Image Transforms Explained

1
IMAGE TRANSFORMS
By
Dr. Suvendu Rup
CSE Dept. IIIT Bhubaneswar
Image Transforms
2
Image transforms can be simple arithmetic or complex

mathematical operations on images which convert images from
one representation to another.
Image (NxN) Image (NxN)

{XN} Transform
{YN}
Y=AX
Image (NxN) Inverse Image (NxN)

{YN} Transform {XN}
X=A-1Y
Ratnakar Dash,
Image Transforms Assistant Professor, CSE dept. of NIT Rourkela
Applications
3
 Preprocessing
 Filtering
 Enhancement
 Image Compression
 Feature Extraction
 Edge detection
 Corner detection
Ratnakar Dash,
Unitary Matrix
4
 Image transformation represents a given image into series

sum of unitary matrices
 A matrix is called unitary if
−1
A =  
A A *T H
Hermitian conjugate
 Unitary Matrices are basis images
Ratnakar Dash,
Properties of Unitary Transform y =Ax
5
 Energy Conservation
 || y ||2 = || x ||2
◼ || y ||2 = || Ax ||2= (Ax)*T (Ax)= x*T A*T A x = x*T x = || x ||2
 Rotation
 A unitary transformation is a rotation of a vector in an N-
dimension space, i.e., a rotation of basis coordinates
Ratnakar Dash,
Contd..
6
 Energy Compaction
 Many common unitary transforms (like DCT, KLT etc.) tend to

pack a large fraction of signal energy into just a few transform
coefficients
 Decorrelation
 Highly correlated input elements ➔ quite uncorrelated output

coefficients
 Covariance matrix E[( y – E(y) ) ( y – E(y) )*T ]
Ratnakar Dash,
What is orthogonal ?
7
 A set of real valued continuous functions

{an(t)} = {ao(t), a1(t),…} is said to be orthogonal over
(t0, t0+T) if
k if m = n
T am (t ). an (t ) dt = 0 if m  n
if k = 1 then an (t ) is orthonormal
Ratnakar Dash,
Example
8
 {sinωt, sin2ωt}
Ratnakar Dash,
Orthogonal Expansion
9
Any arbitrary signal x(t); {t0, t0+T}

Can be represented by series summation of a set of
orthogonal basis functions.

x(t ) =  cn an (t )
n =0
cn nth coefficient of expansion
Ratnakar Dash,
Contd..
10

x (t ) = c a
n =0
n n (t ), t 0  t  t 0 + T

  x(t )a
T
m (t ) dt =  c a
T n =0
n n (t ) am (t ) dt
= c0  a0 (t ) am (t ) dt +
T
c1  a1 (t ) am (t ) dt + ... +
T
cm  am (t ) am (t ) dt + ...
T
  x(t )a m
(t ) dt = kcm   x(t )am(t )dt = cm, if k = 1
T
T
Orthonormal Set
Ratnakar Dash,
Complete, Closed(orthogonal
11
functions)
There is no signal x(t) with x 2 (t )dt  

 T
such that
 x(t )a (t )dt = 0, n = 0,1,..
T
n
 x(t) with  (t )dt  

x 2
T N −1
xˆ (t ) =  cn an (t )
n =0
Such that  x(t ) − xˆ (t )

2
dt    0
Ratnakar Dash,
Discrete Formulation
12
 Let the set of samples represented by {u(n): 0 ≤ n ≤ N-1}

is a vector of dimension N
 Premultiply u by a unitary matrix A of dimension N*N,

we get another vector
v= Au
Transformed vector
Transformation matrix
Ratnakar Dash,
Contd..
13
In the form of series summation

N −1
v(k ) =  a (k , n)u (n) 0  k  N −1
n =0
where A−1 = A*T

As A is a unitary matrix we can get back u as
N −1
u ( n ) =  a  ( k , n )v ( k ) 0  n  N −1
k =0
The columns ofAT
i.e.,

k 
(a  a (k , n ), 0  n  N − 1 )


T
are called the basis vectors of A
Ratnakar Dash,
Example
14
 Let u0  v0  a00 a01 a02 

