Unit 5 Image Transformations – Frequency Domain

UNIT 5
IMAGE TRANSFORMATIONS-
FREQUENCY DOMAIN

Structure Page No.

5.1 Introduction 53
Objective
5.2 Discrete Fourier Transform 54
5.3 Discrete Cosine Tranform 61
5.4 Discrete Wavelet Transform
5.5 Haar Transform 63
5.6 Summary 69
5.7 Solutions/ Answers 69

5.1 INTRODUCTION
As we saw in Unit 2, an image can be transformed, to show or hide
information in the image. Image transformations can be done in both the
spatial and the frequency domain. In the spatial domain, image
transformation is carried out by changing the value of the pixels based
on certain constraints. These transformations can change the
brightness and clarity of the images.

In this unit, we shall focus on the transformations in the frequency

domain. We recall at this point that an image is also a 2D signal and
that the transformations that can be applied on a signal can also be
applied on any image. The transformations in the frequency domain
provide us information on the frequency content of the image. These
transformations can help represent the information in the image in a
more compact form, thereby making it computationally easier to store
and transmit the images. The transformations may also help in
separating the noise and the salient information present in the image.

53
Block 2 DIGITAL IMAGE PROCESSING-II
In Sec. 5.2, we shall focus on very important and useful image
transformations, namely the Discrete Fourier transformation (DFT). We
shall continue our discussion in Sec. 5.3 with the Discrete Cosine
Transformation (DCT). Subsequently, Discrete Wavelet Transform will
be discussed in Sec. 5.4 Thereafter, In Sec. 5.5, Haar transform will be
discussed . As we go through this unit, we shall see the unique
properties of each of these transforms.
Now we shall list the objectives of this unit. After going through the unit,
please read this list again and make sure that you have achieved the
objectives.

Objectives
After studying this unit you should be able to:

 find the Discrete Fourier Transform (DFT)

 compute the Discrete Cosine Transform (DCT)
 find the Discrete Wavelet Transform(DWT)
 find the Haar Transform
 apply the above mentioned transforms

We shall begin the unit with Discrete Fourier Transform(DFT).

5.2 DISCRETE FOURIER TRANSFORM

The Discrete Fourier Transform (DFT) transfers an image from the
spatial domain to the frequency domain. It is one of the most important
transforms in image processing, which enables us to decompose an
image into its sine and cosine components. The output image after
applying the Fourier transformation is represented as a linear
combination of a collection of sine and cosine waves of different
frequencies.

Consider a 1D function, {f ( x ), 0  x  N  1}. The general form of a

transformation is
N 1
g(u )   T (u, x )f ( x ); 0  u  N  1 (1)
x 0

where T ( u , x ) is called the forward kernel of transformation and g ( u ) is

the transformed image.

If the transformation is the Discrete Fourier Transform.

Then,

N 1 x
1 i 2  N
g (u )   e f ( x ); u  0,1, 2,...N  1 (2)
x 0 N

The inverse 1-D DFT will then be,

54
Unit 5 Image Transformations – Frequency Domain

N 1 i 2  u x
f (x )   e N
g(u ) (3)
x 0

As can be seen the signal is written as a linear combination of an

orthogonal set of basis functions. Similarly, an image can be
transformed into a set of “basis images”, which can be used for
representing the image.

We can extend the transform to 2-D image.

Consider an image f ( x , y) of size M  N. The 2-D DFT of f ( x , y) is

defined as follows:

M 1 N 1  ux vy 
i 2    
F(u, v)   f ( x, y)e M N
(4)
x 0 y  0

And the inverse 2-D DFT is given by:

 ux vy 
1 M 1 N 1 i 2   
f ( x , y)  
MN u 0 v 0
F( u , v ) e M N
(5)

The 2-D DFT is separable, symmetric and unitary. In case of square

images, M  N. Many a time in image processing we work with square
images. Additionally if the imagesize is a power of 2, then DFT
implementation becomes very easy. Computational complexity can be
reduced by efficient algorithms such as FFT.

The representation of intensity as a function of frequency is called

‘spectrum’. In the Fourier domain image, each point to a particular
frequency contained in the spatial domain image. The coordinates of
the Fourier spectrum are the spatial frequencies. The spatial position
information of an image is encoded as the difference between the
coefficients of the real and imaginary parts. This difference is called the
“phase angle”. The phase information is very useful for recovering the
original information. The phase information represents the edge
information or boundary information of the objects present in an image.
For applications such as medical image analysis, the phase information
is very crucial in getting information from the image.

Note that while the image values f ( x, y) are going to be real, the
corresponding frequency domain data is going to be complex. There will
be one matrix containing real values R ( u , v) and the other matrix I( u , v)
will contain the imaginary component of the complex value. The
amplitude spectrum or the magnitude for 2D DFT is given by

F( u , v)  [ R 2 ( u , v)  I 2 ( u , v)]1 / 2 (6)

55
Block 2 DIGITAL IMAGE PROCESSING-II
where, R and I are real and imaginary parts of F( u , v) and and all
computations are carried out for the discrete variables u  0,1, 2, ..., M  1
and v  0,1, 2, ..., N  1. The spectrum tells us the relative magnitude at
each frequency.

The power spectrum of the 2-D DFT is defined as

2
P ( u , v )  F( u , v )  R 2 ( u , v )  I 2 ( u , v ) (7)

and the phase spectrum of the 2-D DFT is given by

I( u , v )
ø (u , v )  tan 1 (8)
R (u , v)

Note that the size of the image remains the same as the original image
in spatial domain.Therefore, the magnitude, Fourier (phase) spectrum
andthe power spectrum are all matrices of size M  N.

Remark: We can find 2-D DFT of an image by simply computing a set

of 1-D DFT is for all rows of f ( x , y). Thus, the 2-D DFT of an image
f ( x , y) is
M 1 2  i u x N 1 2  i vy
F(u, v)   e M
 f ( x , y) e N
.
x 0 y 0

M 1 2  i u x
  F( x, v) e M
,
x 0
N 1 2  i vy
where F( x, v)   f ( x, y) e N
.
y 0

Also, the 2-D DFT can also be found using the Eqn. (4) with the
condition of separability as we used in Unit-2.

Let us discuss properties of 2-D DFT.

DFT has several useful properties which makes it an important

transformation.

i) Separability: It is separable because a 2D transform is separable if

T(u, x, v, y)  T1 (u, x )  T2 ( v, y).
ii) Symmetry: It is symmetric because a 2D transform is symmetric if
T1 (u, x )  T2 (u, x ).
iii) Periodicity:The 2-D DFT and the 2-D IDFT are both periodic, that
is, F( u , v)  F( u  M, v)  F( u , v  N )  F( u  M, v  N ) .
iv) Conjugate symmetry: F(u , v)  F * (  u  pM ,  v  qN ), where p and
q being integers. The property of conjugate symmetry implies that
F( u , v)  F(  u , v)

v) If f ( x, y) is real and even then F( u , v) is real and even.

vi) If f ( x , y) is real and odd then F( u , v) is imaginary and odd.
56
Unit 5 Image Transformations – Frequency Domain
vii) Let F be the DFT operator, then
F(f ( x , y)  g ( x , y)  F(f ( x , y)  F(g ( x , y)

However, F(f ( x , y)  g ( x , y)  F(f ( x , y)  F(g ( x , y)

viii) Translation in the spatial domain by ( x 0 , y 0 ) implies
 ux vy 
 j2   0  0 
f ( x  x 0 , y  y 0 )  F(u , v)e  M N 

While translation in the frequency domain by ( u 0 , v 0 ) implies

 xu yv 
j2  0  0 
F(u  u 0 , v  v 0 )  f ( x , y)e  M N 

ix) The average value of the signal is given by

1 M 1 N 1
f ( x , y)   f ( x, y)
MN x 0 y 0
1 M 1 N 1
If we see the value of F(0, 0)   f (x , y)  F(0,0)  f (x, y) .
MN x 0 y0

x) Rotation: Rotating f ( x , y) by  rotates F( u , v) by .

Fig. 1: Rotation in f ( x , y)

Fig. 1 shows how the rotation in the spatial domain (on left) affects the
rotation in the frequency domain (right).

After going through all the properties of DFT, let us see how do we
visualize 2-D DFT. We need to translate the origin of the transformed
image to the center of the image (u , v)  (M / 2, N / 2) to be able to
display the full period of the 2-D DFT. As we saw above, translating
the Fourier image to the center, requires us to use the translation
property of F( u , v) with u 0  M / 2 and v 0  N / 2.
57
Block 2 DIGITAL IMAGE PROCESSING
PROCESSING-II
 i 2  0  0  
 xu yv 

Then, Ff (x, y) e  M N    F(u  u 0 , v  v 0 ) becomes
 
F{f ( x , y )e i ( x  y ) }  F{f ( x , y) ( 1) ( x  y ) }  F( u  M / 2, v  N / 2)

a) Original Image b) Original DFT image c) Translated DFT image

Fig. 2: DFT of an Original Image

We can see in Fig. 2 what changes do we see after DFT. Fig 2. (a) is
the original image in the spatial domain, (b) is the 2-D DFT image, and
(c) is the translated DFT image to show the full period of the 2D DFT of
image in (a).

Let us see how do we visualize the range of 2-D DFT.

In general, the range of values the 2-D DFT F( u , v) is very large.

Therefore, when we attempt to display the values of F( u , v ), smaller
values are not distinguishable because of quantization as can be seen
in Fig. 3 (b). Therefore, to enhance the small values, weapply a
logarithmic transformation given by
D( u , v)  c log(1  F( u, v) )
Where, the parameter 𝑐is chosen so that the range of
D( u , v) is [0, 255].
255
c
log(1  max{ F(u , v) })

a) The original image b) The 2D DFT image

c) The 2D DFT image after log transform

58 Fig. 3: DFT image after log tranform

Unit 5 Image Transformations – Frequency Domain
We can visualise the display of the amplitude of the 2-D DFT after
logarithmic transformation in Fig. 3(b) and Fig. 3(c) respectively for the
original image as shown in Fig. 3(a).

Example 1: Compute the DFT of the 1D sequence f ( x )  [1, 0,  1, 0] .

Solution: Here N  4. Using Eqn. (2) we get

 i , 2   ux
1 3
g(u )   f (x )  e
4 x 0
4
; u  0,1, 2, 3
ux
1 3   2 i 
  f ( x )   e 4  ; u  0,1, 2, 3
4 x 0  
1 3
 
4 x 0
f ( x ) (i) ux ; u  0,1, 2, 3

1
 [f (0)(i) 0  f (10( i) u  f ( 2)(i) 2 u  f (3)(i) 3u ]; u  0,.1, 2, 3
4
1
 [1  0  ( 1)(i) 2 u  0]; u  0,1, 2, 3
4
1
 [1  (i) 2 u ]; u  0,1, 2, 3
4

1
This gives g  [0, 2, 0, 2], which is the DFT of f ( x ).
4
***
Example 2: Construct a DFT matrix of order 2.

Solution: Here N  2.

 i . 2  . ux
1 1
g(u )   f (x ) e
2 x 0
2
; u  0,1.

1 1
 
2 x 0
f ( x ) (1) ux ; u  0,1

1
 [f (0)( 1) 0  f (1) (1) u ]; u  0,1
2
1
 [f (0)  ( 1) u f (1); u  0,1
2

1
Now, g(0)  [f (0)  f (1)]
2
1
g (1)  [f (0)  f (1)]
2
1 1 1
DFT matrix   .
2 1  1
***
Example 3: Compute the 2-D DFT of the 2  2 image

59
Block 2 DIGITAL IMAGE PROCESSING-II
1 1
f ( x , y)   .
1 1

Solution: Let the DFT of f ( x, y) be F( u , v), which is given in Eqn. (4).

1 1  ux vy 
2 i   
F(u, v)   f ( x, y) e  2 2 
; u, v  0, 1
x 0 y 0
1 1
  f ( x, y) (1) ux ( 1) vy ; u , v  0, 1
x 0 y 0

 [f (0, 0)(1) 0 (1) 0  f (0,1)(1) 0 (1) v  f (1, 0)(1) u (1) 0

 f (1,1)(1) u (1) v ]; u, v  0, 1
 [ f (0, 0)  ( 1) v f (0,1)  ( 1) u f (1, 0)  ( 1) u ( 1) v f (1,1)]; u , v  0, 1,1.

F(0, 0)  f (0, 0)  f (0,1)  f (1,0)  f (1,1)  4

F(1, 0)  f (0, 0)  f (0,1)  f (1, 0)  f (1,1)  0
F(0,1)  f (0, 0)  f (0,1)  f (1, 0)  f (1,1)  0
F(1,1)  f (0, 0)  f (0,1)  f (1, 0)  f (1,1)  0

 4 0
Thus, the 2-D DFT of the given image is  .
0 0
F(0,0) is 4 which happens to be the average of all the intensity values in
original image. The other values represent frequency values. But since
there is no variation in values of original image, there is no frequency
involved, and that is why the frequency values in DFT are zeroes.

Alternatively, the 2-D DFT can also be found using the DFT basis matrix
formed by finding 1-D DFT of each row of f ( x , y) and then using that as
kernel.

That is F( u , v)  kernel  f ( x , y)  (kernel) T

We have already found the DFT matrix of order 2 in Example 2.

Therefore,

1 1
kernel   
1  1
1
[we are skipping as in 2-D we
1 1 1 1 1 1 2
Hence, F(u, v)     
1  1 1 1 1  1 have taken
1
in inverse DFT]
MN
2 2 1 1
  
0 0 1  1
 4 0
 
0 0
60 Both the results are same.
Unit 5 Image Transformations – Frequency Domain
***
Try the following exercises.

E1) Does the implementation of a separable and symmetric transform,

such as the DFT in an image requires the sequential
implementation of the corresponding one-dimensional transform
row-by-row and then column-by-column? Justify your answer.

E2) Find the DFT of the sequence f ( x )  [i, 0, i, 1].

E3) Construct a DFT matrix of order 4. Also, check whether DFT
matrix is unitary matrix or not.

E4) Find the inverse 2-D DFT of F( u , v) found in Example 3.

In the following section, we shall discuss discrete cosine transform.

5.3 DISCRETE COSINE TRANSFORM

The Discrete Cosine Transform DCT is a family of unitary
transformations that transforms the real values of input image to
another set of real values. Unlike the DFT that is complex, the DCT is a
real transform because it projects the signal onto real cosinewaves.

