Properties of Wavelets (cont’d)

wide variety of functions (e.g., functions with
discontinuities, functions defined on bounded domains
etc.).
– Well suited to problems involving images, open or closed
curves, and surfaces of just about any variety.
– Can represent functions with discontinuities or corners more
efficiently (i.e., some have sharp corners themselves).
Discrete Wavelet Transform (DWT)
• Need to consider a subset of scales and positions rather than
every possible scale and position.

• Sample the time-frequency plane on a dyadic (octave) grid
(i.e., power of two).

Normalization
1                   t    
CWT f ( , s)           f  t            dt              constant
s   t               s      

 jk (t )  2 j  2 j t  k    j, k  Z
Discrete Wavelet Transform (cont’d)

 jk (t )  2 j  2 j t  k    j, k  Z

for frequency
localization
(i.e., details)

for time localization
Discrete Wavelet Transform (cont’d)
2D array of coefficients

a j ,k   f (t ) *,k (t )
j
t                     (forward DWT)

f (t )   a j ,k j ,k (t )          (inverse DWT)
k       j

where     jk (t )  2 j  2 j t  k    j, k  Z
DFT vs DWT

• Fourier expansion:

(inverse DFT)

one parameter basis

or   f (t )   a l l (t )
l
DFT vs DWT (cont’d)
• Wavelet expansion:
two parameter basis

f (t )   a j ,k j ,k (t )   (inverse DWT)
k   j
Multiresolution (i.e., hierarchical)
Representation
f (t )                                                                       high
resolution

ˆ
f1 (t )

ˆ
f2 (t )
…
ˆ
f s (t )
low
resolution

f (t )     a       j ,k (t )
j ,k               jkt(t )  2 j j 2 jtj  k  j,jk k  Z
 jk ( )  2    2 t k               , Z
k     j
Efficient Implementation Using Filter Banks

• The lower resolution coefficients can be calculated
from the higher resolution coefficients by a tree-
structured algorithm called a filter bank.
ψ(2t - k)

f1 (t )   a1,k 1,k (t )
ˆ                                                   a1,k (j=1)
k       j

ψ(t - k)
f 0 (t )   a0,k 0,k (t )
ˆ
k       j
a0,k (j=0)
Multiresolution Conditions of WT
• If a set of functions can be represented by a weighted
sum of ψ(t - k), then a larger set (including the
original) can be represented by a weighted sum of
ψ(2t - k).

f(t)                   ψ(t - k)

ψ(2t - k)
Efficient Representation Using Details
H3=L2+D3

details D3                                     H2=L1+D2
&
L2=H2

details D2
&                                            H1=L0+D1
L1=H1

details D1
&
L0
Efficient Representation Using Details
(cont’d)

representation:            L0 D1 D2 D3

A wavelet representation of a function consists of a coarse
overall approximation together with detail coefficients that
influence the function at various scales.
Example - Haar basis

• Suppose we are given a 1D "image" with a resolution
of 4 pixels:
[9 7 3 5]

• This image can be represented in the Haar basis as
follows:
Example - Haar basis (cont’d)
• Start by averaging the pixels together (pairwise) to get
a new lower resolution image:

• To recover the original four pixels from the two
averaged pixels, store some detail coefficients.
Example - Haar basis (cont’d)
• Repeating this process on the averages gives the full
decomposition:
Example - Haar basis (cont’d)
• The Harr decomposition of the original four-pixel
image is:

• We can reconstruct the original image to a resolution
by adding or subtracting the detail coefficients from
the lower-resolution versions.
Example - Haar basis (cont’d)

Note small magnitude
detail coefficients!
How to compute the details Di ?
Basis functions:               Vj : space spanned by ψ(2jt - k)

ψ(t - k)        V0                          f (t )     a       j ,k (t )
f(t) ϵ Vj              k     j
j ,k

ψ(2t - k)       V1

…

ψ(2jt - k)        Vj

Multiresolution conditions  nested spanned spaces: V j  V j 1
i.e., if f(t) ϵ V j then f(t) ϵ V j+1
How to compute Di ? (cont’d)
• Let Wj be the orthogonal complement of Vj in Vj+1
- i.e., all functions in Vj+1 that are orthogonal to all functions in Vj