U =  u1 
V =  v1  A =  a10 a11 a12 
u2  v2  a20 a21 a22 

 00 10
a a a 20 
A−1 = AT =  a01 a11 a21 

a02 a12 a22 
U = A−1V = AT V
  
 0
u  0   00 
v a  10 
a  20 
a
U =  u1  = A−1  v1  =  a01  v0 +  a11  v1 +  a21  v2
u2  v2  a02  a12  a22 
basis vectors b1 b2 b3
Ratnakar Dash,
Contd..
15
In case of an image, an image can be represented by a set of basis images

Image u (m, n) 0  m, n  N − 1
N −1 N −1
v(k , l ) =  ak ,l (m, n)u (m, n) 0  k , l  N − 1
m =0 n =0
ak ,l (m, n) N × N matrix/ N2 numbers of such matrices
Inverse Transform
N −1 N −1
u (m, n) =  ak,l (m, n)v(k , l ) 0  m, n  N − 1
k =0 l =0
Ratnakar Dash,
Contd..
16
❑ Set of complete orthonormal discrete basis function satisfies the

following properties
ak ,l (m, n) Orthonormality
N −1 N −1
 ak ,l (m, n)a (m, n) =  (k − k ' , l − l ')


k ',l '
m =0 n =0
Completeness
N −1 N −1
 ak ,l (m, n)ak,l (m' , n' ) =  (m − m' , n − n')

k =0 l =0
Ratnakar Dash,
Computational Complexity
17
 To Compute transform coefficient v(k , l )

N −1 N −1
v(k , l ) =  ak ,l (m, n)u (m, n) 0  k , l  N − 1
m=0 n =0
 For each v(k , l ) no. of complex multiplications and additions

~ O(N2), so for all v(k , l )~O(N4) which is quite expensive for
practical size Images.
 How to reduce computation time??
 Using separable unitary transforms
ak ,l (m, n) = ak (m).bl (n)  a(k , m )b(l , n )
Ratnakar Dash,
Separable Transforms
18
where ak (m), k = 0,1,..N − 1 are 1D orthonormal sets of basis vectors

bl (n), l = 0,1,..N − 1
should be unitary matrices themselves
In most cases we choose A & B to be same. A  a(k , m) and B = a(l , n)
N −1 N −1
v(k , l ) =  a (k , m) u (m, n) a (l , n); V = AUAT
m =0 n =0
N −1 N −1
u (m, n) =  a  (k , m) v(k , l ) a  (l , n); U = AT VA
k =0 l =0
V = AU A T
V T = A [ AU ]T
 These are called 2D separable transformation
 The 2D transform can be performed first by transforming each column of U
and then transforming each row of the result to obtain rows of V
complexity is 2N3
Ratnakar Dash,
Concept of Basis Images
19

 Let ak denotes the kth column of AT
 Define the matrices
  T
A =a a
k ,l k l
 Define the inner product of two NxN matrices F and G

N −1 N −1
F , G =  f (m, n )g  (m, n )
m =0 n =0
Ratnakar Dash,
Concept of Basis Images
20
 Then the transformation equation can be written as

N −1 N −1
v(k , l ) =  ak ,l (m, n)u (m, n)  U , Ak,l
m =0 n =0
N −1 N −1
u (m, n) =  ak,l (m, n)v(k , l )
k =0 l =0
N −1 N −1
 U = u (m, n) =  v(k , l ) Ak,l
k =0 l =0
 U is represented by linear combination of N2 Matrices.

 These matrices are called basis images
Ak,l k , l = 0,1, 2..N − 1
Ratnakar Dash,
Example
21
1 1 1 1 2
A=  
U =  
2 1 − 1  3 4 
1 1 1  1 2  1 1
Transformed Image V =    
2 1 − 1  3 4  1 − 1
1 4 6 1 1
=   
2 − 2 − 2 1 − 1
 5 − 1
=  
− 2 0 
To get basis images, take outer product of the columns of AT

1  1 1  1 1
 ( )

A0,0 = 1 1 =  
2  1 2  1 1
 1  1 −1 T
A0,1 =   = A1,0
2  1 −1
 1 1 −1
A =  
2  −1
1,1
1 
Ratnakar Dash,
Contd..
22
 Inverse Transform gives
T 1 1 − 1 5 − 11
 1
A VA =    
2 1 − 1 − 2 0 1 − 1
1  3 − 11 − 1  1 2
=    =  
2  7 − 11 − 1  3 4
= U  Original Image
Ratnakar Dash,
Some Transformation Techniques
23
 Discrete Fourier Transform