The 1-D DCT is given as:

N 1
 (2x  1)u 
C(u )  a (u ) f ( x ) cos   ; 0  u  N  1, (9)
x 0  2N
 1
 ; u0
 N
where, a (u )  
 2 ; u  1, ..., N  1
 N

Then, the inverse transform is given by

N 1
 (2 x  1) u 
f ( x )   a (u ) C(u ) cos  , (10)
u 0  2N
where a ( u ) is the same function as used for DCT.

Let us visualize the effect of 1-D DCT through Fig. 4, where in the rows
of the 8 8 transformation matrix of the DCT for a signal f ( x ) with 8
samples are shown.

61
Block 2 DIGITAL IMAGE PROCESSING-II

Fig. 4: 1-D DCT

The figures for values of u from 0 to 7 show the various rows of the 8 8
transformation matrix of the DCT for a 1D signal f ( x ) with 8 samples.

Let us now expand 1-D DCT to 2-D DCT.

Consider an image f ( x , y) of size M  N. Then, the 2-D DCT of the

image is defined as:
M 1 N 1
 (2x  1)u   (2 y  1) v 
C(u , v)  a (u ) a ( v) f ( x, y) cos cos ,
x 0 y 0  2M   2 N  (11)
0  u  M  1, 0  v  N  1

And the inverse 2-D DCT is given by

M1 N1
 (2x  1)u   (2 y  1) v 
f ( x, y)   a (u ) a ( v) C(u, v) cos cos
u 0 v 0  2M   2 N  (12)
0  x  M  1, 0  y  N  1

Where a ( u ) is same as defined earlier for 1D DCT.

DCT and DFT are very similar, however, DCT has the advantage over
DFT that DCT is real while DFT is complex. Moreover. DCT has better
energy compaction in comparison to the energy compaction of DFT.
Energy compaction is the ability to pack the energy of the spatial
sequence into as few frequency coefficients as possible. This property
is exploited for image compression and is a very important property.You
can see in the Fig.5, that most of the DCT image is dark, which means
DCT values are concentrated only in few pixels very near the origin.
This indicates that DCT has high compaction.

a) Original image – Lena b) 2D DCT of Lena

Fig.5: The 2D DCT in (b) of the image Lena in (a) shows the high compaction
capability of DCT.
62
Unit 5 Image Transformations – Frequency Domain
Example 4: Compute the discrete cosine transform (DCT) matrix for
order 2.

Solution: Using Eqn. (9), we substitute N  2, and we get

1
(2 x  1) u
C(u )  a (u ) f ( x ) cos ; 0  u  1.
x 0 2 2
 1
 2 ; u0

where a ( u )  
 2  1; u  1
 2

At u  0, we get
1 1 (2 x  1)   0
C ( 0)  
2 x 0
f ( x ) cos
4
1
1
  f (x ) 1
2 x 0
1 1
  f (x)
2 x 0
1
 [f (0)  f (1)]
2
At u  1, we get
1
(2 x  1) 1 
C(1)  1  f ( x )  cos
x 0 4
1
(2x  1) 
  f ( x ) cos
x 0 4
  3 
 f (0) cos  f (1) cos 
 4 4
1 1
 f ( 0)  f (1)
2 2

By collecting the coefficients, we get the required DCT. Therefore, DCT

is
 1 1
 2 2
C( u )  
1 1 
 
 2 2 
***
Now try the following exercises.

E5) Why is DCT important for image compression?

E6) Find the DCT of the matrix of order 4.

63
Block 2 DIGITAL IMAGE PROCESSING-II
So far, we have discussed discrete fourier transform and discrete
cosine transform.
5.4 DISCRETE WAVELET TRANSFORM

In Block-1 of this course we learned about the Spatial domain, it was

learned that the Spatial domain is the normal image space where the
term "the domain" refers to the normal image space that is represented
as a matrix of pixels, . In Spatial domain, the transformation methods
are executed by directly operating on the pixel values of an image.
Adjustments in spatial domain are made to the values in order to obtain
the desired level of improvement.

In earlier sections of this unit we learned about the second type of

domain i.e. the frequency domain, where the pace at which the
individual color components in an image shift is referred as the image's
frequency and in this domain i.e. frequency domain the prime focus is
on the rate at which the pixel values in the spatial domain vary. It is to
be noted that, in any image the color changes very quickly, for the
regions with high frequencies, whereas in regions that contain low
frequencies, the color changes quite gradually.

It is essential to keep in mind that, in contrast to the spatial domain, the

frequency domain does not provide direct operations on the values.
This restriction prevents you from performing some calculations in the
frequency domain. In order to begin the processing of the image, it must
first go through a transformation that restores it to its original frequency
distribution. It is possible to separate the frequency components into
two basic sub-components. Components with a high frequency that
correlate to the edges of an image and components with a low
frequency relates to the smooth regions of an image . This technique
does not result in the production of an image as its end product; rather,
it produces a transformation as its conclusion. It is necessary to carry
out an inverse transformation on the data that was produced as a result
of the processing that was done so that the image can be restored to its
perfect, original form.

Also, we learned that Fourier transform is a powerful tool that has been
available to signal analysts for many years. It gives information
regarding the frequency content of a signal. However, the problem with
using Fourier transforms is that frequency analysis cannot offer both
good frequency and time resolution at the same time. A Fourier
transform does not give information about the time at which a particular
frequency has occurred in the signal. Hence, a Fourier transform is not
an effective tool to analyse a non-stationary signal. To overcome this
problem, windowed Fourier transform, or short-time Fourier transform,
was introduced. Even though a short-time Fourier transform has the
ability to provide time information, multi-resolution is not possible with
short-time Fourier transforms. Wavelet is the answer to the multi-
resolution problem. A wavelet has the important property of not having a
fixed-width sampling window.
64
Unit 5 Image Transformations – Frequency Domain
The technique of wavelet transformation, which is typically utilised for
the analysis of images and the compression of data, will be investigated
in this section. In spite of the fact that the frequency domain
encompasses a number of other mathematical transformations,
including the Fourier transform, the Laplace transform, and the Z
transform, the wavelet transformation method is going to be the one that
is discussed in detail in this section.

Let's begin with acquiring an understanding of what wavelets are and

why we need this transformation before we move on to trying to
comprehend the Discrete Wavelet Transformation, often known as the
DWT. This will help us get a better grasp on the DWT. According to
Wikipedia, "a wavelet is a wave-like oscillation with an amplitude that
begins at zero, rises, and then decreases back to zero." It is best to
think of it as a "short oscillation" in the majority of situations, which is
analogous to what could be captured by a seismograph or heart
monitor.

Let’s try to understand this concept of wavelet in a better way, with the
explanation given below:
A wavelet is a wave-like oscillation that is localised in time; an example
of this type of oscillation is provided further down in this paragraph.
Scale and location are the two fundamental features that wavelets
possess. How "stretched" or "squished" a wavelet is can be defined by
its scale, which can also be referred to as its dilation. This characteristic
is connected to frequency in the sense that it is understood for waves.
The wavelet's position in time can be determined based on its location
(or space).

The magnitude of the wavelet can be calculated by examining the value

of the parameter labelled "a" in the preceding expression. When the
value is decreased, the wavelet will take on an appearance that is more
compressed. As a direct result of this, it is now possible to acquire
information at high frequencies. In contrast, increasing the value of "a"
will cause the wavelet to stretch, which will lead to the collection of
information at low frequencies. The value of the "b" parameter is what
decides where the wavelet is positioned in the image. When the value
of "b" is decreased, the wavelet will shift to the left. Increasing the value
of "b" will cause it to relocate to the right. In contrast to waves, wavelets
only exhibit non-zero behaviour for a small period of time during which
their locations are meaningful. This difference in behavior between
wavelets and waves is called the wavelet scale. When we perform an
analysis of a signal, in addition to being interested in the oscillations that
65
Block 2 DIGITAL IMAGE PROCESSING-II
the signal displays, we are also interested in the locations of those
oscillations.

The fundamental concept here is to determine the proportion of a

wavelet that exists in a signal at a specific scale and location. For those
of you who are familiar with convolutions, this is a perfect example. A
signal is convolved with a set of wavelets operating at a range of
different scales. We go with a wavelet that has a specified scale. After
that, we multiply the wavelet and the signal at each time step, and then
we slide this wavelet across the entire signal, which means we change
where it is located. The result of performing this multiplication provides
us a coefficient that corresponds to that wavelet scale at that time step.
After that, the wavelet scale is increased, and the procedure is carried
out again.

Based on previous explanation, we understood that wavelets are

functions that are concentrated in time and frequency around a certain
location.

Generally, got confused for waves and wavelets but they are different
the fundamental difference between the two is that a wave is an
oscillating function of time or space that is periodic. The wave is an
infinite length continuous function in time or space. In contrast, wavelets
are localised waves. A wavelet is a waveform of an effectively limited
duration that has an average value of zero.
A function () can be called a wavelet if it posses the following properties:
1. The function integrates to zero, or equivalently its Fourier transform
denoted as ψ(ω) is zero at the
onigin:
∞
∫ ∞ ψ(x)dx = 0 (12a)

This implies ψ(ω)| =0 in the frequency domain.

2. It is square integrable, or equivalently, has finite energy:
∞
∫ ∞ |ψ(x)| dx < ∞ (12b)

3. The Fourier transform ψ(x) must satisfy the admissibility condition

given by
∞ | ( )|
C ∫ ∞
= dω < ∞ (12c)
| |

Interpretation of Eqs. (12a), (12b) and (12c)

Equation (12a) suggests that the function is either oscillatory or has a
wavey appearance.Equation (12b) implies that most of the energy in
ψ(x) is confined to a finite interval, or in other words. ψ(x) has good
space localisation. Ideally, the function is exactly zero outside the finite
interval. This implies that the function is a compactly supported function.
Equation (12c) is useful in formulating the inverse wavelet transform.
From Eq. (12c), it is obvious that in ψ(x) must have a sufficient decay in
66
Unit 5 Image Transformations – Frequency Domain
frequency. This means that the Fourier transform of a wavelet is
localized, that is, a wavelet mostly contains frequencies from a certain
frequency band. Since the Fourier transform is zero at the origin, and
the spectrum decays at high frequencies, a wavelet has a bandpass
characteristic. Thus a wavelet is a ‘small wave’ that exhibits good time-
frequency localisation. A family of wavelets can be generated by dilating
and translating the mother wavelet in ψ(x) which is
given by
ψ( , ) (x) = ψ
√
(12d)
Here, a is the scale parameter and b is the shift parameter.
After understanding the concept of wavelets now it’s time to begin with
our topic of Wavelet Transform.

Wavelet transforms can be either continuous or discrete, depending on

how they are implemented. The Continuous Wavelet Transform (CWT)
uses every wavelet that is feasible across a range of scales and places,
meaning that it has an endless number of scales and locations to
choose from. This is the primary distinction between these two types.
While the Discrete Wavelet Transform (DWT) uses a limited number of
wavelets, which are defined according to a specific range of sizes and
locations, this set of wavelets is not limited in any way. Few ore
comparisons between CWT and DWT are given below”

CWT- Continuous Wavelet DWT- Discrete Wavelet

Transform Transform
1 Scale At any scale Dyadic scales
2. Translation At any point Integer point
3. Wavelet Any wavelet that satisfies Orthogonal, biorthogonal,
minimum ...
criteria
4. Large Small
Computation
5. Detection Easily detects direction, Cannot detect minute
orientation object if not finely tuned
6. Application Object Detection, Pattern Compression, De-
recognition noising, Transmission
Feature extraction Characterisation

So, the wavelet transform can be broadly classified into (i) continuous
wavelet transform, and (ii) discrete wavelet transform. For long signals,
continuous wavelet transform can be time consuming since it needs to
integrate over all times. To overcome the time complexity, discrete
wavelet transform was introduced. Discrete wavelet transforms can be
implemented through sub-band coding. The DWT is useful in image
processing because it can simultaneously localise signals in time and
scale, whereas the DFT or DCT can localise signals only in the
frequency domain. 67
Block 2 DIGITAL IMAGE PROCESSING-II
It is to be noted that apart from image processing, the DWT is quite
promising tool for the Signal processing also. After the suggestion of
Mallat's, that signals may be represented at several resolutions using
wavelet decomposition, Discrete Wavelet Transform (DWT) emerged as
an extremely flexible tool for signal processing. Because the energy of
wavelets is concentrated in time while still retaining the wave-like
(periodic) characteristics, we discovered that wavelets make it possible
to perform time and frequency analysis of signals at the same time. This
was one of the key takeaways from the investigation into wavelets. As a
consequence of this, wavelet representation offers a flexible
mathematical tool for the analysis of transient, time-variant (non-
stationary), signals that are not statistically predictable, particularly in
the region of discontinuities. This quality is characteristic of images that
have discontinuities at the edges. In DWT, a digital signal splits up into
its component sub-bands, so that the lower frequency sub-bands have
finer frequency resolution and coarser time resolution compared to the
higher frequency sub-bands. .

The wavelet transformation technique overcomes the limitations of the

Fourier method. The Fourier transformation, despite the fact that it deals
with frequencies, does not reveal any facts regarding the passage of
time. In accordance with the Heisenberg's Uncertainty Principle , we
can either have a high frequency resolution but a low temporal
resolution, or vice versa. The introduction to the
Heisenberg's Uncertainty Principle is given below:

The Heisenberg uncertainty principle was originally stated in physics,

and claims that it impossible to know both the position and momentum
of a particle simultaneously. However, it has an analog basis in signal
processing. In terms of signals, the Heisenberg uncertainty principle is
given by the rule that it is impossible to know both the frequency and
time at which they occur. The time and frequency domains are
complimentary. If one is local, the other is global. Formally, the
uncertainty principle is expressed as
1
(∆𝑡) (∆𝜔) ≥
4
In the case of an impulse signal, which assumes a constant value for a
brief period of time, the frequency spectrum is infinite; whereas in the
case of a step signal which extends over infinite time, its frequency
spectrum is a single vertical line. This fact shows that we can always
localise a signal in time or in frequency but not both simultaneously. If a
signal has a short duration, its band of frequency is wide and vice
versa.

The Wavelet Transform offers a number of important benefits, including

the following:

 The Wavelet transform has the ability to concurrently

extract local spectral and temporal information.
 A selection of different wavelets from a variety to choose
68
Unit 5 Image Transformations – Frequency Domain

The first significant benefit is one that we have gone over in some detail
already. This is most likely the most important advantage of utilising the
Wavelet Transform. This may be preferable to employing a method
such as a Short-Time Fourier Transform, which needs slicing a signal
into segments and then applying a Fourier Transform to each individual
segment.

