DIGITAL IMAGE TRANSFORMS 159

where x and y range from 0to N- 1

T:be in other transforms, the
transformation of image fx, y) can be performed using DST as
F* AfA
where A is the kernel matrIX and the
original can be recovered as
A'FA = AFA
As 4-= A,the forward and inverse sine transforms are same.
DST transform is fast and its
computational complexity is O(n log n). As its energy
compaction is good, DST iS very useful for compressing images.
45 WALSH TRANSFORM
The Walsh transform is given as
1 N-1

x=0 i=0

The term h(x, y) is given as

n

h(z, y) =j-yj%o0)
N =0
h(x, v) is called the kernel of the Walsh transform. Here b{x) represents the ith bit of the
binary representation of x. For example, if n=4and z=8 (the binary representation of x=
1000), then b{x) can be denoted as
box) = 0, b,(r) =0, byl(r) =0, b,() = 1
The inverse Walsh transform is given as
N-1 n-1

x=0 /=0

Here the inverse kernel is given as

h(a, u) =|]-y4A.a)i=0

Ine TD Walsh transform can also be extended to two dimensions. The Walsh
transformation is given as
.y) T(-owb*ho .)
N-I N-1

W(u, v) = re0 y=0 i=0

160 DIGITAL IMAGE PROCESSING

The inverse Walsh transform is given as

n-1

N i=0

Itcan beobserved that the Walsh kernel is separable and symmetric.

g (x, y, u, v) =g(,u) g), v)
= h,(x, u) h,(y, )
n-1
where
g,(*,u) = 1
(,u) =Il-)o) i=0

n-1
and
8,(y, v) = h, (y,) =I-iya0Mo)
VN i=0
For example, the Walsh transform for the 2 x 2 kernel is given as
+
Wax=
and
+
0(+ + +
I|+
W= 2 + +

3+
For x 2 and u =2, the
value of the kernel is given as
x=0010, b, =0, b,=1,b,=0, b4 =0
u=001 0, b, =0, b, = 1, b,=0, b,
=0
N

The same result can 4

be
1. Write the obtaned using the
of x. following shortcut:
binary
2. Write
thebinary representation
representation of u in reverse.
 DIGITAL IMAGE TRANSFORMS 101

3. Find the number of overlapping of ls. If it is Oor cven, the sign is positive. Elsc, it is
negative.
For example, for the aforementioned exanmple
u= 0100 (in reverse)
x= 00 10

By perfoming bit-level comparison, it can be observed that there is no overlapping of

ls among uand x. In other words, the number of overlappings is zero. Therefore the sign
ispositive.

4.6 HADAMARD TRANSFORM

The basis function of the Hadamard transfornm is also +1 and -1, like the Walsh transform.
The kernelof the 1D Hadamard transform is given as

8(r, u)=-)
N

The 1D Hadamard transform H(u) is given as

1 N-1
H)= x=0

where N = 2"
The inverse Hadamard transform f(x) is given as
N-1
f(x) =H(u) (-1)=
r=0

where N = 2" and the term (-1) is the inverse kernel, given by h(r, u). Like the
Walsh transform, the 2D Hadamard transform is defined as
N-I N-]

Hu, v) =N L0 ye0
jr, y(-)"
The inverse Hadamard transform is given as
N- -|

i=0
162 DIGITAL IMAGE PROCESSING

The Hadamard kernel is also separable and symmetric. Hence,

g(x, y, u, v) =g,(x, u) g, y, v)
=h (x, u) h,(y, v)
The Hadamard matrix is symmetrical and orthogonal. Inimage processing,the
transforn differs from the Walsh transform only in the order of the basis
Hadamard transform of an image fis denoted as
Hadamar
functions.The,
g-AxfxA
The matrix A is related to the Hadamard matrix as
1
A= H,
VN
Here N is the dimension of the image.
The basic Hadamard matrix is given as
The Hadamard matrices of higher dimensions can be represented
recursively as
H,, HN HN
H -HN
For example,the basic Hadamard transform for the
matrix H,is given as
(1 1 1
1 -1 1 -1
1 1 -1 -1
1 1 -1
This is represcnted by the
Kronecker product as
In
H,4- H,2 H,2, 2
general,the Hadamard matrix for H2. 20 is given as

The inverse
Hadamard transform is given as
f= A'gA'
The inverse
Hadamard transform exactly produces the
original matrix.
 DIGITAL IMAGE TRANSFORMS 103

hample&2 ethat Haamani transtorm works for the followingimage:

transtimationis given as T AxFxA.