 Discrete Cosine Transform
 Karhunen-Loeve Transform
 Haar Transform
 Walsh Transform
 Hadamard Transform
Ratnakar Dash,
Fourier Transform (1D)
24
Continuous Fourier Transform (CFT)


W ( f ) = Fw(t ) =  w(t ) e − j 2ft
dt
−
W( f ) = X( f )+ j Y( f )
W ( f ) = W ( f ) e j ( f ) Frequency, [Hz]
Phase
Amplitude Spectrum
Spectrum
Inverse Fourier Transform (IFT)


w (t) = F W (f ) =  W (f ) e + j2ft df
-1
−
Ratnakar Dash,
Discrete Fourier Transform (1D)
25
Equal time intervals

◼ Discrete Domains
❑ Discrete Time: n = 0, 1, 2, 3, …………, N-1
❑ Discrete Frequency: k = 0, 1, 2, 3, …………, N-1
Equal frequency intervals
◼ Discrete Fourier Transform

N −1  2 
− j  nk
X [ K ] =  x[n] e  N 
; K = 0, 1, 2,….., N-1
n =0
◼ Inverse DFT
N −1  2 
1 j  nk
x[n] =
N
 X [K ] e
K =0
 N 
; n = 0, 1, 2,….., N-1
Ratnakar Dash,
2D – Discrete Fourier Transform
26
▪ Image: f(x, y), 0≤x≤N-1, 0≤y≤N-1

 xu yv 
− j 2  + 
▪ Transform kernel: g ( x, y, u , v) = e  N N 
▪ Discrete Fourier Transform of f(x, y)is

N −1 N −1  ux vy 
− j 2  + 
 f (x, y )e
1
F (u , v) = N N
N x =0 y =0
N −1 N −1  ux vy 
j 2  + 
f ( x, y ) =
1
N
 F (u, v)e
u =0 v =0
 N N 
Ratnakar Dash,
Contd..
27
In general, Transform kernel can be represented as:

2 2
−j ux + vy
−j (ux + vy )
wN = e N
w N =e N
Discrete Fourier Transform of f(x, y) can be

represented as
N −1 N −1
f ( x, y )wNux + vy
1
F (u , v) =
N
 x =0 y =0
N −1 N −1
f ( x, y ) =
1
N
 F (u ,
u =0 v =0
v ) wN
− (ux + vy )
Ratnakar Dash,
Contd..
28
Let R(u,v), I(u,v) be real part and imaginary part of

F(u,v)
Modulus： (
F (u , v) = R (u , v ) + I (u , v )
2 2
)12
 I (u , v ) 
Phase：  (u , v) = tan −1  
 R (u , v ) 
Power Spectrum： E (u , v) = R 2 (u , v ) + I 2 (u , v )
Ratnakar Dash,
Amplitude and Phase
29
F (x)
amplitude
phase
original
F (x)
Ratnakar Dash,
Images and their spectrums
30
Image domain Frequency domain Image domain Frequency domain
Ratnakar Dash,
Properties of 2-D Fourier Transform
31
• Separability
• Linearity
• Scaling
• Shift Theorem
• Rotation
• Periodicity and conjugation
• Convolution
• Correlation
• Average
Ratnakar Dash,
Separability
32
❑ The implementation steps for the two-dimensional DFT

may be visualized as shown in the diagram below
Row Column
Transforms Transforms
f(x,y) F(x,v) F(u,v)

Multiply by N
Ratnakar Dash,
Separability (contd. )
33
f(x,y) F(u,v)
f(x,y) F(x,v) F(u,v)
Ratnakar Dash,
Linearity
34
❑ DFT is a linear operator
F[ a f1(x,y)+b f2(x,y)] = aF1(u,v)+bF2(u,v)
Ratnakar Dash,
Linearity (contd..)
35
f1(x,y) f2(x,y) 2f1(x,y)+3f2(x,y)
F1(u,v) F2(u,v) Phase Spectrum (2F1(u,v)+3F2(u,v))
Ratnakar Dash,
Scaling
36
1 u v
f (ax, by ) ⎯⎯
⎯→ DFT
F , 
ab  a b 
Ratnakar Dash,
Scaling (contd..)
37
f(x,y) F(u,v)
f(2x,3y) F(u,v)
Ratnakar Dash,
Shift Theorem
38
f ( x − x0 , y − y0 ) F (u, v ) exp[ − j 2 (ux0 + vy0 ) / N ]
f ( x , y ) exp[ j 2 (u0 x + v0 y ) / N ] F ( u − u0 , v − v 0 )
Ratnakar Dash,
Shift Theorem (contd..)
39
Original Image Spectrum Image