Vj+1 = Vj + Wj
How to compute Di ? (cont’d)
• If f(t) ϵ Vj+1, then f(t) can be represented using basis
functions fromVj+1:

f (t )   ck (2 j 1 t  k )
k

•Alternatively, f(t) can be represented using basis functions
from Vj and Wj:
Vj+1 = Vj + Wj

f (t )   ck (2 j t  k )   d j ,k (2 j t  k )
k                      k
How to compute Di ? (cont’d)

Think of Wj as a means to represent the parts of a
function in Vj+1 that cannot be represented in Vj

Vj+1          f (t )   ck (2 j 1 t  k )
k

f (t )   ck (2 j t  k )   d j ,k (2 j t  k )
k                       k

Vj,                        Wj
How to compute Di ? (cont’d)
•   Vj+1 = Vj + Wj - using recursion on Vj:

Vj+1 = Vj-1+Wj-1+Wj = …= V0 + W0 + W1 + W2 + … + Wj

if f(t) ϵ Vj+1 , then:
f (t )   ck (t  k )   d j ,k (2 j t  k )
k                 k    j

V0                  W0, W1, W2, …
basis functions        basis functions
Wavelet expansion (Section 7.2)

• Any given decomposition of a signal into wavelets involves
pair of waveforms (mother wavelets):
encode low                              encode details or
resolution info
φ(t) ψ(t)            high resolution info

• The two shapes are translated and scaled to produce
wavelets (wavelet basis) at different locations and on
different scales.
φ(t-k) ψ(2jt-k)
Wavelet expansion (cont’d)

• f(t) is written as a linear combination of φ(t-k) and
ψ(2jt-k) :

f (t )   ck (t  k )   d j ,k (2 t  k )
j

k                 k    j

scaling function       wavelet function

Note: in Fourier analysis, there are only two possible values
of k ( i.e., 0 and π/2); the values j correspond to different
scales (i.e., frequencies).
Wavelet expansion (cont’d)

• Haar scaling and wavelet functions:

computes average     computes details
1D Haar Wavelets

• Think of a one-pixel image as a function that is constant
over [0,1)

examples:    0           1

• We will denote by V0 the space of all such functions.
1D Haar Wavelets

• Think of a two-pixel image as a function having two
constant pieces over the intervals [0, 1/2) and [1/2,1)

examples:    0    ½     1

• We will denote by V1 the space of all such functions.

• Note that   V0  V1

ϵ V1
1D Haar Wavelets (cont’d)
• V j represents all the 2j-pixel images
• Functions having constant pieces over 2j equal-sized
intervals on [0,1).
width: 1/2j

examples:

ϵ Vj                      ϵ Vj
1D Haar Wavelets (cont’d)

V0, V1, ..., V j define a hierarchical representation
i.e.,   V j  V j 1

V0 low resolution

V1
…
VJ high resolution
1D Haar Wavelets (cont’d)

• Let’s define a basis for V j :
mother scaling function:

0   1
1D Haar Wavelets (cont’d)
1D Haar Wavelets (cont’d)
1D Haar Wavelets (cont’d)
• Suppose W j is the orthogonal complement of V j in V j+1
– i.e., functions in Vj+1 which are orthogonal to Vj
1D Haar Wavelets (cont’d)
• Let’s define a basis ψ ji for W j :
mother wavelet function:

j=1
1D Haar Wavelets (cont’d)

basis for V 1 :

dot product
basis W 1 :                         j=1   is zero!
1D Haar Wavelets (cont’d)

Basis functions ψ ji of W j
form a basis in V j+1
Basis functions φ ji of V j
1D Haar Wavelets (cont’d)

V3 = V2 + W2
1D Haar Wavelets (cont’d)

V2 = V1 + W1
1D Haar Wavelets (cont’d)

V1 = V0 + W0
1D Haar Wavelets (cont’d)

φ(t)

ψ(t)
Example - Haar basis (revisited)
Decompositions of f(x)

