Image Processing /Chapter 3

Chapter 3
Image Transformations
Introduction
Two dimensional unitary transforms play an important role in image processing. The
term image transform refers to a class of unitary matrices used for representation of
images.
UNITARY MATRICES and UNITARY (ORTHOGONAL) TRANSFORM
A square matrix = [ , , , . . ] ( for the ith column vector of A) is unitary if

its inverse is equal to its conjugate transpose, i.e., .

In particular, if a unitary matrix is real , then and it is orthogonal.

Both the column and row vectors ( , = 1 ,2 ,3 . . , ) of a unitary matrix are
orthogonal (perpendicular to each other)
other) and normalized (of unit length), or
orthonormal,, i.e., their inner product satisfies:
1 =
( , ) = = , =
0
These n orthonormal vectors can be used as the basis vectors of the n
n-dimensional
vector space.
Any unitary (orthogonal) matrix A can define a unitary (orthogonal) transform of a vector
= [ , , , . . ]

= = ( )

= ( )

1D Unitary Transforms:
represent by a vector of
A one dimensional sequence u(n) where n=0,1,2,3, N-1, is represented
size N.
A unitary transformation of this 1 dimensional discrete sequence is given by
=
Where A is a unitary matrix and this can be represented expanded in the form

for (3.1)

The transformed vector is the representation of the Transformation coefficients v(k)

which is the transform of u(n).
The matrix A is called forward transformation kernel
kernel.
The transformation is said to be unitary if matrix A is unitary.

The matrix A will be said to be a unitary matrix if it obeys the relation that .
So, this says that A into A conjugate transpose, that should be same as A conjugate
1

transpose A which will be same as an identity matrix.

Page

= = = where I is the identity matrix .

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

transform: =
This implies the inverse transform =
and is given by

for (3.2)

Equation (3.2)
2) can be viewed as a series representation of sequence u(n) .

or is the inverse transformation kernel.

The columns of i.e., the vectors

are called the "basis vectors" of A.

The series coefficients v(k) give a representation of original sequence u(n).
The u(n)
n) is now represented in the form of series summation of a set of vectors or
orthonormal basis vectors.
Two dimensional Orthogonal and Unitary transforms:
As applied to image processing, a general orthogonal series expansion for an N x N
image is a pair of transformations.
transformations
For a 2 dimensional image u(m,
u n) the transformed image v (k,l) using the unitary
matrix A is written in the expanded form as,

(3.3)

The inverse transformation to get the original 2 dimensional matrix from the
transformation coefficient matrix can be written as.

(3.4)

image transform ak,l(m,n) is a set of complete orthonormal discrete basis

functions(defined
(defined in later section).
section)
Separable Unitary Transforms:
The number of multiplications and additions required to compute transform coefficients

in equation(3.3) is . This is too large for practical size images.

separable, if ak,l(m,n) can be represented in the form
If the transform is referred to be separable

ak,l(m,n) ak(n)bl(n)
where {ak(m) k=0,1,N-1} and { bl(n) , l=0,1N-1} are 1-D
D complete orthogonal sets of
basis vectors.
On imposition of completeness and orthonormality properties we can show that
2
Page

and are unitary matrices.

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

i.e = = and = =

Often one chooses same as , so that

and = (4.1.5)

(4.1.6)
and =

Eqn (4.1.5) can be written as = .

Eqn (4.1.5) can be performed by first transforming each column of and then

transforming each row of the result to obtain rows of .

Basis Images:

Let denote column of . Let us define the matrices and

matrix inner product of two matrices and as

=
Then equ (6) and (5) give a series representation, for the images as ,

= and =

Any image can be expressed as linear combination of matrices. for

k,l=0,1,2,.... N-1
1 are called "basis images".

Therefore any image can be expanded in a series using a complete set of basis
images.

Example: Let = ; =

Transformed image = = and Basis images are found as outer

3

product of columns of i.e

Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

= =

= =

= =

= =

The inverse transformation = = =

The output image is now represented by summation basis images multiplied by
corresponding input.

v = u(0,0) x + u(0,1) x + u(1,0) x + u(1,1) x

1x + 2x + 3x + 4x

Properties of Unitary transforms :

1) Energy conservation :

In the unitary transformation, = ,

=
Proof

= = = .
This implies, unitary transformation preserves signal energy or equivalently the length

of vector in dimensional vector space. That is , every unitary transformation is

simply a rotation of in dimensional vector space. Alternatively , a unitary

transform is a rotation of basis coordinates and components of are projections of

on the new basis. Similarly , for 2D unitary transformations, it can be proved that

=
4
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

Example: Consider the vector = and =

This transformation = can be written as = ; = where , new

basis vectors,denote the columns of and , are projections of in the new

coordinate system.