TNNN NNaionis given as

transformed image.
takohsened that the ornginal image is retrieved from the

transform matrix is called sequency.

The number of sign changes in the rows of the
number of sign changes in the rows of the Hadamard kernel is not sequential.
Te
ensure that the number of sign changes is
Tae Hadamard kemel can be modified to reordering the Walsh kernel.
Hadamard kernel can be obtained by
ending. The

UAR TRANSFORM
-1, or 0. The kernel of the Haar
consists of elements+1,
c tasts of the Haar transtorm
Tsform isgenerated using the following procedue:
1. Fndthe order N. Let n = log N.
* Detemine pand q.
(2)p ranges from 0to n - 1.
(b)Ifp =0. then q =0 or q= I.
Else, I sqs
3. The value k is determined as
and Z 0Z, 1/2, 2/2, ... (N - IZ
k= +q- I
4 Ifk=0. then
 DIGITAL IMAGE PROCESSING
164

h,(z) = ho(z)=for Ze (0, 1)

VN
Else
Szl9)/27
|2P if 9-1
h(:)=Ap(2)= 2P 2P

2P 2P
=0 for Z e[0, 1]
For example, let us consider the base for N= 2.
IfN = 2, n= log N=log 2 = 1
So the value ofp =0. Therefore, g takes the value 0 or 1.
Therefore, k=2 +q-1.When p=0 and g =0, the value of k=0. When p=0a
= 1, the value of k= 1.

So when k= 0, 20
ho (0) =

and when k = 1, 20
h(0) =
Therefore the 2 x 2 Haar matrix is given as

A,=
Similarly any higher-order matrix can be
computed. The matrix for N= 4 is given
1
1 -1
A,= -1
V4-V2 -V2
4.8 SLANTTRANSFORM
0
-V2 -V2)
The slant
hence verytransform is uscd
widely is orthogona/s
fast. Its kernel can be image
generated compression. Slant transform is
recursively like Hadamard ttr:ansform.
 DIGITAL IMAGE TRANSFORMS 165
transform of the order 2 x 2is
The slant defined as

Ingeneral, an Nx N matrix can be recursively given as

1
1
by
Sy = 1
l(N/2)-2 0
SNI2
0 SN2)
|-by aN
l(N/2)-2 0

Here by =(1+4aD-1) 2and ay =2xby Xa-1, I is an identity matrix, and 0 is the

matrix of zeros. Let us generate kernel for N=4, AsN= 2", n=2. Let us substitute this in
the equation to get

b, =

1 2
and
a, =2xxl=
Substituting this in the aforementioned recursive equation, one gets
0
A
A, =
b b
-1
|X
0 A 4)
b,- -
1 1
2 2
1 -1 0 0
X
X
-1
0 0 1
0 0 1 -1)
2 2

is given as follows:
Therefore, the final slant transform of the order 4 x 4
 DIGITAL IMAGE PROCESSING -S- -
166
-
3

A,
1 -I -1
1 3 1

The matrnn is real, not symmctric, and unitary.

The fnwardslant transform is given as
F= Af
and the inversc slant transformisgiven as
f= A" xF
-A xF
This logc can be extended to 2D images also. The forward discrete Slant transío
gven as follows:
F= Axfx A"
The inverse discrete slant transform is given as follows:
f= 4' xFxA
The applscation of discrete slant form is exactly similar to previous
transforms
computational complexity of discrete slanttransform is O(N log, M).
49 SD ANDLTRANSFORMS
The sitgular-valuc decomposition (SVD)
epend on the concepts of eigen values andtransform and Karhunen-Loeve (KL) ra
epedent The concepts of eigen values andeigen vectors. These transformatiots
Ltibe a vector of input sarmples andA be eigen vectors are explainedIas follows
a transformation matrix
vec s oy Thas can be that ma}s
represented as
l ete be an ther vector that is
perpendicular to it Let it be
Then

AÀ
The etor A À0
s
vector The calar
termed as egen vector of the ds
value À s
known as latent valuc called the egen valuematr1x A. It is also known
it s also

assoCiated with the