The second essential benefit appears to be more of a logistical

consideration. In the end, the most important thing to take away from
this is that there is a large variety of wavelets from which to choose in
order to get the one that most closely matches the characteristic shape
that you are seeking to extract from your signal.

In comparison to the Fourier Transform, the Wavelet Transform has the

primary benefit of being able to extract local information that is both
spectral and temporal in nature. Analyzing electrocardiogram (ECG)
readings, which comprise periodic and transient signals of relevance, is
an example of a real-world use of the Wavelet Transform.

As a result, we realised that non-stationary signals are the ideal

candidates for the use of the wavelet transform. By applying this
transformation, one can obtain a high temporal resolution for high-
frequency components while maintaining a decent frequency resolution
for low-frequency components. This technique begins with a mother
wavelet, which could be a Haar, Morlet, or Daubechies, among other
options. After that, the signal is essentially recast as scaled and shifted
iterations of the mother wavelet. We will discuss Haar transformation
in the subsequent section 5.5 of this unit

Important points:

 The wavelet transform is used to decompose a time series; this

results in waves that are not only localised in frequency but also
in time.
 One of the most significant drawbacks of the Fourier Transform
is that it collects global frequency information, which refers to
frequencies that are present throughout an entire signal. There
are some applications, such as electrocardiography (ECG), in
which the signals include brief intervals of distinctive oscillation,
that this form of signal decomposition would not suit very well.
The Wavelet Transform is an alternate method that may be used,
and what it does is it decomposes a function into a group of
wavelets.
 A simple comparison between Wavelet Transform and Fourier
Transform: The Fourier transform can be thought of as a special
case of the wavelet transform. A function is decomposed by the
Fourier transform into sine and cosine waves, which serve as the
base functions. Although the period lengths of the sine and
cosine waves change, the base functions remain the same
across the entire interval. The wavelet transform, on the other
hand, applies scaling as well as shifting to the basic functions.
69
Block 2 DIGITAL IMAGE PROCESSING-II
Also, the sine and cosine waves are not required to be the base
functions, despite the fact that there are some well-known base
functions. Instead, we are free to choose whichever functions we
want to use as base functions, so long as they fulfills the
fundamental condition of a wavelet, which is that it has a finite
amount of energy.

5.5 HAAR TRANSFORM

The Haar transform is a wavelet transform. Wavelet transforms are

based on small waves called wavelets which are of varying frequencies
and limited duration. These are different from the Fourier transform,
where the basis functions are sinusoids. Haar transform is a transform
whose basis functions are orthonormal wavelets. The Haar transform
can be expressed as
T  HFH T (13)
where, F is an N  N image matrix, H is the N  N Haar transform matrix
and T is the resulting N  N transform.
The Haar transform, H, contains the Haar basis functions, h k ( t ). They
are defined on a continuous interval, t  [0,1] for k  0,1, ..., N  1, where
N  2 n. Then, H is generated by uniquely decomposing the integer k as
k  2 p  q  1, where, 0  p  n  1 and when p  0, q  0,1; p  0 then,
1  q  2n .

For example, when N  4, k will take the values k  0,1, 2, 3. For these
the corresponding values of p and q have to satisfy that k  2 p  q  1.

Therefore, we compute the values of k, p and q in Table 1.

Table 1
k 0 1 2 3
p 0 0 1 1
q 0 1 1 2

0 1 N  1
Let t take the values from the set  , ,..., .
N N N 
Then, the Haar basis functions are recursively defined as:
 For k  0, the Haar function is defined as a constant

h 0 (t)  1/ N (14)

 When k  0, the Haar function is defined by

70
Unit 5 Image Transformations – Frequency Domain
 p/2 (q  1) (q  0.5)
2 ; p
t
2 2p

1  p/ 2 (q  0.5) q
h k (t )   2 ; p
t p (15)
N 2 2
0 ; othewise


As can be seen from the definition of the Haar basis functions, for the
non-zero part of the function, the amplitude and width is determined by
p while its position is determined by q.

We now show how the Haar transform matrix can be computed at

n  m / N , where n  0,1,..., N  1 to form the N  N discrete Haar
transform matrix through the following examples.
Example 5: For, N  2, compute the discrete Haar transform of a 2  2
matrix.
Solution: Here, N  2, we know that N  2 n .
Substituting the value of N, we get 2  2 n , which gives n  1.
Since, 0  p  n  1, we get 0  p  0.

Therefore, p  0, and hence q  0, 1, 2.

We determine the value of k using the relation k  2 p  q  1, we obtain

p 0 0
q 0 1
k 0 1
1
for k  0, h 0 ( t )  [using Eqn. (14)]
2
1  1 ;t0
for k  1, h, ( t )  
2  1 ; t  1 / 2
1 1 1
Thus, Haar transform is h 2   .
2 1  1
***
Example 6: For N  8, the 8  8 discrete Haar transform matrix.
Solution:As you know we need to find various parameters of Haar
transform. So, we find them as follows:
i) Here N  8,
ii) N  2n  n  3
iii) when p  0, q  0,1
p  1, q  1, 2
p  2, q  1, 2, 3, 4
iv) All the possible values of k for each set of p and q are given below:
p 0 0 1 1 2 2 2 2
q 0 1 1 2 1 2 3 4
71
Block 2 DIGITAL IMAGE PROCESSING-II
k 0 1 2 3 4 5 6 7

 1 2 3 4 5 6 7
v) Accordingly, t  0, , , , , , , .
 8 8 8 8 8 8 8
Now, we compute h k ( t ) for each k and t .
For k  0, h 0 ( t )  0. for all t .
 p / 2 q 1 q  0.5
 2 ; 2p  t  2p

1  p / 2 q  0.5 q
In general, h k ( t )   2 ; p
t p . (16)
8 2 2
 0 ; otherwise


Now, let us find each h k ( t ) for each of the interval of t for a particular k
using Eqn. (16) in the following table:
For k  1
Parameters h k (t) Haar Transform
k, q, p after
simplification
k  1,  0 1  0.5 1 1
 1; 2 0  t  2  1  t  2 h1 (t )  ; for
q  1, 2 2

p0 1  1  0. 5 1 1 1 2 3
h1 (t )   1;  t  0   t 1 t  0, , ,
8 2 2 2 8 8 8
 0 ; otherwise 1
 h1 (t)  ; for
 2 2
4 5 6 7
t , , ,
8 8 8 8
k  2,  1 1
 2 ;0t  4 h 2 ( t )  ; for
q  1, 2

p 1 1  1 1 1
h 2 (t )   2 ;  t  t  0,
8 4 2 8
 0 ; otherwise 1
h 2 (t )  ; for
 2

2 3
t ,
8 8
h 2 ( t )  0; for
4 5 6 7
t , , , .
8 8 8 8
k  3,  1 3 1
 2 ; 2t4 h 3 ( t )  ; for
q  2, 2

p 1 1  3 4 5
h 3 (t )   2 ;  t  1 t ,
8 4 8 8
 0 ; otherwise 1
h 3 (t)  ; for
 2

6 7
t ,
8 8
72
Unit 5 Image Transformations – Frequency Domain
h 3 ( t )  0; for
1 2 3
t  0, , , .
8 8 8
k  4,  1 1
;0t h 4 (t)  ; for
q  1,  2 8 2

p2 1  1 1 t 0
h 4 (t)   2 ; t 1
8 8 4 h 4 (t)  ; for
 0 ; otherwise 2
 1
 t
8
h 4 (t )  0; for
2 3 4 5 6
t , , , , ,
8 8 8 8 8
7
8
k  5,  1 3 1
h 5 (t)  ; for
q  2,  2; 4t8 2

p2 1  3 1 2
h 5 (t )   2 ;  t  t .
8 8 2 8
 0 ; otherwise 1
h 5 (t)  ; for
 2

3
t
8
h 5 ( t )  0; for
1 4 5 6 7
t  0, , , , ,
8 8 8 8 8
k  6,  1 5 1
h 6 (t)  ; for
q  3,  2 ;2  t  8 2

p2 1  5 3 4
h 6 (t )   2 ;  t  t
8 8 4 8
 0 ; otherwise 1
h 6 (t)  ; for
 2

5
t
8
h 6 ( t )  0; for
1 2 3 6 7
t  0, , , , ,
8 8 8 8 8
k  7,  3 7 1
h 7 (t)  ; for
q  4,  2 ;4  t  8 2

p2 1  7 6
h 7 (t )   2 ;  t  1 t
8 8 8
 0 ; otherwise 1
h 7 (t)  ; for
 2

7
t
8
h 7 ( t )  0; for
73
Block 2 DIGITAL IMAGE PROCESSING-II
1 2 3 4 5
t  0, , , , , .
8 8 8 8 8
The Haar transform is given in the following matrix.
 1 1 1 1 1 1 1 1 
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
 
 1 1 1 1 1 1 1 1 
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
 1 1 1 1 
 0 0 0 0 
 2 2 2 2 
 0 1 1 1 1 
0 0 0
h k ( t )   1 1
2 2 2 2 

 0 0 0 0 0 0 
 2 2 
 0 1 1
0 0 0 0 0 
 2 2 
 1 1 
 0 0 0 0 0 0 
 2 2 
 0 1 1 
0 0 0 0 0
 2 2 

The plot of these 8 basis functions are shown in Fig. 6.

Fig. 6: Haar Basis Functions

As can be seen by Fig.6, all non-zero Haar functions h k ( t ), k  0

consists of a square wave and its negative version, and the
parametersp defines the magnitude and width of the shape while q
specifies the position (or shift) of the shape. This gives the unique
property to the Haar transform that it not only represents the signal at
different scales based on the different frequencies, but also represents
their locations across time.
Moreover, an important property of the Haar transform matrix is that it is
real and orthogonal, that is, H  H * and H 1  H T . The orthogonal
property of the Haar transform allows the analysis of the frequency
components of the input signal. The Haar transform can also be used
for analyzing the localized feature of the signal.

74 a) Original Image b) Output of Haar Transform

Unit 5 Image Transformations – Frequency Domain

Fig. 7: Haar Transform

Fig. 7 (b) shows the output of Haar Transform of the image in Fig. 7 (a).

Try the following exercises.

E7) Let X  [ x[ 0], x[1], x[ 2], x[3]] T  [1, 2, 3, 4] T . Then, X is a 4-point

signal. Find the Haar transform coefficients and show that the
signal can be expressed as a linear combination of the basis
functions by the inverse transform.

E8) For N  4, compute h 4 , which represents the 4  4 discrete Haar

transform matrix.

Now let us, summarize what we have discussed in this unit.

3.5 SUMMARY
In this unit, we discussed transformations which convert the spatial
domain image to the frequency domain. We saw that these transform
provide a variety of information based on the frequency content of the
image.We discussed in depth three very important image transforms,
namely the Discrete Fourier transformation (DFT), the Discrete Cosine
Transformation (DCT) and the Haar transform. We also discussed the
properties of each of these transforms, which shall help us in using
them for image filtering in the frequency domain.

3.6 SOLUTIONS AND ANSWERS

E1) Consider an image f ( x , y) of size M  N and the generic image
transform T where, x indicates row and y indicates column.
Then,
M 1 N 1
g (u , v)   T (u , x , v, y) f ( x , y)
x 0 y0

If, T is separable and symmetric, then we can write g( u , v) as

M 1 N 1
g ( u , v)   T1 ( u , x )T2 ( v, y) f ( x , y)
x 0 y 0
M 1 N 1
 g (u , v)   T1 (u , x ) T2 ( v, y) f ( x , y)
x 0 y 0

N 1
Then,  T (v, y) f ( x, y) is the same as applying the one-
y0
2

dimensional transform long the x -th row of the image. By doing

this for each of the M rows of the image, we obtain an
intermediate image F( x , v).
M 1
Then, g(u , v)   T1 (u , x ) F( x, v)
x 0 75
Block 2 DIGITAL IMAGE PROCESSING-II
Therefore, we can see, the above sum corresponds to applying
the one dimensional transform along the v -th column of the
intermediate image ( x , v).
Therefore, we have shown that the implementation of a
separable and symmetric transform in an image requires the
sequential implementation of the corresponding one-dimensional
transform row-by-row and then column-by-column (or the
inverse). We explain this in the Fig. 9:

Fig. 9

 i  2  ux
1 3
E2) g (u )   f ( x )e
4 x 0
4
; u  0, 1, 2, 3

1
 [i  i( i) 2 u  ( i) 3u ]; u  0, 1, 2, 3
4

1
 [i  i(1) 2 u  (i) u ]; u  0, 1, 2, 3
4

1
g  [1  2i, i,  1  2i,  i].
4

E3) Here N  4.
 i 2  ux
1 3
g(u )   f (x ) e 4 ; u  0, 1, 2, 3
4 x 0
1
g (u )  [f (0)  ( i) u f (1)  ( i) 2 u f ( 2)  (i) 3u f (3)]; u  0, 1, 2, 3.
4
1
g (0)  [f (0)  f (1)  f (2)  f (3)]
4
1
g (1)  [f (0)  i f (1)  f ( 2)  i f (3)]
4
1
g ( 2)  [f (0)  f (1)  f (2)  f (3)]
4
1
g (3)  [f (0)  i f (1)  f ( 2)  i f (3)]
4
1 1 1 1
1  i  1 i 
1
Hence, the DFT matrix is A   .
4 1  1 1  1
 
1 i 1  i
You may check if A  A *T  I.
76
Unit 5 Image Transformations – Frequency Domain
 ux vy 
1 1 1 2 i  
E4) Using Eqn. (5), f ( x , y)  
2  2 x 0 y 0
F( u , v ) . e  2 2 
; u , v  0,1

1 1 1

4 x 0 y  0
F(u , v) (1) ux (1) vy ; u , v  0,1

1
 
 f (0,0)  (1) v F(0,1)  (1) u F(1, 0)  (1) u (1) v F(1,1) ; u, v  0,1
4
1 4 4 1 1
which gives f ( x, y)    .
4 4 4 1 1

E5) In general, in most of the images, large part of the signal energy
lies at the low frequencies which appear in the upper left corner
of the DCT image. Since the higher frequencies present in the
lower right of the image are small enough to be neglected, the
original image can be represented in less number of coefficients,
thereby achieving compression. Therefore, as DCT has good
compaction property, it can represent the original image in less
number of coefficients and therefore, storage and transmission of
the image is better and faster. Moreover, the original image can
be recreated close to the original from the most important
components of the DCT.