= = =
2) Energy Compaction Property:
Most unitary transforms have a tendency to pack a large fraction of average energy of an
image into relatively few transform coefficients. Since total energy is preserved this

implies that many transform coefficients will contain very little energy. If and

denote the mean and covariance of vector then corresponding quantities for are :

= =

And
=

=
Variances of the transform coefficients are given by the diagonal elements of

i.e. =

Since is unitary , it implies:

= = =
5

and
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

= =

The average energy of transform coefficients tends to be unevenly

distributed, although it may be evenly distributed for input sequence .For a 2D

random field , with mean and covariance , its transform

coefficients satisfy the properties ,

=
and

= =

If covariance of is separable i.e

=
Then variances of transform coefficients can be written as a separable product,

where

= ;

Properties of Unitary transforms :(contd.)

3) Decorrelation :
When input vector elements are highly correlated , the transform coefficients tend to be

uncorrelated. That is , the off-diagonal terms of covariance matrix tend to be small

compared to diagonal elements.
4) Other properties : (i) The determinant and eigenvalues of a unitary matrix have
unity magnitudte.
(ii) Entropy of a random vector is observed under unitary transformation.Since entropy
is a measure of average information of the random vector, this means information is
preserved under unitary transformation.
6
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3
Example :

Given the entropy of an Gaussian random vector with mean and covariance

, as :

To show is invariant under any unitary transformation.

Let =

= =

Use the definition of , we have

= =
Now

=
Hence

= is invariant under any unitary transformation.

Example (Energy compaction and decorrelation ):

Consider zero mean vector that is unitarily transformed to,

where and represents the correlation between and . It is

required to find the energy of the transformed coefficients.
Given
7
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3
We have

From the matrix

Total average energy of 2 is distributed equally between and .

But from we note , and .

The total average energy is 2, but average energy in is greater than that in .

If , then of total average energy is packed in first sample.

Correlation between and is

Now

It follows that:

that is, is less in absolute value than for .

For we find
correlation between transform coefficients has reduced.
8
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

Fourier Representations
Any periodic function can be expressed as the sum of sines and/or cosines of different
frequencies, each multiplied by a different coefficient (we now call this sum a Fourier
series). Functions that are not periodic (but whose area under the curve is finite) can be
expressed as the integral of sines and/or cosine
cosines multiplied by a weighing function. The
formulation in this case is the Fourier transform.
A function, expressed in either a Fourier series or transform, can be reconstructed
(recovered) completely via an inverse process, with no loss of information.
This is one of the most important characteristics of these representations because it
allows us to work in the "Fourier domain" and then return to the original domain of the
function without losing any information.
Fourier Transform
Let f (x) be a continuous function
unction of a real variable x. The Fourier transform of f (x),
denoted as F { f (x) } or F(u), is defined by the equation

where
The reverse transformation is defined as

The Fourier transform of a real function is generally complex, that is,

F(u) = R(u) +j I(u)
where R(u) and I(u), respectively, are the real and imaginary components of F(u).
The F(u) can be expressed in exponent form as.

F(u) is called Fourier spectrum of f (x) and (u)

(u) is its phase angle. The variable u
is called the frequency variable.
varia
Fourier transform can be easily extended to the 2D situation, where the above
relations become:
Example : Find the Fourier transform of the function in Fig. (a)
9

Fig. (a) Time domain signal Fig. (b) Fourier Transform Fig. (c) Fourier Spectrum (Magnitude of FT)
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3
The fourier transform of f(t) is

The result in the last step of the preceding expression is known as the sinc function.

where sinc(0) == 1, and sinc(m) == 0 for all other integer values of m.

Fourier transform contains complex terms, and the magnitude of the transform (a real
quantity), is called the Fourier spectrum or the frequency spectrum is

Figure (c) shows a plot of |F(u)| as a function of frequency.