Mesh Plot
Tx = 50, Ty =50 Spectrum Image

Mesh Plot
Ratnakar Dash,
Rotation
40
In polar coordinates, f (x,y) and F(u,v) can be represented by f(r,

θ) and F(w,φ) alternatively. Then
f (r, θ+ θ0) F(w,φ + θ0)
(a) a sample image (b) its spectrum (c) rotated image (d) resulting spectrum
Ratnakar Dash,
Periodicity and conjugation
41
Periodicity ：
F(u,v) = F(u+aN,v+bN)
f (x,y) = f (x+aN,y+bN)
Where a, b = 0, ±1, ± 2,……
Conjugation：
F(u,v) = F*(-u,-v)
|F(u,v)| = |F(-u,-v)|
Ratnakar Dash,
Convolution
42
2-D convolution is defined as:

M −1 N −1
1
f e ( x, y )  g e ( x, y ) =
MN
 f
m =0 n =0
e (m, n) g e ( x − m, y − n)
Then: f(x, y)*g(x, y) F(u, v)G(u, v)
f(x, y)g(x, y)F(u, v)*G(u, v)
Ratnakar Dash,
Convolution (contd..)
43
Original image Spectrum
Convolved image(convolved with mean filter) Spectrum

Ratnakar Dash,
Correlation
44
2-D correlation is defined as：

M −1 N −1
1
f e ( x, y)  ge ( x, y) =
MN
  e (m, n) ge ( x + m, y + n)
f *
m= 0 n = 0
Correlation Theorems：
f ( x , y )  g ( x , y )  F * ( u, v ) G ( u, v )
f ( x , y ) g ( x , y )  F ( u, v )  G ( u, v )
*
Ratnakar Dash,
Average
45
Average of 2D image f (x, y) is defined as：

_ N −1 N −1
1
f ( x, y) = 2
N
  f ( x, y)
x =0 y =0
Let u=v=0 , then:

N −1 N −1
1
F (0,0) =
N
  f ( x, y)
x =0 y =0
_
1
f ( x, y) = F (0,0)
N
Ratnakar Dash,
Contd..
< < image
and its
spectrum
46
Fast Fourier Transform
47
recursive algorithm
• decimation in time = odd even in freq. domain

• decimation in freq. domain = odd even in time
N → N log N
4 2
Ratnakar Dash,
Visualizing the DFT
48
 Consider the power spectrum of the 1D square wave

350
300
250
200
150
100
50
0
0 1 2 3 4 5 6 7
Ratnakar Dash,
Visualizing the DFT(circular shifting)
49
 The FT is centered about the origin