 0.5 0. 5 0.5 0.5 

 0.65 0.27  0.27  0.65
E6) C(u )   
 0.5  0.5  0.5 0.5 
 
0.27  0.65 0.65  0.27 

E7) X  [ x [ 0 ], x [1 ], x [ 2 ], x [ 3 ] ]T  [1, 2, 3, 4]T be the 4-point signal.

Then, we shall use the basis matrix, H 4 to compute the Haar
transform coefficients.

 1 1 1 1  1   5 
 1 1  1  1  2   2 
1   
2 2  2 0 0  3  1 2 
     
 0 0 2  2 4  1 2 

The inverse transform will be:

1 1 2 0   5 
   2 
1 1 1  2 0   
2 1  1 0 2   1 2 
   
1  1 0  2   1 2 

  1   1  2   0   1 
          
1   1   1  1  2  1  0    2 
 5 2   
2   1   1 2 0  2  2   3 
          
  1   1  0   2    4 
77
Block 2 DIGITAL IMAGE PROCESSING-II
As can be seen, X is a linear combination of the basis vectors.

 1 2 3
E8) Here N  4; n  2; p  0,1; t  0, , , . Let us write all the values
 4 4 4
of Haar transform in the following table:

k , q, p h k (t ) h k ( t ) after
simplification

p  0, q  0, k  0 1 1 1
h 0 (t)   for all t h 0 (t ) for
4 2 2
1 2 3
t  0, , , .
4 4 4

p  0, q  1, k  1  1 1 1
 1; 0  t  2 h 1 ( t )  ; t  0,
2 4

1 1
h 1 ( t )   1;  t  1 1 2 3
2 2 h1 (t)  ;t  ,
2 4 4
 0 ; otherwise



p  1, q  1, k  2  1 1
h 2 (t)  ;t  0
 2 ;0  t  4 2

1 1 1
h 2 ( t )   2 ;  t  1 1
2 4 2 h 2 (t)  ;t 
 0 ; otherwise 2 4

 2 3
h 2 ( t )  0; t  , .
4 4

p  1, q  2, k  3  1 3 1 2
h 3 (t)  ;t 
 2; 2  t  4 2 4

1 3
h 3 ( t )   2 ;  t  1 1 3
2 4 h 3 (t)  ;t 
 0 ; otherwise 2 4

 1
h 3 ( t )  0; t  0, .
4

Hence,

 1 1 1 1 
 2 2 2 2 
 1 1 1 1 
 
h k (t )   2 2 2 1 
 1 1
0 0 
 2 2 
 1 1 
 0 0 
 2 2 
78
 UNIT 6
IMAGE ENHANCEMENT &
FILTERING IN FREQUENCY
DOMAIN

Structure Page No.

6.1 Introduction 153

Objectives
6.2 Shifting the Centre of the Spectrum 154
6.3 Image Smoothing in Frequency Domain 158
6.4 Image Sharpening in Frequency Domain 165
6.5 Image Degradation 172
6.6 Image Degradation/Restoration Model 174
6.7 Noise Models 176
6.8 Restoration in the Presence of Noise Only – Spatial Filtering 181
6.9 Periodic Noise Reduction 187
6.10 Estimation of Degradation Function 194
6.11 Inverse filtering 197
6.12 Wiener filtering 199
6.13 Summary 169
6.14 Solutions/Answers 169

6.1 INTRODUCTION
In the previous units of this course, we have considered an image in the
spatial domain which is the form in which camera captures it. In this unit, we
shall view the image as a signal and apply the well known filtering techniques
used in signal processing. The only difference here will be that image would
be considered as 2-D signal (along x and y axes). We shall see that this view
of the image results in number of benefits over the spatial domain treatment.

In this unit, we will discuss various enhancement techniques in frequency

(fourier) domain. We discuss the basic issues associated with frequency
domain filtering. We also discuss various low pass and high pass filters in
frequency domain with their applications and advantages in image
enhancement.
153
Unit 6 Image Enhancement & Filtering-Frequency Domain
By discussing the enhancement techniques in frequency (fourier)
domain, we have looked at image improvement without bothering about
source which caused a degradation in the quality of the image. If the
source is known to us, it is possible to improve the quality of the image
in a better way. Thus, it is required to discuss the concept of Image
restoration/degradation.

Image restoration is a pre-processing method that suppresses a known

degradation. Image acquisition devices introduce degradation because
of defects in optical lenses, non- linearity of sensors, relative object
camera motion, blur due to camera mis-focus, atmospheric turbulence
etc. Restoration tries to reconstruct an image that was degraded by a
known degradation function. Iterative restoration techniques attempt to
restore an image by minimizing some parameter of degradation,
whereas blind restoration techniques attempt improve the image without
knowing the degradation function. Like image enhancement, image
restoration also aims to improve image quality, but it is more objective
process where as enhancement is a subjective process. Noise is
visually unpleasant, it is bad for compression and bad for analysis.
Restoration involves modeling of these degradations and applying
inverse process to recover the original image.

So we learned that Image restoration is the process of retrieving an

original image from degraded image. The idea is to obtain an image as
close to the original image as possible. This is possible by removing or
minimizing degradations. This is often difficult in case of extreme noise
and blurs, and often called an inverse problem. An inverse problem
aims to find the cause and extent of degradation.

In this unit we will also learn that the image Restoration involves
modeling of the degradations and by applying inverse process we can
recover the original image. For the restoration process, it is mandatory
that we estimate the degradation process accurately. Else we will not be
able to remove it.

Now, we shall list the objectives of this unit. After going through the unit,
please read this list again and make sure that you have achieved the
objectives.

Objectives
After studying this unit, you should be able to
 define images in frequency domain
 perform filtering in frequency domain
 apply different types of image smoothing filters
 apply different types of image sharpening filters
 describe images degradation models
 state difference between restoration and enhancement
 apply different noise models
 estimate degradation function
 apply Inverse filtering
 apply Wiener filtering
154
Unit 6 Image Enhancement & Filtering-Frequency Domain
Let us begin with shifting the centre of the spectrum.

6.2 BASICS OF FILTERING IN FREQUENCY

DOMAIN - SHIFTING THE CENTRE OF THE
SPECTRUM
To start with we understood that any signal (periodic or non periodic) can be
expressed as the summations of sines and/or cosines multiplied by a
weighting function. This is carried out by applying Fourier transform on the
image.

In 1D signal, the fourier transform takes the form


F(u )   f ( x ) e i 2  ux dx


F(u ) can be expressed in polar coordinates:

F( u )  F( u ) e j  ( u )

 
1
where F(u )  R 2 ( u )  I 2 (u ) 2 (magnitude or spectrum)
 I( u ) 
(u )  tan 1   (phase angle or phase spectrum)
 R (u ) 
R ( u ) : the real part of F(u )
I( u ) : the imaginary part of F(u )
The various benefits of frequency domain analysis are the following:
1) It is convenient to design a filter in frequency domain. As filtering is more
intuitive in frequency domain, designing an appropriate filter is easier.
2) Implementation is very efficient with fast DFT via FFT.
3) Convolution in spatial domain reduces to multiplication in frequency
domain which is a much simpler process.
However, the image in spatial domain is not continuous but consists of
discrete values. The discrete version of fourier transform is
2  ux
1 N 1
F(u )   f ( x )e N
, u  0,1,..., N  1
N x 0
Also, since the image is two dimensional signal, we need 2D Fourier
transform. For a N  N image it takes the form:
 ux  vy 
1 N1 N1  j2  
F( u , v)    f ( x, y) e  N 
,
N x 0 y  0

where u and v are the frequencies along x and y axes and take the values
0,1, 2,..., N  1 .
In the spatial domain we consider the origin to be located at top left corner of
the image. For better display in the frequency domain, it is common to shift the
origin to centre of the image.
Periodicity of Fourier transform is given by
v ( k , l)  v ( k  M , l)  v ( k , l  N )  v ( k  M , l  N ) (1)
155
Unit 6 Image Enhancement & Filtering-Frequency Domain
u ( m, n )  u ( m  M , n )  u ( m, n  N )  u ( m  M , n  N ) (2)

Fig 1(a) shows that the values from N / 2 to N  1 are the same as the value
from N  1 to 0. As DFT has been formulated for value of k in the interval
[0, N  1] , the result of this formulation yield two back to back half periods in
this interval. To display one full interval between 0 to N  1 as shown in
Fig. 1(b), it is necessary to shift the origin of transform to the point k  N / 2 .
To do so we have to take advantage of translation property of Fourier
transform.

 M N
v(m, n ) (1) m n 
FT
v k  , l   (3)
 2 2

(a) Spectrum of f(x) without shifting centre.

(b) Spectrum of f(x) after shifting centre.

Fig. 1

Fig. 2 (a) and (b) show how the origin shifts from left corner of the image to
centre of the image.
Basic Property of images in Frequency Domain

The forward transform of input image u (m, n ) is given by

1 N1 N1
v ( k , l)  
N m 0 m 0
u (m, n ) WNkm WNl n 0  k, l  N  1 (4)

156
Unit 6 Image Enhancement & Filtering-Frequency Domain

(a) Change of centre in the spectrum of an image.

(b) Change of centre in the spectrum of an image.

Fig. 2
Following properties of the Fourier transform are observed

i) Each term of v(k , l) contains all the values of u ( m, n ) modified by the

values of exponential terms.
ii) Frequency is directly related to the rate of change of grey level values.
iii) DC value or the average grey level value in an image is the slowest
varying components corresponding to u  0, v  0 . It is also the largest
component in the frequency domain.
iv) Smooth variation of grey levels corresponds to low frequency components.
Slow varying components can be the background of an image, hair of a
person, skin, or texture etc.
v) Faster grey level changes correspond to high frequency components.
These can be the edges/boundary of the objects or noise present in the
image.
vi) As we move away from origin, higher frequencies are encountered.

Fig. 3 shows the variation in frequency of a centred spectrum.

Fig 3: Frequency variation in an image.

Also, note that the rotation of an image in spatial domain causes exactly same
rotation in frequency domain.
157
Unit 6 Image Enhancement & Filtering-Frequency Domain
 Rotating f ( x , y) by  rotates F( u , v) by  .

Once the image is transformed in frequency domain, it is easy to carry out

image processing operations on it. We apply a low pass filter, if we are
interested in only slowly varying components of the image (like object shapes),
and we wish to suppress high frequency components (like noise). If we are
interested in highlighting the edges or special textures, we can employ high
pass filters, which will allow high frequency components to be displayed.
Filtering in frequency domain is multiplication of a suitable filter H ( u , v) by
image in Fourier domain F( u , v) to result in G (u , v) . By taking inverse Fourier
Transform of G ( u , v) we get the image back in spatial domain.

Generally, the filters are centred and are symmetric about the centre. Input
image should also be centred. Following steps are followed to carry out
filtering in frequency domain (Fig. 4):

xy
Step 1: Multiply input image f ( x , y) by (1) to move the origin in the
transformed image to
M N
u and v 
2 2
Step 2: Compute F( u , v) , Fourier transform of the output of step 1.
Step 3: Multiply filter function H (u , v) to F( u , v) to get G (u , v) .
Step 4: Take inverse Fourier transform of G ( u , v) to get g ( x , y) .
Step 5: Take the real part of g ( x , y) to get g r (x, y)
Step 6: Multiply the result of step 5 by (1) x  y to shift the centre back to origin
and enhanced image is generated.

Fig. 4: Block Diagram of Filtering in Frequency Domain.

Types of Frequency Domain Filters

Frequency domain filters are categorized into three types.

1. Smoothing filters
2. Sharpening filters
3. Homomorphic filters
Smoothing filters are low pass filters and are used for noise reduction. It blurs
objects. Sharpening filters are high pass filters and produce sharp images with
dark background. Laplacian and high boost filters are used to produce sharp
images. Homomorphic filters are based on illumination and reflectance model,
and create a balance between smoothing and sharpening filtering effect. This
classification is shown in Fig. 5.
158
Unit 6 Image Enhancement & Filtering-Frequency Domain

Fig. 5: Types of Frequency Domain Filters.

Try the following exercises.

E1) Write the steps involved in frequency domain filtering with the help of
block diagram.

E2) Explain how image enhancement is better in the frequency domain as

compared to spatial domain.

In the following section, we will discuss smoothing filters in the frequency

domain.

6.3 SMOOTHING FILTERS

Smoothing filters are low pass filters (LPF). Edges, sharp transitions and noise
in the grey levels contribute to high frequency contents in an image. A low
pass filter only passes low frequency and blocks the high ones. It removes
noise but in the process introduces blurring as a side effect in the image.
The basic model of filtering is

G ( u , v)  H( u , v) F( u, v) (5)

where F( u , v)  Fourier transform of the image to the filtered, H( u, v) 

Transfer function of the filter, and G (u , v)  Enhanced image where high
frequency components have been alternated.

The transfer function H ( u , v) is of three types

a) Ideal LPF
b) Butterworth LPF
c) Gaussian LPF
Ideal filters has sharp slope in transition band whereas Gaussian filter has
smooth slope in transition. Butterworth filter has a parameter called filter order
which controls the slope of transition band. Higher value of filter order leads to
ideal filter.
6.3.1 Ideal Low Pass Filters (ILPF)
Low pass filter removes all frequencies above a certain frequency components
D 0 . Ideal low pass filter is defined by the transfer function

1 D(u , v)  D 0
H ( u , v)  
0 D(u , v)  D 0
159
Unit 6 Image Enhancement & Filtering-Frequency Domain
 M
1/ 2
 N 
1/ 2

Where D(u , v)   u   v   

 2  2  
M N
D( u , v) is the distance from point (u , v) to the centre  ,  . If size of an
 2 2
M N
image is M  N , then the centre is at  ,  . Filter transfer function is
 2 2
symmetric about the midpoint.

D 0 is non negative quantity specifying the frequency content to be retained. It

is also called cut off frequency. Fig. 6(a), Fig. 6(b) is the plot of ILPF and
Fig. 6 (c) is perspective plot and Fig. 6(d) is filter displayed as an image.

(a) Transfer Function (b) Plot of Ideal LPF

c) Prospective plot of ILPF (d) ILPF displayed as an image with D0 = 10

Fig. 6

Choice of cut off frequency in an ideal LPF

1. The cutoff frequency D 0 decides the amount of frequency components

passed by the filter.