φ0,2(x)

V2               φ1,2(x)

φ2,2(x)

φ3,2(x)
i ( x)  i, j ( x)
j
Decompositions of f(x) (cont’d)

φ0,1(x)

V1and W1   φ1,1(x)
V2=V1+W1
ψ0,1(x)

ψ1,1(x)
ij ( x)  i, j ( x)

 ij ( x)   i , j ( x)
Example - Haar basis (revisited)
Decompositions of f(x) (cont’d)

φ0,0(x)

V0 ,W0 and W1              ψ0,0(x)

V2=V1+W1=V0+W0+W1
ψ0,1(x)

ψ1,1(x)
ij ( x)  i, j ( x)

 ij ( x)   i , j ( x)
Example - Haar basis (revisited)
Example
Example (cont’d)
Orthogonality and normalization

• The Haar basis forms an orthogonal basis
– not always true for other wavelet bases
• It can become orthonormal through the following
normalization:

i j ( x)  2 j  (2 j x  i)
 i ( x)  2 y(2 x  i)
j               j       j

since     i , j ( x)  2 j ,  i , j ( x)  2 j
Filter banks (analysis)
• The lower resolution coefficients can be calculated
from the higher resolution coefficients by a tree-
structured algorithm called a filter bank.

down-sampling

h0(-n) is a lowpass filter and h1(-n) is a highpass filter
Example - Haar basis (revisited)

[9 7 3 5]
low-pass,                   high-pass,
down-sampling               down-sampling

(9+7)/2 (3+5)/2      (9-7)/2 (3-5)/2       V1 basis functions
Filter banks (analysis) (cont’d)
Example - Haar basis (revisited)

[9 7 3 5]

low-pass,            high-pass,
down-sampling        down-sampling

V1 basis functions

(8+4)/2    (8-4)/2
Convention for illustrating the 1D
Haar wavelet decomposition
x     x   x   x   x   x   … x   x
average
detail
…

re-arrange:

V1 basis functions

re-arrange:
Convention for illustrating the 1D
Haar wavelet decomposition (cont’d)

average
x   x   x   x   x   x   … x   x    detail

…
Examples of lowpass/highpass
analysis filters

Haar

Daubechies
Filter banks (synthesis)
• The higher resolution coefficients can be calculated from
the lower resolution coefficients using a similar structure.

: up-sampling
between samples)
Filter banks (synthesis) (cont’d)
Examples of lowpass/highpass
synthesis filters

Haar (same as
for analysis):

Daubechies
Computing the 2D Haar Wavelet Transform

• The 2D Haar wavelet decomposition can be computed
using 1D Haar wavelet decompositions (i.e., 2D Haar
wavelet basis is separable).

• Two decompositions
– Standard decomposition
– Non-standard decomposition

• Each decomposition corresponds to a different set of
2D basis functions.
Standard Haar wavelet decomposition

• Steps

(1) Apply the 1D Haar wavelet transform to each row of
pixel values.

(2) Apply the 1D Haar wavelet transform on the columns
of the row-transformed array.
Standard Haar wavelet decomposition
(cont’d)
average
(1) row-wise Haar decomposition:          detail

re-arrange terms
xxx    …      x                     …                     …
xxx    …      x                     …
…        …     .       …          …         .
…      …         .
xxx    ...   x                      …
Standard Haar wavelet decomposition
(cont’d)
average
(1) row-wise Haar decomposition:                  detail

from previous slide:                      row-transformed result

…                                            …
…
…
…          …           .                   …          …       .
…
Standard Haar wavelet decomposition
(cont’d)
average
detail
(2) column-wise Haar decomposition:

row-transformed result                  column-transformed result
…                                           …
…                                           …
…
…         …       .                        …          …       .
…                                           …
Example

row-transformed result
…
…

…           …       .
…
Example (cont’d)

column-transformed result
…
…

…         …       .
…
2D Haar basis for standard decomposition
To construct the standard 2D Haar wavelet basis, consider
all possible outer products of 1D basis functions.