 But the DFT is centered about N/2
 We need to correct with a circular shift operation
 Or, multiply by (-1)k prior to taking the transform
350
300
250
200
150
100
50
0
0 1 2 3 4 5 6 7
Ratnakar Dash,
DFT basis images
50
Ratnakar Dash,
Ratnakar Dash,
Ratnakar Dash,
Ratnakar Dash,
Ratnakar Dash,
Ratnakar Dash,
Ratnakar Dash,
Example
58
img=imread('lena.bmp','bmp');
subplot(121);imshow(img);
title('original image ')
fimg=fftshift(fft2(img));
subplot(122); imshow(abs(fimg)/10000)
title(' transformed image ')
original image transformed image
Ratnakar Dash,
Discrete Cosine Transform (DCT)
59
❑ DCT expresses a sequence of finite data points in
terms of a sum of cosine functions oscillating at
different frequencies.
❑ These cosine functions are treated as its basis
function.
❑ Fast algorithm exists.
❑ Most popular in image compression application
because of its energy compaction property.
❑ Adopted in JPEG.
❑ The periodicity implied by DCT implies that it
causes less blocking effect than DFT.
❑ Can be implemented by 2n points FFT.
Contd..
60
Transform kernel of two-dimensional DCT is
 (2 x + 1)u   (2 y + 1)v 
C ( x, y, u, v) =  (u )  (v)  cos    cos  2N 
 2N   
where
1 2
 (0) = ,  (i) = , 1  i  N
N N
Clearly, the kernel for the DCT is both separable and
symmetric, and hence the DCT may be implemented as a
series of one dimensional DCTs.
Ratnakar Dash,
Contd..
61
Forward transform
N −1 N −1
  (2 x + 1)u    (2 y + 1)v 
F (u, v) =  (u ) (v) f ( x, y) cos  cos 
x =0 y =0  2N   2N 
Inverse Transform
N −1 N −1
  (2 x + 1)u    (2 y + 1)v 
f ( x, y) =   (u ) (v) F (u, v) cos  cos 
u =0 v =0  2 N   2 N 
where u, v, x, y = 0, 1, 2, …, N-1
Forward and inverse transformations are same
Ratnakar Dash,
62
 clc;
 clear
 N=8
 x=(0:N-1)';
 C=cos((2*x+1)*x'*pi/(2*N))*sqrt(2/N);
 C(:,1)=C(:,1)/sqrt(2);
Ratnakar Dash,
Kernel matrix to generate DCT
63
Basis Images
0.353553 0.490393 0.46194 0.415735 0.353553 0.277785 0.191342 0.097545
0.353553 0.415735 0.191342 -0.09755 -0.35355 -0.49039 -0.46194 -0.27779
0.353553 0.277785 -0.19134 -0.49039 -0.35355 0.097545 0.46194 0.415735
0.353553 0.097545 -0.46194 -0.27779 0.353553 0.415735 -0.19134 -0.49039
0.353553 -0.09755 -0.46194 0.277785 0.353553 -0.41573 -0.19134 0.490393
0.353553 -0.27779 -0.19134 0.490393 -0.35355 -0.09755 0.46194 -0.41573
0.353553 -0.41573 0.191342 0.097545 -0.35355 0.490393 -0.46194 0.277785
0.353553 -0.49039 0.46194 -0.41573 0.353553 -0.27779 0.191342 -0.09755
Ratnakar Dash,
Basis images of an 8x8 DCT
64
Ratnakar Dash,
Properties of DCT
65
Separability
2D DCT/ IDCT expression shows that they are separable.
So they can be implemented as two 1-D DCT/IDCT
Fast DCT
FDCT is possible in the same manner as FFT
Periodicity
Magnitude of DCT coefficient is periodic with period 2N
Example
66
Original Image DCT Coefficients
DCT Coefficients
(Zoomed) Mesh Plot
Ratnakar Dash,
Image Reconstruction in DCT
67
Original image Reconstructed image
Using only 1
DCT coefficient
Using 3 DCT
coefficients
Coefficients are selected
out of 8x8 window
Ratnakar Dash,
Contd..
68
Using 6 DCT Using 15 DCT

coefficients coefficients
Using 10 DCT Using 64 DCT

coefficients coefficients
Ratnakar Dash,
Karhunen-Loeve Transform
69
 K-L transform is a linear combination of orthogonal

functions of ensembles of input images.
 Basis functions are image dependent
 No fast algorithm exists
 Usually used for comparison
 Transform kernel is not fixed unlike other transforms
 Based on second order statistical properties of image
 Also called Hotelling transform or method of principal

components
Ratnakar Dash,
Contd..
 x1 
70
x 
 Covariance matrix of a population of vectors X =  2
: 

C x = E ( X −  x )( X −  x )
T
  
 xn 
M
1
x =
M
X
k =1
k
C x  real & symmetric
So, Orthogonal vectors of Cx can be calculated Ratnakar Dash,

Contd..
71
❑ Calculate eigenvectors and eigen values of Cx

❑ Sort eigen vectors depending on decreasing values of
eigen values
❑ Select some of the eigenvectors
❑ These eigenvectors are considered as transformed matrix
❑ Calculate the transformed images
Y = A( X −  x )
❑ Y is called as principal component coefficients.