2. Smaller the value of D 0 , more are the number of frequency components

eliminated by the filter.

3. In general, D 0 is chosen such that most of the frequency components of

interest are passed while unnecessary components are eliminated.

4. A useful way to establish a set of standard cut off frequencies is to

compute circles having a certain percentage of the total image power.
M 1 N 1
PT   P(u , v)
u 0 v 0
160
Unit 6 Image Enhancement & Filtering-Frequency Domain
2
Here P(u , v)  F(u, v) is the total image power.

5. Consider a circle of radius D 0 () as the cut off frequency with respect to a
threshold a such that

 P(u, v)   P
u v
T .

6. Thus, we can fix a threshold α which tells how much of the total energy is
retained and obtain an appropriate cut off frequency D 0 () .

Properties of ILPF

As it is an ideal filter, it is non-real, non-actual and non-physical. But it can be

stimulated in computers. Fig 7 and 8(b) to (c) show low pass filtered images
with different cut off frequencies. As the filter radius increases, less and less
power and information is removed which resulted in less blurring. A very
noticeable effect that can be seen in the output image is ringing.

Now, the question arises that what is ringing?

Ringing is undesirable and unpleasant lines around the objects present in the
image Fig. 7 (b). As the cut of frequency D 0 increases, effect of ringing
reduces. Ringing is a side effect of ideal lpf .

Why is there Ringing in Ideal LPF?

Ideal LPF function is a rectangular function as shown in Fig. 6-X. The inverse
Fourier transform of a rectangular function, is a sinc function. We can observe
two distinctive characteristic of sin c function:
1. A dominant component at the origin which is responsible for blurring.
2. Concentric circular components are responsible for ringing which is the
characteristic of an ideal LPF.

Fig. 6-X: Fourier Inverse of Rectangular Transfer Function

1
Radius of the centre component 
cut off frequency
1
Number of circles per unit distance from origin 
cut off frequency

Thus, as the cut off frequency (D 0 ) is increased, blurring as well as ringing

reduces. The examples are given in Fig. 7 and Fig. 8.

161
Unit 6 Image Enhancement & Filtering-Frequency Domain

(a) Original image (b) Output of ILPF with D0=30 (c) Output of ILPF
with D0= 50
Fig. 7

(a) Original image (b) Output of ILPF with D0=50 (c) Output of
ILPF with D0=80
Fig. 8

6.3.2 Butterworth Low Pass Filters (BLPF)

The Butterworth filter replaces the sharp cutoff of Ideal LPF by a smooth
cutoff. Frequency response of BLPF does not have a sharp transition between
pass band and stop band. It is more appropriate for image smoothing and
does not introduce ringing effect for lower order filters. Transfer function of
BLPF is given by

1
H ( u , v)  ,
1  D( u , v) / D 0 
2n

M N
where the cut off frequency or distance from the centre D 0   , , and
 2 2
1/ 2
 2
M  N 
2

the filter order is n, and D( u, v)   u     u   

 2  2  

Fig. 9 (a) and Fig. 9 (b) show the transfer function of BLPF. Fig. 9 (c) is the
plot of BLPF and Fig. 9 (d) is BLPF displayed as an image.

(a) Plot of BLPF (b) Plot of BLPF transfer function

162
Unit 6 Image Enhancement & Filtering-Frequency Domain

(c) Plot of BLPF (d) BLPF displayed as an image

Fig. 9
Transfer function of BLPF does not have sharp transition near the cut off. For
n  1 , the transition is very smooth. As the filter order increases, the transfer
function approaches towards ideal LPF. No ringing is visible on the image
filtered by BLPF for n  1. Noise is reduced and blurring is observed in all the
images. For n  2 , ringing is un-noticeable, but it can becomes more
significant for higher values of n. Fig. 10 shows the increasing effect of ringing
as n increases from 1 to 20.

Fig. 10: Spatial Representation of BLPF of order 1, 2, 5 and 20 and

Corresponding Intensity Profile

The output corresponding to the change in the values of D o and n are shown
in Fig. 11.

(a) Output of BLPF for D0= 30 (b) Output of BLPF for D0= 40

(c) Output of BLPF for n = 4, D0 = 30 (d) Output of BLPF for n = 20 D0 = 30

Fig. 11 163
Unit 6 Image Enhancement & Filtering-Frequency Domain
6.3.3 Gaussian Low Pass filters (GLPF)
Still a better variant of the low pass filter is the Gaussian Low Pass filter which
have smooth transition between pass band and stop band. It does not
introduce any ringing in the output image. The transfer function of GLPF is
given by
2
( u ,v ) / 2 2
H ( u , v)  e  D ,

where D( u , v) is the distance from the origin of Fourier transform, and  is the
measure of spread/dispersion of the Gaussian curve.

Larger the value of  , larger is the cut off frequency and the filter is milder.
Let   D 0 then transfer function is given by

2
( u ,v ) / 2 D0 2
H ( u , v)  e  D ,

where D 0 is the cut off frequency.

When D(u, v)  D 0 , the amplitude of transfer function is down to 0.607 of its

maximum value of 1.

Fig. 12 (a) is GLPF transfer function, Fig. 12 (b) is plot of GLPF and Fig. 12 (c)
is GLPF displayed as an image. Fig. 13 (a) to Fig. 13 (c) are GLP filtered
images. No ringing is observed in the output, but only blurring is visible. As the
cut off frequency increase, blurring reduces. No ringing in the output is a very
big advantage of GLPF. These filters can be used in situations where no
artifacts are desirable (eg. medical imaging). In medical imaging, GLPF is
preferred over ILPF/ BLPF.

(a) GLPF Transfer Function for Various Values of D (b) Plot of GLPF

(c) GLPF Displayed as an Image

Fig. 12 164
Unit 6 Image Enhancement & Filtering-Frequency Domain

(a) Output of GLPF for D0 = 10 (b) Output of GLPF for D0 = 300

(c) Output of GLPF with D0 = 50

Fig. 13
Ideal Butter worth Gaussian
Transfer H ( u , v)  H( u , v)  H ( u , v)  e  D
2
( u ,v ) / 2 D 2
Function
1, D( u , v)  D 0 1

1  D(u , v) / D 0 
24
0, D( u , v)  D 0
Applications Reduce noise Reduce noise Reduce noise
Problems Blurring Blurring, Ringing Blurring,
Ringing for higher order No ringing
filters ( n  2)

Let us discuss some of the applications of Low pass filters in frequency

domain.

6.3.4 Applications of Low Pass Filters

LPF are generally used as a preprocessing step before an automatic
recognition algorithm. It is also used to reduce noise in images. Few examples
are listed below.
1. Character Recognition: Input to an automatic character recognition
system is generally of poor quality. Input may contain noise due to
improper acquisition system or there may be gaps in the alphabets.
(broken alphabets). Because of these problems, character recognition
system fails to give expected results consistently. Hence, LPF is used as
a preprocessing step to blur the image. Blurring is used to bridge small
gaps in the alphabets. This is done using GLPF with D0 = 80. This
increases chances of getting correct result from automatic character
recognition system.
2. Object Counting: Object counting is to count the number of objects in an
image. The output of an object counting algorithm may give wrong output
because of poor quantity of input image. If there are small gaps in the
boundary of objects, automatic algorithm will not give expected results.
165
Unit 6 Image Enhancement & Filtering-Frequency Domain
Blurring is used to fill in small gaps in the boundary of objects which helps
in producing correct results.

3. Printing and Publishing Industry: Unsharp masking is used in

publishing industry to sharpen image where blurred version of an image is
subtracted from the image itself to get sharpen image.

4. “Cosmetic” processing is another use of low pass filter prior to printing.

Blurring is used to reduce the sharpness of fine skin lines and small
blemishes on human face. Smoothened images look very soft and
pleasing and face looks younger.

Try the following exercises.

E3) Give the formula for transform function of a Butterworth low pass filter.

E4) Explain and compare ideal low pass filter and Butterworth filter for image
smoothing.

E5) Explain smoothing frequency domain filters. What is ringing effect?

E6) Discuss the applications of image smoothing filters.

In the following section we will discuss sharpening filters.

6.4 IMAGE SHARPENING IN FREQUENCY

DOMAIN
In the Fourier transform of an image, high frequency contents correspond to
edges, sharp transition in grey levels and noise. Low frequency contents
correspond to uniform or slowly varying grey level values.

High pass filtering is achieved by attenuating low frequency components

without disturbing high frequency components. High pass filter (HPF) can also
be viewed as reverse operation of low pass filter. Transfer function of HPF is
given by

H h p ( u , v)  l  H l p ( u , v)

Here, H ( u , v) is the transfer function of a LPF. Thus

Smoothing  LPF  attenuates high frequency components

Sharpening  HPF  attenuates low frequency components

Here, we discuss only real and symmetric filters. Following sharpening filters
are discussed in this section:

1. Ideal high pass filter

2. Butterworth high pass filter
3. Gaussian high pass filter

High pass filters are used for enhancing edges. These filters are used to
extract edges and noise is enhanced, as a side effect.

166
Unit 6 Image Enhancement & Filtering-Frequency Domain
6.4.1 Ideal High Pass Filter (IHPF)
Transfer function of a 2D IHPF is given by

1, if D(u , v)  D 0
H (u , v)  
 0, if D(u , v)  D 0
Here, D 0 is the cut off frequency and D( u , v) is the distance from the origin of
the Fourier transform. Fig. 14 (a) and Fig. 14 (b) is the IHPF and its transfer
function respectively. Fig. 14 (c) is plot of IHPF and Fig. 14 (d) is IHPF as an
image. Note that the origin (0,0) is at the centre and not in the corner of the
image. The abrupt transition from 1 to 0 of the transfer function H ( u , v) cannot
be realized in practice. However, the filter can be simulated on a computer.
This filter sets to all frequencies inside the circle of radius D and passes all
frequencies above D0 without any attenuation. Ringing is clearly visible in the
output (Fig. 15 (b), and Fig. 16(c)) other than sharp edges and boundaries.
Output image looks very dark and dull as the high value DC component
G (0, 0) is eliminated.

(a) Plot of IHPF (b) Transfer function of IHPF

(c) Plot of IHPF (d) IHPF displayed as an image

Fig. 14

(a) Output of IHPF for D0= 50 (b) Output of IHPF for D0 = 60

Fig. 15 167
Unit 6 Image Enhancement & Filtering-Frequency Domain
6.4.2 Butterworth High Pass Filter (BHPF)
Butterworth filter does not have sharp transition between passband and stop
band. The slope depends on the order of the filter. Transfer function of BHPF
is
1
H (u, v)  2n
,
 D0 
1  
 D( u , v ) 
where n is the order of the filter, D 0 is the cut off frequency and D( u , v) is the
distance from the origin of Fourier transform.
Fig. 16 (a) and Fig. 16 (b) are BHPF transfer function and Fig. 16 (c) and
Fig. 16 (d) are plot and image display of BHPF.
Frequency response does not have a sharp transition as in the ideal HPF.
Thus, less distortion is seen in the output with no ringing effect even for
smaller values of cut off frequencies. This filter is more appropriate for image
sharpening than ideal HPF as there is no ringing in output.

Fig. 16(b) is the plot of GHPF for D 0  30, n  2, and Fig. 16 (c) GHPF
displayed as an image. Fig. 17(a) and Fig.17 (b) are the output of GHPF for D0
= 30 and 130 respectively for n  2 . It is clear from the output, as D0
increases, more and more power is removed from the output image. Thus,
output looks sharper for higher value of D 0 . Fig. 17(d) is the output for
D 0  30 n  20, ringing is clearly visible in the output. As n increases, ringing
in butterworth filter increases.

(a) Transfer function of BHPF (b) Transfer function of BHPF

(c) Plot of BHPF (d) BHPF displayed as an image

Fig. 16

(a) Output of BHPF with D0 = 130, n = 2 (b) Output of BHPF with D0= 30, n = 2
168
Unit 6 Image Enhancement & Filtering-Frequency Domain
Fig. 17
6.4.3 Gaussian High Pass Filter (GHPF)
Gaussian high pass filters have smooth transition between passband and
stopband near cutoff frequency. The parameter D is a measure of spread of
the Gaussian curve. Larger the value D 0 , larger is the cut off frequency.
Transfer function of GHPF is
 D 2 ( u , v)
H ( u , v)  1  e 2
,
2D 0
where D 0 is the cut off frequency and D( u , v) is the distance from origin of
Fourier transform.

(a) GHPF transfer function (b) Plot of GHPF

(c) GHPF displayed as an image

Fig. 18

Fig. 18(a) is GHPF Transfer function filter. Plot and image are displayed in
Fig. 18 (b) and Fig. 18 (c). Output in Fig. 19 is much smoother than previous
two filters with no ringing effect.

(a) Output of GHPF with D0=30 (b) Output of GHPF with D0=120
Fig. 19

Let us compare these three high pass filters in frequency domain filters in the
following table.
Table 1
Ideal Butterworth Gaussian
Transfer H ( u , v)  H ( u , v)  H ( u , v) 
function
1, D ( u , v)  D 0 1 D 2 (u , v)
 2n 1 e
0, D ( u , v)  D 0
2
 D  2D 0
1  
 D( u , v) 
Application Edge Edge enhancement Edge
enhancement enhancement
169
Unit 6 Image Enhancement & Filtering-Frequency Domain
Problems Ringing No Ringing No Ringing
Try the following exercises.

E7) How many types of high pass filters are there in frequency domain? List
them.
E8) Give the formula for transform function of a Gaussian high pass filter.

Now, Its time to discuss the concept of image degradation.

6.5 IMAGE DEGRADATION

Overall our objective is to improve image. For that it is important to
understand image degradation if we want to remove it. Degradations
are of three types

a) Noise
b) Blur
c) Artifacts

Let us define these one by one.

a) Noise is a disturbance that causes fluctuations in pixel values.

Pixel values show random variations and can cause very
disturbing effects on the image. Thus suitable strategies should be
designed to model and remove/ reduce noise. Original image is
shown in Fig. 20(a) and noisy image with added Gaussian noise is
shown in Fig. 20(b).

(a) Original (b) Noisy Image

Fig. 20

b) Blur is a degradation that makes image less clear.This makes

image analysis and interpretation very difficult. Motion blur is a
very common cause of blurring where blur occurs due to the
movement of object or camera. Fig. 21 (a) shows original image
and Fig. 21 (b) shows blurred image.