φ0,0(x)

ψ0,0(x)
V2=V0+W0+W1

ψ0,1(x)

ψ1,1(x)
2D Haar basis for standard decomposition
To construct the standard 2D Haar wavelet basis, consider
all possible outer products of 1D basis functions.

φ0,0(x), φ0,0(x)       ψ0,0(x), φ0,0(x)                      ψ0,1(x), φ0,0(x)

ij ( x)  i, j ( x)    ij ( x)   i , j ( x)
2D Haar basis of standard decomposition

V2                             ij ( x)  i, j ( x)

 ij ( x)   i , j ( x)
Non-standard Haar wavelet decomposition

• Alternates between operations on rows and columns.

(1) Perform one step of horizontal pairwise averaging and
differencing in each row.

(2) Apply vertical pairwise averaging and differencing in
each column of the result.

(3) Repeat the process for the quadrant containing averages
in both directions.
Non-standard Haar wavelet decomposition
(cont’d)

pairwise horizontal    pairwise vertical
Haar decomposition:    Haar decomposition:
xxx   …     x              …                        …
xxx   …     x              …                        …
…        …       .       …        …         .
…     …     .
xxx   ...   x              …                        …

Note: averaging/differencing
of detail coefficients shown
as red again!
Non-standard Haar wavelet decomposition
(cont’d)
re-arrange terms
…
…              horizontal pairwise   vertical pairwise
Haar decomposition    Haar decomposition
…       …          .
…
…
…
…
…
…       …     .
…
Example

re-arrange terms
…
…
…       …          .
…
…
Example (cont’d)

…
…

…   …   .
…
2D Haar basis for non-standard decomposition

• Defined through 2D scaling and wavelet functions:

 ( x, y )   ( x) ( y )     0,0 ( x, y )   ( x, y )
0

 ( x, y )   ( x) ( y )      kj, f ( x, y )  2 j  (2 j x  k , 2 j y  f )
 ( x, y )   ( x) ( y )    kj, f ( x, y )  2 j (2 j x  k , 2 j y  f )
 ( x, y )   ( x) ( y )     kj, f ( x, y )  2 j (2 j x  k , 2 j y  f )

Three sets of detail coefficients!
2D Haar basis for non-standard decomposition
(cont’d)

0,0 ( x, y )   ( x, y ) LL: average LL
0
Detail coefficients

LH: intensity variations along
    j
k, f   ( x, y )  2  (2 x  k , 2 y  f )
j       j         j
columns (horizontal edges)

kj, f ( x, y )  2 j (2 j x  k , 2 j y  f )    HL: intensity variations along
rows (vertical edges)
 kj, f ( x, y )  2 j (2 j x  k , 2 j y  f )   HH: intensity variations along
diagonals
2D Haar basis for non-standard decomposition
(cont’d)

ij ( x)  i, j ( x)
V2
 ij ( x)   i , j ( x)
Convention for illustrating
2D wavelet decomposition

LL   LH
LH            HH
HL    HH

LL             HL

The wavelet transform can be applied again on the lowpass-lowpass
version of the image, yielding seven subimages.
Forward/Inverse DWT
(using textbook’s notation)

LL                                 i=H,V,D

H LH
V HL
D HH
Forward 2D DWT (filter bank)

H LH
V HL
D HH
Inverse 2D wavelet transform

• Enlarge the wavelet transform data to its original size
(insert zeros between each value).

• Convolve each of the four subimages with the
corresponding inverse (lowpass, highpass) filters.

• Sum the results to obtain the original image.
Inverse 2D DWT (filter bank)

H LH
V HL
D HH
Applications

• Noise filtering
• Image compression
– Fingerprint compression
• Image fusion
G. Bebis, A. Gyaourova, S. Singh, and I. Pavlidis, "Face
Recognition by Fusing Thermal Infrared and Visible
Imagery", Image and Vision Computing, vol. 24, no. 7, pp.
727-742, 2006.
• Recognition/Classification
• Image retrieval

```
DOCUMENT INFO
Shared By:
Categories:
Stats:
 views: 37 posted: 5/11/2011 language: English pages: 141