❑This transformation is called K-L transform
Ratnakar Dash,
Properties of Y
72
y = 0
C y = AC x A T
Ratnakar Dash,
Example
73
Input matrix
  3  4  4  4  5  5  5  6 
x =          
  4  3  4  5  4  5  6  5 

Mean Vector
 4.5 
x = 



 4.5 
Ratnakar Dash,
Contd..
74
 0.75 0.375 
CX =
 0.375 
 0.75 
Compute Eigen values
(0.75 −  ) 2 = (0.375) 2
 0.75 −  = 0.375
  = 0.75  0.375
1 = 1.125
2 = 0.375
Ratnakar Dash,
Contd..
75
We get two Eigen vectors

1 1
e = 
1
1
2  
1 
1 
e =
2  − 1

2  
Transform Matrix
→ A= 1 1 1 
1 − 1
2  
Ratnakar Dash,
Contd.
76
 Reconstructed
X = A Y + X
T
 K no of Eigen vectors
Y = Ak ( X −  X )
ˆ
X = A Y +
T
k X
Ak = k  n
Y = k 1
X = n 1 Ratnakar Dash,
Properties of K-L Transform
77
 Decorrelation
 E[ y yH ]= E[ (UH x) (UH x)H ]= UH E[ x xH ] U = diag{1, 2, … , N}
 Note: Other matrices (unitary or nonunitary) may also

decorrelate the transformed sequence
 Minimizing MSE under basis restriction

 If only allow to keep m coefficients for any 1 m N, what’s
the best way to minimize reconstruction error?
→ Keep the coefficients w.r.t. the eigenvectors of the first m
largest Eigen values
Ratnakar Dash,
Principal Component Coefficients
78
Lena image Energy compaction in lena image
Ratnakar Dash,
Image Reconstruction in K-L Transform
79
1 PC 10 PCs 20 PCs
Original image
50 PCs All PCs
Ratnakar Dash,
Haar Wavelet Transform
 One dimensional transformation on each row followed by one

dimensional transformation of each column.
 Extracted coefficients would be

 Approximation
 Vertical
 Horizontal
 Diagonal
 Approximation coefficients are further decomposed into the next level
 4 level decomposition is used
Ratnakar Dash,
Contd..
Graphical Representation - Wavelet decomposition

(level = 2)
Ratnakar Dash,
Approach
 For a 2X2 matrix
a b a + b a − b
x =   x =  
c d  c + d c − d 
1 a +b+c + d a −b +c −d 
y=  
2 a +b−c −d a −b−c + d 
Ratnakar Dash,
Example
2 2 9 5 4 14 0 4
2 0 5 2 Column wise Summation 2 7 2 3
6 7 4 7 13 11 -1 -3
3 4 8 8 7 16 -1 0
Row wise Summation
Approximation Horizontal
3 10.5 1 3.5 6 21 2 7
Finding Average
10 13.5 -1 -1.5 20 27 -2 -3
1 3.5 -1 0.5 2 7 -2 1
3 -2.5 0 -1.5 6 -5 0 -3
Vertical Diagonal
Properties of Haar transform
84
 The Haar transform is real and orthogonal. Therefore,

Hr=Hr*
 The Haar transform is a very fast transform.
 On an Nx1 vector it can be implemented in O(N) operations
 The Haar transform has poor energy compaction for Images
Ratnakar Dash,
85
 %Initialization
 H=[1]; NC=1/sqrt(2);%normalization constant
LP=[1 1]; HP=[1 -1]; % iteration from H=[1]
 for i=1:Level
H=NC*[kron(H,LP);kron(eye(size(H)),HP)];
 end
 HaarTransformationMatrix=H;
Ratnakar Dash,
86
Ratnakar Dash,
Basis images of Haar Transform
87
Ratnakar Dash,
1-step Haar Decomposition
88
Original image Haar transformed image
Ratnakar Dash,
2-step Haar Decomposition
89
Ratnakar Dash,
Discrete Walsh Transform (1-D)
n −1 bi ( x ) bn−1−i ( u )
1
g ( x, u ) =  (−1)
 N =0
i 

1− D Forward Kernel
where,
N → No. of Samples
n → No. of bits needed to represent x and u
b k ( z ) → k th bit in digital representation of z
N −1 n −1 bi ( x ) bn−1−i ( u )
1
W (u ) =  f ( x) (−1)
 N x =0
i =0