170
Unit 6 Image Enhancement & Filtering-Frequency Domain

(a) Original Image (b) Blurred image

Fig. 21

c) Artifacts or distortions are extreme intensity or color fluctuations

that can make image meaningless. Distortions involve geometric
transformations such as translation, rotation or change in scale.

Now, the question arises that what are the sources which contribute to
image degradation. Image degradation (as shown in Fig. 22) can
happen due to

a) Sensor Distortions: Involves quantization, sampling, sensor noise,

spectral sensitivity, de-mosaicking, non linearity of sensor etc.
b) Optical Distortions: are geometric distortion, blurring due to
camera mis-focus.
c) Atmospheric Distortions: are haze, turbulence etc.
d) Other Distortions: Low illumination, relative motion between object
and camera etc.

Fig. 22: Typical Degradation Sources

The processes that are used to remove degradation are mainly image
enhancement and image restoration.

Restoration is the process of inverting a degradation using knowledge

about its nature, whereas enhancement is a process that aims to
improve ‘bad’ quality image so that it looks better. Restoration is
distinguished from enhancement, as degradation can be considered as
an external influence. Table 2 lists the differences between
enhancement and restoration.
171
Unit 6 Image Enhancement & Filtering-Frequency Domain
Table 2: Enhancement v/s Restoration

Enhancement Restoration
1. It gives better visual representation It remove effects of sensing
environment
2. No model required Mathematical model of
degradation
3. It is a subjective process It is an objective process
4. Contrast stretching, histogram Inverse filtering, wiener filtering,
equalization etc are some denoising are some restoration
enhancement techniques techniques.

Try the following exercises.

E9) What are the factors that can cause image degradation.

In this section we will discuss image degradation/restoration model.

6.6 IMAGE DEGRADATION/RESTORATION

MODEL
Consider the block diagram given in Fig. 23 shows the block diagram of
degradation/restoration model. Degradation function h (x, y) and noise
n(x, y), operate on input image f ( x, y) to generate a degraded and noisy
image g(x, y).

Fig. 23: Block diagram of degradation/restoration model

The notations used in the block diagram are

f ( x, y)  original image
h (x, y)  degradation function
n (x, y)  additive noise
g(x, y)  degraded and noisy image
f̂ ( x , y)  restored image

The objective of restoration process is to estimate f̂ ( x , y) from the

degraded version g( x, y), when some knowledge of degradation
function H and noise n is available. The degraded image g( x, y) as
shown in Fig. 24 can be expressed mathematically as

g(x, y)  h (x, y) *f (x . y)  n (x, y)

172
Unit 6 Image Enhancement & Filtering-Frequency Domain
This equation is in spatial domain and * represents convolution
operation. An equivalent frequency domain representation is graphically
shown in Fig. 25 and expressed as

G(u, v)  H(u, v) F(u, v)  N(u, v)

Here G(u, v)  F[g( x, y)]

H(u, v)  F[h (x, y)]
F(u, v)  F[g(x, y)]

N(u, v)  F[n(x, y)]

F(u, v)  H 4 (u, v)[G (u, v)  N(u, v)]

Fig. 24: Image Degradation Model (Spatial Domain)

Fig. 25: Image Degradation Model (Frequency Domain)

Restored image can be obtained by the above equation. The problems

in implementing this equation are

1) The noise N is unknown. Only the statistical properties of noise

can be known.
2) The operation H is singular or ill posed. It is very difficult to
estimate H.
173
Unit 6 Image Enhancement & Filtering-Frequency Domain
Try an exercise.

E10) Explain in detail an image degradation model.

In the following section, we shall discuss noise models in detail.

6.7 NOISE MODELS

Major source of noise in digital images is during image acquisition. Non-
ideal image sensors and poor quality of sensing elements contribute to
majority of noise. Environmental factors such as light conditions,
temperature of atmosphere, humidity, other atmospheric disturbances
also account for noise in images. Transmission of image is also a
source of noise. Images are corrupted with noise because of
interference in the channel, lightning and other disturbances in wireless
network. Human interference also plays a part in addition of noise in
images.
Properties of Noise

Spatial and frequency characteristics of noise are as follows:

1) Noise is assumed to be ‘white noise’ (it could contain all possible
frequency components), as such, fourier spectrum of noise is
constant.
2) Noise is assumed to be independent in spatial domain. Noise is
‘uncorrelated’ with the image, that is, there is no correlation
between pixel value of image and value of noise components.
The spatial noise descriptor is the statistical behavior of the intensity
values in the noise component. Noise intensity is considered as a
random variable characterized by a certain probability density function
(PDF).

Restoration techniques are oriented towards modeling the degradation

(noise in this case) and restore an image to the original state. Most
types of noise are modeled as known PDFs. Based on the estimated
parameters from the noisy image, a particular noise PDF is chosen.
Noise models are divided into two categories:
a) Noise which is independent of spatial location: Gaussian,
Rayleight, Gamma, Exponential, Uniform noise are examples of
this category.
b) Noise which is dependent on spatial location: Periodic noise is
example of this type of noise.
Following are the most commonly occurring noise models:

Gaussian Noise
Gaussian noise model is most frequently used in practice. The PDF of a
Gaussian random variable ‘z’ is given by
174
Unit 6 Image Enhancement & Filtering-Frequency Domain
2
( z  )
1 
p( z )  e 22
, (2)
2

where z  intensity/grey level value

  mean (average) value of z

  standard deviation

Plot of p(z) with respect to z is shown in Fig. 26. 70% of its values are
in the range [(  ), (  )] while 95% of the values are in the range
[(  2), (  2)] . DFT of Gaussian (normal) noise is another
Gaussian process. This property of Gaussian noise makes it most
often used noise model. Some examples where Gaussian model is the
most appropriate model are electronic circuit noise, sensor noise due to
low illumination or high temperature, poor illumination.

Fig. 26: PDF of Gaussian Noise Model

Gaussian noise is useful for modeling natural processes which

introduce noise (e.g. noise caused by the discrete nature of radiation
and the conversion of the optical signal into an electrical one –
detector/shot noise, the electrical noise during acquisition – sensor
electrical signal amplification, etc.).

Rayleigh Noise

Radar range and velocity images typically contain noise that can be
modeled by the Rayleigh distribution. Rayleigh distribution is defined by

2
2
( z a )

 (z  a ) e b za
p(z )   b , (3)
0 za


Mean density is given as   a   b / 4 and variance is given by

b ( 4  )
2  .
4
175
Unit 6 Image Enhancement & Filtering-Frequency Domain
Plot of PDF is shown in Fig 27. As it is clear that the curve doesn’t start
from origin and is not symmetrical with respect to the centre of the
curve. Thus, Rayleigh density is useful for approximating skewed (non-
uniform) histograms. This is mainly used in range imaging.

Fig. 27: PDF of Rayleigh Noise

Erlang (Gamma) Noise

Erlang noise is given by

 a b z b 1  nz
 e z0
p( z )   ( b  1) ! , (4)
0 z0

b
Where a and b are positive integers, mean density is given by  
a
b
and variance is  2  .
a2

When the denominator is a gamma function, the PDF describes the

gamma distribution. Plot is shown in Fig. 28.

Fig. 28: PDF of Erlang Noise

Uniform Noise

Uniform noise is specified as

176
Unit 6 Image Enhancement & Filtering-Frequency Domain
 1
 if a  z  b
p(z)   b  a
0 otherwise

Then mean and variance of uniform noise is given by

ab (b  a ) 2
 , 2 
2 12

Fig. 29: PDF of Uniform Noise.

Fig. 29 shows the plot of PDF of uniform noise. Uniform noise is least
used in practice.

Impulse (Salt and Pepper) Noise

Impulse (salt and pepper) noise is specified as

Pa za
p(z )   , (6)
Pb zb

Fig. 30 shows the plot of PDF of impulse noise. If b  a, intensity (grey

level) ‘b’ will appear as a light dot on the image and ‘a’ appears as a
dark dot. This is a ‘bipolar’ noise, If Pa  0 or Pb  0  unipolar noise
Generally, a and b values are saturated (very high or very low value),
resulting in positive impulses being white (salt) and negative impulses
being black (pepper). If Pa  0 and Pb exists, this is called ‘pepper
noise’ as only black dots are visible as noise. If Pb  0, only Pa exists,
this is called ‘salt noise’ as only white dots are visible on the image as
noise.

Impulse noise occurs when quick transitions happen, such as faulty

switching takes place. Noise parameters are generally estimated based
on histogram of small flat area of noisy image.

The salt & pepper noise is generally caused by malfunctioning of

camera’s sensor cells, by memory cell failure or by synchronization
errors in the image digitizing or transmission.

177
Unit 6 Image Enhancement & Filtering-Frequency Domain

Fig. 30 PDF of Uniform noise.

Fig. 31 shown an example of impulse noise with p  0.1 added to input

image and togenerate a noisy image g. Noise level p  0.1 means that
approximately 10% of pixelsare contaminated by salt or pepper noise
(highlighted by box)

Fig. 31: Numerical Example of Adding Impulse Noise with P = 0.1.

Fig. 32 shows the flower image with different types of noise. It is very
easy to identify the effect of different types of noise on the images.
Fig. 32 (a) shows original image, Fig. 32 (b) shows image with
Gaussian noise. Fig. 32 (c) shows image with salt and pepper noise and
Fig. 32 (d) shows image with uniform noise. The amount of noise added
can also vary. If the amount of noise added is more, it becomes very
difficult to remove it.

(a) Original image (b) Image with Gaussian noise

178
Unit 6 Image Enhancement & Filtering-Frequency Domain

(c) Image with salt & pepper noise (d) Image with uniform noise

Fig. 32

Let us discuss an important type of noise.

Periodic noise is a spatially dependent noise. During Image acquisition,

electrical or electromechanical interference may cause such type of
periodic noise. A strong periodic noise can be seen in the frequency
domain as equi-spaced dots at a particular radius around the centre
(origin) of the spectrum. Fig. 33 shows image with periodic noise.

Fig. 33: Image with periodic noise

Now, you may like to try the following exercises.

E11) What is noise? How noise can be eradicated?

E12) Explain the types of noises based on its probability distribution.

In the following section, we will discuss restoration in the presence of

noise only-spatial filtering.

6.8 RESTORATION IN THE PRESENCE OF NOISE

ONLY – SPATIAL FILTERING
If H is an identity operator and degradation is only due to additive noise,

179
Unit 6 Image Enhancement & Filtering-Frequency Domain
g(x, y)  f (x, y)  n( x, y)
G(u, v)  F(u, v)  N(u, v)

As noise is unknown, f (x, y)  g(x, y)  n(x, y) is not realistic. Thus,

spatial filtering is used when additive random noise is present. Mean
and median filters are used for noise removal. Band reject and band
pass filters are used for periodic noise removal.

6.8.1 Mean Filters

Spatial Smoothing concepts are explained in earlier unit 4 of this
course. Now, Consider Fig. 34, S x , y  Sub image window of size m  n
centred at (x, y). Fig. 35 shows 3 3 and 5 5 sub images. Mean filter
computes average value of the corrupted image g( x, y) in the area
defined by S x , y

1
f̂ ( x, y)   g(s, t )
mn s , tSxy

Such filter smooths local variations in an image, thus reducing noise

and introducing blurring. This filter is well suited for random noise like
Gaussian, uniform noise.

Fig. 34: Graphic Illustration of Sub-Image in An Image

(a) S( x, y) of size 3 3 (b) S( x, y) of size 5 5

Fig. 35: Sub-Image of Various Sizes

Thus, new value at ( x, y) in image in Fig. 36 is

180
Unit 6 Image Enhancement & Filtering-Frequency Domain
{g(s, t )}  19  [30  10  20  10  250  25  20  25  3]  46.7  47

Fig. 36: Example of Mean Filtering.

Let us apply this in the following example.

Example 1: Show effect of 3 3 mean filter on a simple image in

Fig. 37 (a) and Fig. 38 (a)

Solution: As explained in Unit 4, a 3 × 3 mean filter is overlapped with

image and output for that particular pixel is derived and then filter centre
is moved to the next pixel. We generate a lower size image because the
filter mask doesn’t overlap fully on first and last row and column.

(a) (b)

Fig.37: Input and Output Image.

Mean filter removes random noise by introducing blurring. Random

noise value of 20 in Fig. 37 (a) is removed from the resultant image (b).
But, importantly the edge is also diluted and blurred. In the second
image Fig. 38 (a), which has fairly constant values, the pixel values
remain more or less same in Fig. 38 (b).

(a) (b)

Fig.38 Input and Output Image.

181
Unit 6 Image Enhancement & Filtering-Frequency Domain
Gaussian noise is added to the input image Fig. 39 (a). 3  3, 5  5 and
7  7 mean filters are applied to the noisy image and the output images
are displayed in Fig. 39 (b), (c), (d). As it is clear from the output, 3 3
filter (Fig. 39 (b)) does not remove the noise completely. Noise is still
seen in the image but blurring is less. In 5  5 (Fig. 39 (c)) filtering more
noise is removed but image gets blurred. In 7  7 (Fig. 39(d)), too much
blurring is seen in the output.

(a) Original Image (b) Filtered Image by

Mean Filter 3 3

(c) Filtered Image by (d) Filtered Image by

Mean Filter 5  5 Mean Filter 7  7.

Fig. 39

Let us discuss median filter.

6.8.2 Median Filters

Median filter replaces the pixel value by the median of the pixel values
in the neighbourhood of the centre pixel ( x, y) . The filtered image is
given by

f̂ ( x , y)  median{g (s, t )}
( s , t )Sxy

Fig. 40 shows the procedure of applying 3 3 median filter on an image.

As impulse noise appears as black (minimum) or white (maximum) dots,
taking median effectively suppresses the noise.

182
Unit 6 Image Enhancement & Filtering-Frequency Domain
Thus, median filter provides excellent results for salt and pepper noise
with considerably less blurring than linear smoothing filter of the same
size. These filters are very effective for both bipolar and unipolar noise.
But, for higher noise strength, it affects clean pixels as well and a
noticeable edge blurring exists after median filtering.

Fig. 40: Example of Median Filtering.

To understand this clearly, see the following example.

Example 2: Show the effect of 3 3 median filter on a simple image
given in Fig. 41(a) and Fig. 41(b).

(a) Input image (b) Output image

Fig. 41
Solution: When a 3  3 median filter is implemented, all 9 pixels around
‘hotspot’ are arranged in ascending/ descending order. Center pixel is
taken as output and center pixel is replaced by the output. This process
is repeated for the entire image.