The 2-D
D Continuous Fourier Transform Pair
Let f(u,v) be a continuous function of two continuous variables, u and v
v.
The two-dimensional,
dimensional, continuous Fourier transform pair is given by the expressions

Where u and v are the frequency variables. When referring to images x and y are
interpreted to be continuous spatial variables. As in the I-D
D case, the domain of the
variables u and v defines the continuous frequency domain.
Figure given below shows a 2-D
2 D function (analogous to the 1
1-D case in previous
Example: Compute 2D Fourier Transform.
10
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

The spectrum is

Let us consider a continuous function as shown in Figure .

The function is discretized at regular intervals, by taking N samples, x units apart as

shown in Figure .Then the function f(x) after discretization can be written as f(x) = f(xo
+ xx) where x takes values 0, 1,2,..., N-1.
The sequences f(0), f(1),..., f(N -1) represents N uniformly spaced samples of the
continuous function.
The 1-dimensional DFT pair of the sampled function is given by

The values u = 0, 1,2,..., N - 1 in equation correspond to the samples of continuous

transform at values 0, u, 2 u,. .., (N -1) u.

The terms u and x

x are related by the expression
11
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

12
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3
The two dimensional Discrete Fourier Transform ( 2D DFT )
If f (x, y) is an M N array, such as that obtained by sampling a continuous function of
two dimensions at dimensions M and N on a rectangular grid,
Then its two dimensional Discrete Fourier ttransform
ransform (DFT) is the array given by

and the inverse DFT (IDFT) is

Properties of the 2-D

D Discrete Fourier Transform
Relationships Between Spatial and Frequency Intervals
Suppose that a continuous function fa(x,y) is sampled to form a digital image, f(x, y),
consisting of M x N samples taken in the x
x- and y-directions,
directions, respectively.
Let x and y denote the separations between samples .
The function f(x, y) represents the samples of the function fa(xo + xx,
x yo + yy)
for x = 0,1,2,..., M-1 and y = 0,1,2,..., N -1.
Then, the separations between the corresponding discrete, frequency domain variables
are given by

Note that the separations between samples in the frequency domain are inversely
proportional both to the spacing between spatial samples and the number of samples.

Separability
separability property: the Fourier transform F(u, v) for the function f(x, y) can be
obtained in two steps by successively applying one-dimensional Fourier transform
The DFT pair that we have already discussed in equations is represented here for
convenience.
13
Page

Consider the term

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

Which is the one-dimensional

dimensional transform of f(x,y) considering x as constant. DFT is
computed for frequency values v = 0,1.. N-1
1 with y as time variable.
Now the expression for 2D DFT is

This is again one-dimensional

dimensional transform of F
F(x,v) computed for frequency values u =
0,1.. N-1 with x as the time variable. Here v is treated as constant.
Therefore, the two-dimensional
dimensional function F(x, v) is obtained by taking
aking a transform along
each row of f(x, y). The final result F(u, v) is then obtained by taking a transform along
each column of F(x,v) .
This can be illustrated as shown in Figure.

The same results can also be obtained by taking transforms along the columns of f(x, y)
and then along the rows using columns result.
14
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

Periodicity
As in the 1D case, the 2D Fourier transform (and its inverse) are infinitely periodic in
the u and v directions; that is,

Or

Where and are positive integers.

Symmetry:

.
If we plot a DFT with its origin at one corner of the display window, viz, the location
(0,0), only one fourth of the complete Fourier spectrum can be viewed. In order to view
the complete we need to shift the centre of the spectrum at another location (M/2, N/2).
15
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

16
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3

17

i.e,
Page

Nagaraja N S,SSE, Mukka

 Image Processing /Chapter 3
Convolution Property
This property forms the basic link between the spatial domain and the frequency
domain.
If f(x, y) is the image and h(x,y) is a 3 x 3 mask (or 5 x 5, 7 x 7) ...) then the modified
output g(x,y) is represented in spatial domain by the convolution operation
g(x, y) =f(x, y) * h(x, y)
The convolution in 1 dimension is given by

Convolution in the time domain is equivalent to multiplication in the frequency domain;

If we extend this to the 2-D domain,
we get

g(x, y) = f(x, y) * h(x, y) in the spatial domain is equivalnt,

to
G(u, v) = F(u, v) x H(u, v) in the frequency domain.

Hence, to find G(u,v), we take the 2-D DFT of the input image, then, we find the transfer
function of the filter, H(u, v), and simply multiply the two. We finally take the inverse
Fourier transform of G(u, v) to get the modified image, i.e., g(x, y).
18
Page

Nagaraja N S,SSE, Mukka