1− D Forward Transform Ratnakar Dash,
Contd..
n −1 bi ( x ) bn−1−i ( u )
h( x, u ) =  (−1)
=0 Difference
i  lies in
1− D Inverse Kernel Multiplicative
bi ( x ) bn−1−i ( u ) Factor
N −1 n −1
f ( x) =  W (u ) (−1)
x =0
 i =
 0


1− D Inverse Transform
Ratnakar Dash,
Discrete Walsh Transform (2-D)
n −1 {bi ( x ) bn−1−i ( u ) + bi ( y ) bn−1−i ( v )}

1
g ( x, y, u, v) =  (−1)
Ni =0


2 − D Forward Kernel
N −1 N −1 n −1 {bi ( x ) bn−1−i ( u ) + bi ( y ) bn−1−i ( v )}

1
W (u, v) =   f ( x, y ) (−1)
N y =0 x =0
i = 0

2 − D Forward Transform
Ratnakar Dash,
Contd..
n −1 {bi ( x ) bn−1−i ( u ) + bi ( y ) bn−1−i ( v )}

1
h( x, y, u, v) =  (−1)
Ni =0


2 − D Inverse Kernel
N −1 N −1 n −1 {bi ( x ) bn−1−i ( u ) + bi ( y ) bn−1−i ( v )}

1
f ( x, y ) =   W (u, v) (−1)
N y =0 x =0
i = 0

2 − D Inverse Transform
Ratnakar Dash,
Property
• Separable
1
W (u ) = [Weven (u ) + Wodd (u )]
2
1
W (u + M ) = [Weven (u ) − Wodd (u )]
2
where,
N
u = 0,1,..., − 1
2
N
M=
2
Ratnakar Dash,
Basis images of Walsh Transform
95
Ratnakar Dash,
Hadamard Transfrom
n−1
1 {bi ( x )bi (u ) +bi ( y )bi ( v )}

g ( x, y, u, v) = (−1) i=0
N  
2 D Forward Kernel
n−1
1 {bi ( x )bi (u ) +bi ( y )bi ( v )}

h( x, y, u, v) = (−1) i=0
N  
2 D Inverse Kernel
Ratnakar Dash,
Hadamard Matrix
n−1
1  bi ( x )bi (u )
g ( x, u ) = (−1) i=0
N   
Hadamard Matrix
u
0(0) 1(7) 2(3) 3(4) 4(1) 5(6) 6(2) 7(5)
0 + + + + + + + +
1 + - + - + - + - sign changes
are not
2 + + - - + + - -
ordered (shown
3 + - - + + - - + in backet)
x
4 + + + + - - - -
5 + - + - - + - +
6 + + - - - - + +
7 + - - + - + + - Ratnakar Dash,
Hadamard transform
98
 Transform matrices can be recursively generated
1 1 
H2 =  
 1 −1 
H n = H n −1  H1
H 3 = H 2  H1 , H 2 = H1  H1
Note: Hadamard Coefficients need reordering to concentrate

energy
Ratnakar Dash,
Contd..
 N=2
1 1 
H2 =  
1 − 1
H N HN 
=
H −N 
H 2N
H N
Ratnakar Dash,
Ordered Hadamard Transformation
n −1
1  bi ( x ) pi ( u )
g ( x, u ) = (−1) i=0
N
where,
p0 (u ) = bn −1 (u )
p1 (u ) = bn −1 (u ) + bn − 2 (u ) bi (u) is changed to pi(u)
p2 (u ) = bn − 2 (u ) + bn −3 (u )
.
.
.
pn −1 (u ) =Image u ) + b0 (u )
b1 (Transforms Ratnakar Dash,
Assistant Professor, CSE dept. of NIT Rourkela
Contd..
N=8
0 1 2 3 4 5 6 7
0 + + + + + + + +
1 + + + + - - - -
2 + + - - - - + +
3 + + - - + + - -
x
4 + - - + + - - +
5 + - - + - + + -
6 + - + - - + - +
7 + - + - + - + -
Ratnakar Dash,
Basis images of Hadamard Transform
102
Ratnakar Dash,
Properties of Hadamard Transform
103
 Unlike Other Transform, the elements of the basis vectors of