(a) Input image (b) Output image

Fig. 42
***
It is clear from Example 2, Fig. 41 (a) and Fig. 41 (b) that if noise
183
Unit 6 Image Enhancement & Filtering-Frequency Domain
strength is low in noisy image, output is completely clean. But if noise
strength is more (more number of noisy pixels in the image), output is
not completely noise free as can be seen in Fig. 42 (a) and Fig. 42 (b).

Let us see the effect of the median filter.

Salt and pepper noise is added to an image given in Fig. 43 (a), noisy
image is shown in Fig. 43 (b). 3 3 mean filter and 5 × 5 median filter is
applied on it. As it is clear from the result, (Fig. 43 (c) and Fig. 43 (d))
mean filter is not effective in removing salt and pepper noise. But
median filter completely removes salt and pepper noise without
introducing any blur.

(a) Original Image (b) Noisy Image

(c) Filtered Image with Mean Filter (d) Filtered Image with Media Filter
Fig. 43
Salt and pepper noise with density of 0.3 is added to an image. The
noisy image (Fig. 44 (a)) is filtered using 3  3, 5  5 and 7  7 , median
filter. The results in Fig. 44 (b), (c), (d) show that 3 3 median filter is
unable to remove the noise completely as the noise density is high. But
5 5 and 7  7 median filters remove noise completely but some
distortions are seen specially in Fig. 44 (d).

(a) Noisy Image (b) Filtered Image with 3 3

Median Filter
184
Unit 6 Image Enhancement & Filtering-Frequency Domain

(c) Filtered Image with 5  5 (d) Filtered Image with 7  7

Median Filter Median Filter
Fig. 44

Now, in the following section, we shall discuss noise reduction.

6.9 PERIODIC NOISE REDUCTION

Periodic noise is spatially dependent noise and it occurs due to
electrical or electromagnetic interference. It gives rise to a regular noise
pattern in an image. Frequency domain (fourier domain) techniques are
very effective in removing periodic noise. Basic steps in frequency
domain filtering remain same as discussed above. Here, we are
discussing two frequency domain filters; namely band reject filter and
band pass filter.

6.9.1 Band Reject Filter

Removing periodic noise from an image involves removing a particular
range of frequencies from the image. Transfer function of ideal band
reject filter is

 W
1 D( u , v)  D 0 
2

 W W
H (u , v)  0 D0   D( u , v )  D 0  , (7)
 2 2
 W
1 D( u , v)  D o 
 2

where W is the width of the band (band width), D 0 is its radial centre
and D(u, v) is the distance from the origin and is given by

 M
1/ 2
 N 
1/ 2

D(u , v)   u     v   
 2  2  

D(u, v) is the distance measured from the point (u, v) to the centre
M N M N
 ,  . If size of an image is M N, then the centre is at  ,  .
 2 2  2 2

185
Unit 6 Image Enhancement & Filtering-Frequency Domain

Fig 45: Frequency response of ideal band reject filter

Transfer function of butter worth band reject filter of order ‘n’ is given by
1
H ( u , v)  2n
(8)
 D( u , v) W 
1  2 2
 D (u , v)  D 0 

Gaussian band reject filter is given by

1  D 2 ( u , v )  D 20 
 
2  D ( u , v ) W 
H(u , v)  1  e (9)

Fig. 46 gives the plots of ideal, butterworth and gaussian band reject
filters.

(a) Ideal (b) Butterworth (order 1) (c) Gaussian

Fig. 46: Plots of Band Reject Filters

6.9.2 Band Pass Filter

Band pass filter performs just opposite to band reject filter. The transfer
function of band pass filter can be obtained from band pass filters.

H bp (u , v)  1  H br (u , v),
(10)

Where, H bp is transfer function of band pass filter and H br is transfer

function of band reject filter.

Ideal band pass filter is given by

 W
0 D(u , v)  D 0  2

 W W
H (u , v)  1 D 0   D( u , v)  D 0  , (11)
 2 2
 W
0 D(u , v)  D 0  2

186
Unit 6 Image Enhancement & Filtering-Frequency Domain
Where D(u, v) is the distance from origin, W is the band width
D0 is the radial centre or the cut off frequency.

Fig. 47 shows the transfer function of ideal band pass filter.

Butterworth band pass filter of order ‘n’ is given by

H( u , v)  1  H( u , v) butterworth band reject (12)

2n
 D( u , v) W 
 2 
1 D (u , v)  D 02 
H(u , v)  1  2n
  2n
(13)
 D( u , v ) W   D( u , v) W 
1  2 2
1  2 2 
 D ( u , v)  D 0   D ( u , v)  D 0 

(a) (b) (c)

Fig. 47: Frequency Response of Ideal Band Pass Filter

Similarly, Gaussian band pass filter is given by

H( u , v)  1  H( u , v) gaussian band pass

 
2
 D 2 ( u , v )  D02 
1 / 2  
 1  1  e  
D ( u , v ) W 
 
 
 D 2 ( u , v )  D 02  2
1 / 2  
 D ( u , v ) W 
e

6.10 ESTIMATION OF DEGRADATION FUNCTION

In order to restore the image, we need to estimate the degradation
function. There are three principal ways to estimate the degradation
function to be used in restoration:

1) Observation
2) Experimentation
3) Mathematical modelling

Once the degradation function has been estimated, then, restoration is

a de convolution process also called Blind Deconvolution.
187
Unit 6 Image Enhancement & Filtering-Frequency Domain
6.10.1 Observation
In restoration using observation, we assume that an image g(x, y) is
degraded with an unknown degradation function H. We try to estimate H
from the information gathered from the image itself. For example, in
case of burred image, a small rectangular section of image containing a
part of object and background is taken (Fig. 48). To reduce the effect of
noise, the chosen part should be such that it shows presence of a
strong signal. We try to un-blur that sub-image manually as much as
possible and generate f s (x, y) from g s ( x, y).

Fig. 48: Estimation by image observation

Here, g(x, y) is the original sub image

f̂ s ( x , y) is the restored version of g s (x, y)

Thus, degradation can be estimated for the sub image by

g ( u , v)
H s ( u , v)  s
F̂s ( u , v)
From the characteristics of H s (u, v), we try to deduce the complete
degradation function H s (u , v) based on the assumption of position
invariance. For example, if H s (u , v) has a Gaussian shape, we can
construct H(u, v) on a larger scale with the same (Gaussian)shape. This
is a very involved process and is used in very specific situations.

6.10.2 Experimentation
It is possible to estimate the degradation function accurately if the
equipment used to acquire the degraded image is available. The
processes is shown in Fig. 49.

188
Unit 6 Image Enhancement & Filtering-Frequency Domain

Fig. 49: Estimation by experimentation

We now list the steps performed for estimation.

Step 1: First step is to adjust the equipment by varying the system

setting such that the image obtained is similar to the degraded
image that needs to be restored.

Step 2: Second step is to obtain the impulse response of the

degradation by imaging an impulse using the same system
setting, since a linear space – invariant system is completely
characterized by its impulse response. An impulse is simulated
by a maximally bright dot of light. As shown in the Fig. 2,
impulse response is given by

G ( u , v)
H( u , v )  ,
A
where G(u, v)  DFT[g(x, y)]  DFT [degraded impulse], and A is the
constant describing the strength of the impulse.

6.10.3 Modelling
Modelling is used to estimate the degradation function. Scientists have
studied several environmental conditions and other processes which
cause degradation, and have formulated several fundamental models
for the degradation functions. Degradation model based on atmospheric
turbulence blur is given as
2
 v2 )5 / 6
H ( u , v)  e  k ( u
2
 y2 )5 / 6
h ( x , y)  e  k ( x ,
189
Unit 6 Image Enhancement & Filtering-Frequency Domain
where k is a constant that depend on the nature of blur. Various values
used for the constant k along with their type of turbulence are given as

k  0.0025 for server turbulence

k  0.001 for wild turbulence
k  0.00025 for low turbulence

This is commonly used in remote sensing and axial imaging

applications. Degradation model for uniform out of focus blur (optical
blur).

1 L L
 ;  x 0 , y0 
h ( x , y)   L2 2 2 (1)
0 ; otherwise

T
H ( u , v)  sin (ua  vb) e  j( ua  vb) (2)
(ua  vb)

Blur can be due to motion (camera and object moving with respect to
each other). With suitable values of T, a and b , blurred image can be
generated using this transfer function. Fig. 50(a) shows original image
and Fig. 50(b) shows blurred image.

(a) Original Image (b) Blurred Image

Fig. 50
Try an exercise.

E13) What are the different methods of estimation of image

degradation function?

In the following section, we shall discuss inverse filtering.

6.11 INVERSE FILTERING

Inverse filter is also known as reconstruction filter. Deblurring is very
important in restoration applications because blurring is visually
annoying. It is bad for analysis, and de-blurred images have plenty of
190
Unit 6 Image Enhancement & Filtering-Frequency Domain
applications. Applications include astronomical imaging, law
enforcement (identifying criminals), biometrics etc. In this unit we will
discuss Inverse filtering, pseudo-inverse filtering and Wiener filtering for
deblurring.

Fig. 51: Block diagram of degradation/restoration model

As in the absence of noise, degradation model becomes,

G(u, v)  F(u, v) H(u, v) (3)

Simplest approach to restoration is direct inverse filtering, where we can

compute an estimate F̂( u, v), of the transform of the original image
simply by dividing transform of the degraded image by the degradation
function.

G ( u , v)
F̂(u, v)  (4)
H ( u , v)

In presence of noise, degradation model as shown in Fig. 51 becomes

G(u, v)  F(u, v) H(u, v)  N(u, v) (5)

After applying inverse filtering

F̂(u , v)  H R ( u , v) G (u , v) (6)

Substituting the values of H R ( u , v) and G(u, v), we get

1
F̂(u, v)  F(u, v)H(u, v)  N(u, v)
H ( u , v)
F(u, v) H(u, v) N(u, v)
 
H ( u , v) H ( u , v)
N( u , v)
 F(u, v)  (7)
H ( u , v)

Thus, in case of noisy degraded images, output is also noisy.

N( u , v)
If H(u, v)  0 
H ( u , v)

noise is amplified and it dominates the output.

Limitations of inverse filtering are:

191
Unit 6 Image Enhancement & Filtering-Frequency Domain
1) It is an unstable filter
2) It is sensitive to noise. In practice, inverse filter is not popularly
used.

To remove the limitations of inverse filter, pseudo inverse filters are

used. Pseduo Inverse filter is defined as,

 1
 H ( u , v) 
H R ( u , v)   H ( u , v)
0 H ( u , v) 

where,  is a small value.

Pseudo-inverse filters eliminates the first problem of inverse filters

(un-stability).

As H(u, v)  0, H R (u, v)  0.

Hence, it does not allow H R ( u , v )  .

It is a stable filter. This filter also has a problem of noise amplification.

Inverse filtering is applied to the image blurred with a Gaussian
(Fig. 52(b)). The output image (Fig. 52(c)) is very close to the original
image (a). Then same inverse filtering is used when the blurred image
is subjected to additive noise with different strengths. Outputs are
shown in Fig. 52(d), (e) and (f). As noise increases, the output of filter
goes down.

(a) Original Image (b) Image Blurred with a Gaussian

(c) Inverse Filter Applied (d) Inverse Filter Applied to

to Noiseless Blurred Blurred Image Plus

192
Unit 6 Image Enhancement & Filtering-Frequency Domain

(e) Inverse Filter Applied (f) Inverse Filter Applied to

to Blurred Image Plus Noise (0.1) Blurred Plus Noise (0.5)

Fig. 52
Try an exercise.

E14) Explain in brief the inverse filtering approach and its limitations in
image restoration.

In the following section, we shall discuss wiener filter.

6.12 WIENER FILTER

Wiener filter is also known as minimum mean square error. This
approach includes both the degradation function and power spectrum of
noise characteristics in developing the restoration filter. Wiener filter
restores the image in the presence of blur as well as noise.

This method is founded by considering image and noise as random

variables and objective is to find as estimate f̂ of the uncorrupted image
f such that the mean square error between them is minimized. This
error is given by

e 2  E{(f  f̂ ) 2 }), (9)

where, E{.} is the expected value of the argument. Noise and image are
assumed to be uncorrelated. Filter transfer function is given by
Sfg (u, v)
H R ( u , v)  , (10)
Sgg (u, v)

where, S fg (u , v) is the power spectral density of recovered image and

noisy image and Sgg (u , v) is the power spectral density of noisy image

Sfg (u, v) H * (u, v) Sff (u, v)

H R ( u , v)   2
, (11)
Sgg (u, v) H(u, v) Sff (u, v)  Snn (u, v)

where, H(u, v)  degradation function.

2
H( u , v)  H( u , v) H * ( u , v)
193
Unit 6 Image Enhancement & Filtering-Frequency Domain
S nn (u , v )  power spectral density of noise
S ff ( u , v)  power spectral density of undergraded image.

If S ff ( u , v ), S nn ( u , v) and H(u, v) is known H R ( u , v) is completely known.

Wiener filter works very well for specific applications and is not suitable
for general images. For example, if a wiener filter H R ( u , v) is working
well for faces, same filter would not work for landscapes etc. Now we
discuss several cases to test wiener filter.
Case 1: When, there is no noise S nn (u , v)  0

H * (u , v) Sff ( u , v)
H R ( u , v)  2
H (u , v) Sff ( u, v)  0
1
H R ( u , v)   inverse filter
H ( u , v)

Thus, if there is no noise, Wiener filter = Inverse filter.

Case 2: Noise is present, but there is no degradation, H(u, v)  1

1.Sff (u , v)
 H R ( u , v) 
1.Sff (u , v)  S nn (u , v)
S (u , v) / S nn (u , v)
 ff
Sff (u , v)
1
Snn (u , v)
SNR

SNR  1
Sff (u , v) PSD of signal
SNR  Signal to noise ratio  
S nn (u , v) PSD of noise
SNR
 SNR  1 H R (u, v)  1
SNR
a) If signal to noise ratio is high
Thus, if SNR high, wiener filter acts like pass band and allows all the
signal to pass through without any attenuation.

b) If SNR  1, if Signal to noise ratio is low, then

SNR
H R ( u , v)   SNR
1
 a very low value
0

Thus, if SNR is low and noise level very high, H R ( u , v )  0, acts as a

stop band for signal and doesn’t allow signal to pass, thus attenuating
noise. If noise is high in the signal, wiener filter reduces it after filtering.