the Hadamard transform take only the binary values +1 , -1,
therefore well suited for DSP.
 Hadamard transform H is real, symmetric, and orthogonal
 Hadamard transform is a fast transform. The one dimensional
transform can be implemented in O(N log2 N) additions and
subtractions
Ratnakar Dash,
Similarity between Walsh and Hadamard Transform
104
 Basis images of Walsh and Hadamard transforms

are same.
 So Hadamard transform is also known as Walsh
transform or Hadamard-Walsh transform.
Ratnakar Dash,
Image Reconstruction
105
Ratnakar Dash,
Image Reconstruction
106
Ratnakar Dash,
107

Image Transforms Explained

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Image Transforms Explained

Uploaded by

Copyright:

Available Formats

1

Image transforms can be simple arithmetic or complex

Image (NxN) Image (NxN)

Image (NxN) Inverse Image (NxN)

 Image transformation represents a given image into series

 A matrix is called unitary if

 Unitary Matrices are basis images

◼ || y ||2 = || Ax ||2= (Ax)*T (Ax)= x*T A*T A x = x*T x = || x ||2

 Many common unitary transforms (like DCT, KLT etc.) tend to

 Highly correlated input elements ➔ quite uncorrelated output

 A set of real valued continuous functions

Any arbitrary signal x(t); {t0, t0+T}

cn nth coefficient of expansion

 x(t) with  (t )dt  

Such that  x(t ) − xˆ (t )

 Let the set of samples represented by {u(n): 0 ≤ n ≤ N-1}

 Premultiply u by a unitary matrix A of dimension N*N,

In the form of series summation

where A−1 = A*T

The columns ofAT

 Let u0  v0  a00 a01 a02 

A−1 = AT =  a01 a11 a21 

In case of an image, an image can be represented by a set of basis images

ak ,l (m, n) N × N matrix/ N2 numbers of such matrices

❑ Set of complete orthonormal discrete basis function satisfies the

 ak ,l (m, n)a (m, n) =  (k − k ' , l − l ')

 ak ,l (m, n)ak,l (m' , n' ) =  (m − m' , n − n')

 To Compute transform coefficient v(k , l )

 For each v(k , l ) no. of complex multiplications and additions

 Using separable unitary transforms

ak ,l (m, n) = ak (m).bl (n)  a(k , m )b(l , n )

where ak (m), k = 0,1,..N − 1 are 1D orthonormal sets of basis vectors

 Define the inner product of two NxN matrices F and G

 Then the transformation equation can be written as

 U is represented by linear combination of N2 Matrices.

To get basis images, take outer product of the columns of AT

 Inverse Transform gives

 Discrete Fourier Transform

Continuous Fourier Transform (CFT)

Inverse Fourier Transform (IFT)

Equal time intervals

◼ Discrete Fourier Transform

▪ Image: f(x, y), 0≤x≤N-1, 0≤y≤N-1

▪ Discrete Fourier Transform of f(x, y)is

In general, Transform kernel can be represented as:

Discrete Fourier Transform of f(x, y) can be

Let R(u,v), I(u,v) be real part and imaginary part of

Image domain Frequency domain Image domain Frequency domain

❑ The implementation steps for the two-dimensional DFT

f(x,y) F(x,v) F(u,v)

f(x,y) F(x,v) F(u,v)

❑ DFT is a linear operator

F[ a f1(x,y)+b f2(x,y)] = aF1(u,v)+bF2(u,v)

f1(x,y) f2(x,y) 2f1(x,y)+3f2(x,y)

F1(u,v) F2(u,v) Phase Spectrum (2F1(u,v)+3F2(u,v))

f ( x − x0 , y − y0 ) F (u, v ) exp[ − j 2 (ux0 + vy0 ) / N ]

Original Image Spectrum Image

Tx = 50, Ty =50 Spectrum Image

In polar coordinates, f (x,y) and F(u,v) can be represented by f(r,

Where a, b = 0, ±1, ± 2,……

2-D convolution is defined as:

Then: f(x, y)*g(x, y) F(u, v)G(u, v)

◼ || y ||2 = || Ax ||2= (Ax)T (Ax)= xT AT A x = xT x = || x ||2