194
Unit 6 Image Enhancement & Filtering-Frequency Domain
M 1 N 1

 F(u, v)
2

SNR  u 0 v 0
M 1 N 1

 N(u, v)
2

u 0 v 0

SNR gives a measure of the level of information bearing signal power

(i.e. of the original, undegraded image) to the level of noise power.
Images with low noise tend to have high SNR and conversely, the same
image with higher noise level has a high SNR.
The mean square error is given by

1 M 1 N 1
MSE  [f (x, y)  f̂ (x, y)]2
MN x 0 y 0

Here, f (x, y) is the original image and f̂ ( x, y) is the restored image.

Wiener filter is also called minimum mean square error (MMSE) or

Least square (LS) filtering because it minimizes the error between the
image and its estimate.

H * ( u , v) Sff (u , v)
H R ( u , v)  2
H (u , v) Sff ( u , v)  S nn ( u , v)

H * (u , v)

2 S (u , v)
H ( u , v)  nn difficult to estimate
S ff ( u , v)

Wiener filter in this form is not very useful, as it is difficult of estimate

noise power spectrum S nn and undergraded image power spectrum Sff ,

S nn (u , v)  difficult to estimate

S ff (u , v)  difficult to estimate

To solve this, we can do an approximation

S nn (u , v)
 k (approximated by constant k)
Sff (u , v)
H * ( u , v)
H R (u , v)  2
H (u , v)  k
And estimated restored image

F̂(u, v)  H R (u , v) G (u , v)

 H * ( u , v) 
 2
 G (u , v)
 H ( u , v)  k 

 1 H ( u , v) 
2

 *  G (u , v)
H ( u , v) H ( u , v ) 2  k 
 
195
Unit 6 Image Enhancement & Filtering-Frequency Domain
k is chosen experimentally and iteratively for best results. In Fig. 53,
small noise is added to a blurred image, which is restored by wiener
filter in Fig. 53(b).If the amount of added noise is increased Fig. 53(c),
the restored image by wiener filter (Fig. 53(d)) is not good. Thus, it is
apparent that the wiener filter only works well when the noise is small.

(a) Blurred Image with Small (b) Image Restored by

Additive noise by Wiener Filter

(c) Blurred Image with Increase (d) Image Restored by

Additive Noise Wiener Filter

Fig. 53

(a) Original Image (b) Blurred Image

(c) Restored image

Fig.54: Applying Wiener Filter

Image is blurred using linear motion = 15, angle = 5 shown in Fig. 54(b).
Wiener filter is used to deconvolve the blurred image. The output
(Fig. 54(c)) is not clear as the wiener filter does not use any prediction
about noise density.

Now, try an exercise.

E15) Discuss the minimum mean square error (Wiener) filtering. 196
Unit 6 Image Enhancement & Filtering-Frequency Domain

Now, we summarise what we have studied in the unit.

6.13 SUMMARY
In this unit, we have discussed the following points.
1. Image characteristics in frequency domain
2. Filtering in frequency domain
3. Basic steps of frequency domain filtering
4. Various low pass and high pass filters
5. Various image smoothing filters in frequency domain
6. Various image sharpening filters in frequency domain
7. Sources of degradation.
8. Difference between enhancement and restoration.
9. Image degradation/restoration model.
10. Various types of noises with their pdfs.
11. Mean and median filters for noise reduction
12. Band reject and band pass filters for periodic noise reduction.
13. Methods of estimation of degradation function.
14. Inverse filtering.
15. Wiener filtering.

6.14 SOLUTIONS AND ANSWERS

E1)

M N
u and v  .
2 2
1. Multiply input image f ( x , y) by (1) x  y to centre the transform to
2. Compute F( u , v) , Fourier transform of the output of step 1.
3. Multiply filter function H ( u , v) to F( u , v) to get G ( u , v) .
4. Take inverse Fourier transform of G ( u , v) to get g ( x , y) .
5. Take the real part of g ( x , y) to get g r ( x , y)
6. Multiply the result of step 5 by (1) x  y to shift the centre back to
origin and enhanced image is generated.
E2) Image enhancement can be done very effectively in frequency
domain. High frequency noise, undesirable breakages in the edges
and other imperfections can be taken care by filtering in frequency
domain. Low pass and high pass filters are implemented with ease
and perfection in frequency domain.
197
Unit 6 Image Enhancement & Filtering-Frequency Domain
1
E3) H ( u , v)  ,
1  D( u , v) / D 0 
2n

where D 0  Cut off frequency or distance from the centre n  filter

M N
order  , 
 2 2
E4)
Ideal Butterworth Gaussian
Transfer H ( u , v)  H ( u , v)  H ( u , v)  e  D
2
( u ,v ) / 2 D2
function
1, D ( u , v)  D 0 1

1  D(u , v) / D 0 
24
0, D ( u , v)  D 0
Application Reduce noise Reduce noise Reduce noise
Problems Blurring Ringing Blurring, Ringing for Blurring no ringing
higher order filters

E5) Smoothing filters are low pass filters (LPF). Edges, sharp transitions
and noise in the grey levels contribute to high frequency contents in an
image. A low pass filter only passes low frequency and blocks the high
ones. It removes noise and introduces blurring as a side effect in the
image.

Ringing is undesirable and unpleasant lines around the objects present

in the image. As the cut of frequency D 0 increases, effect of ringing
reduces. Ringing is a side effect of ideal lpf.

E6) LPF are generally used as a preprocessing step before an automatic

recognition algorithm. It is also used to reduce noise in images. Few
examples are listed below.

Character Recognition, Object counting, Printing and publishing

industry, “Cosmetic” processing etc.

E7) The sharpening filters are listed as follows:

1. Ideal high pass filter

2. Butterworth high pass filter
3. Gaussian high pass filter

High pass filters are used for enhancing edges. These filters are used
to extract edges and noise is enhanced, as a side effect.

E8) Gaussian high pass filters have smooth transition between passband
and stop band near cut off frequency. The parameter D is a measure
of spread of the Gaussian curve. Larger the value D 0 , larger is the cut
off frequency. Transfer function of GHPF is

 D 2 ( u , v)
H ( u , v)  1  e ,
2D 02
198
Unit 6 Image Enhancement & Filtering-Frequency Domain
where D 0  cut off frequency and D( u , v) is the distance from origin of
Fourier transform.

E9) Image degradation can happen due to

a) Sensor distortions: Involves quantization, sampling, sensor

noise, spectral sensitivity, de-mosaicking, non linearity of
sensor etc.
b) Optical distortions: are geometric distortion, blurring due to
camera mis-focus.
c) Atmospheric distortions: are haze, turbulence etc.

d) Other distortions: Low illumination, relative motion between

object and camera etc.

E10) Fig shows the block diagram of degradation/restoration model.

Degradation function h ( x, y) and noise n( x, y), operate on input
image f ( x, y) to generate a degraded and noisy image g(x, y).

Fig.: Block diagram of degradation/restoration model

f ( x, y)  original image
h (x, y)  degradation function
n (x, y)  additive noise
g(x, y)  degraded and noisy image
f̂ ( x , y)  restored image

E11) Noise is a disturbance that causes fluctuations in pixel values.

Pixel values show random variations and can cause very
disturbing effects on the image. Thus suitable strategies should
be designed to model and remove/ reduce noise. Major source of
noise in digital images is during image acquisition. Non-ideal
image sensors and poor quality of sensing elements contribute to
majority of noise. Environmental factors such as light conditions,
temperature of atmosphere, humidity, other atmospheric
disturbances also account for noise in images. Transmission of
image is also a source of noise. Images are corrupted with noise
because of interference in the channel, lightning and other
disturbances in wireless network. Human interference also plays
a part in addition of noise in images.

Properties of Noise
199
Unit 6 Image Enhancement & Filtering-Frequency Domain

Spatial and frequency characteristics of noise are as follows:

1) Noise is assumed to be ‘white noise’ (it could contain all

possible frequency components), as such,Fourier spectrum of
noise is constant.

2) Noise is assumed to be independent in spatial domain. Noise

is ‘uncorrelated’ with the image, that is, there is no
correlation between pixel value of image and value of noise
components.

Based on noise properties and types of noise, different filters are

used to reduce/remove noise.

E12) Gaussian Noise

Gaussian noise model is most frequently used in practice. The

PDF of a Gaussian random variable ‘z’ is given by

( z u ) 2
1 
p( z)  e 22
,
2

where z  intensity/grey level value

  mean (average) value of z
  standard deviation

Rayleigh Noise

Radar range and velocity images typically contain noise that can
be modeled by the Rayleigh distribution. Rayleigh distribution is
defined by

2
2
( z a )

 ( z  a )e b za
p(z )   b
0 za

b ( 4  )
Mean density is given   a   b as  2 
4 4
Erlang (Gamma) Noise

Erlang noise is given by

 a b z b 1  z
 e z0
p( z )   ( b  1)!
0 z0

a and b are positive integers. Mean density is given by

b b
and variance  2  2
is  
a a

200
Unit 6 Image Enhancement & Filtering-Frequency Domain
Uniform Noise

Uniform noise is specified as

 1
 if a  z  b
p(z )   b  a
0 otherwise

Then mean and variance of uniform noise is given by

a  b 2 (b  a ) 2
 , 
2 12

Impulse (Salt and Pepper) noise

Impulse (salt and pepper) noise is specified as

Pa za
p( z)  
Pb zb

E13) There are three methods of estimation of degradation function:

a) Observation
b) Experimentation
c) Modelling
E14) In presence of noise, degradation model as shown in figure 4
becomes

G(u, v)  F(u, v) H(u, v)  N(u, v)

After applying inverse filtering

F̂( u, v)  H R (u , v) G ( u , v)

Substituting values of H R ( u , v) and G(u, v)

1
F̂(u, v)  F(u, v) H(u, v)  N(u, v)]
H ( u , v)
F(u, v) H (u, v) N (n, v)
 
H ( u , v) H ( u , v)
N ( u , v)
 F(u, v) 
H ( u , v)

Limitations of inverse filtering are:

1) It is an unstable filter
2) It is sensitive to noise. In practice, inverse filter is not
popularly used.
E15) This approach includes both the degradation function and power
spectrum of noise characteristics in developing the restoration
201
Unit 6 Image Enhancement & Filtering-Frequency Domain
filter. Wiener filter restores the image in the presence of blur as
well as noise.

This method is founded by considering image and noise as

random variables and objective is to find as estimate f̂ of the
uncorrupted image f such that the mean square error between
them is minimized. This error is given by

e 2  E{(f  f̂ ) 2 })

Where E{.} is the expected value of the argument. Noise and

image are assumed to be uncorrelated.

S fg (u , v)
 H R (u , v) 
S gg ( u , v )

F̂( u, v)  H R (u , v) G ( u , v)
 H * (u , v) 
 *
 G (u , v)
 H (u , v)  k 
 1 H ( u , v) 
2

 * 2
 G (u , v)
 H (u , v) H ( u , v)  k 

k is chosen experimentally and iteratively for best results.

202
Unit 6 Image Enhancement & Filtering-Frequency Domain

203
Unit 4 Colour Image Processing

UNIT 7
COLOUR IMAGE PROCESSING

Structure Page No.

7.1 Introduction 73
Objectives
7.2 Human Vision System 74
7.3 Colour Fundamentals 76
7.4 Colour Models 79
RGB Model
CMY and CMYK Model
HSI Model
7.5 Pseudo-colour models 87
7.6 Summary 88
7.7 Solutions / Answers 88

7.1 INTRODUCTION
The purpose of this unit is to introduce the concepts related to colour
image processing. As we have been working with grayscale images till
now, we would like to have an in-depth discussion of how colour images
are formed and the various colour models that exist.

We shall first discuss the human vision system in Sec. 7.2. A healthy
vision system is capable of seeing the world in colour.

As you read further, we shall discuss the various colour models that
exist and the advantages and disadvantages of each in Sec. 7.3.

Finally, we shall discuss pseudo-colour processing, also called false

colour, which is the process of assigning colours to grey values based
on specified conditions. We shall discuss colour processing in Sec. 7.5.
We finally summarise the discussion in Sec. 7.6 and in Sec. 7.7, we
give the solutions/answers/hints to the exercises.

Now we shall list the objectives of this unit. After going through the unit,
please read this list again and make sure that you have achieved the
objectives.
73
Block 1
Objectives
After studying this unit, you should be able to:

 to differentiate between thousands of thousands of colours and their

shades in the colour images.
 To define the different colour models and use them as per the
requirements.
 To apply different pseudo colour models

A colour image is a powerful source of information. Human visual

system has the ability to differentiate between hundreds of colours and
their shades. Therefore, colour images contain a large amount of extra
information compared to grey-scale images, that give a better
understanding of the contents of the image, for example, in object
detection and segmentation. If an image is captured by a full-colour
sensor, then the resulting image is a full colour image.

A grayscale image can be converted into a colour image using the

technique of pseudo-colour processing, where each intensity is
assigned a colour.

Full colour image processing is primarily used in most applications such

as visualisation and publishing. We start with discussion on human
vision system in the following section.

7.2 HUMAN VISION SYSTEM

The human eye is nearly spherical in shape with a 20mm diameter on
an average, as shown in Fig. 4.1. The three main parts of the eye are:

i) The cornea and sclera outer cover,

ii) the choroid and
iii) the retina.

Let us discuss them one by one briefly.

Fig.1: Structure of the human eye source

74
Unit 4 Colour Image Processing
i) As you see in Fig. 1, the sclera is an opaque member that
encloses theoptic globe all around, except at the anterior end,
which is covered by the cornea. The cornea is a tough,
transparent cover of the anterior chamber.
ii) Choroid is the layer under the sclera. The membrane choroid
contains a network of blood vessels. These blood vessels form
the major source of nutrition for the eyes. If the choroid is
damaged and inflamed, it can restrict blood flow in the eye,
resulting in serious damage of the eye cells. The role of the
choroid is also to control the amount of light entering the eye as
well as reduce the backscatter inside the eye.

The choroid is divided into two parts:

a. The ciliary muscles which relax and tighten to enable the

lens to focus by changing its shape,
b. The iris diaphragm, that contracts and expands to control
the amount of light that enters the eye.

The lens is a transparent, biconvex structure that helps to refract

light into the eye such that the image is formed on the retina. The
lens is flexible and ca change shape to change the focal length of
the eye. THSI ensures that objects at various distances can be
focussed upon and their images can be formed on the retina.

iii) The retina is the innermost membrane of the eye. It lines the wall
of the complete posterior portion of the eye. The retina can be
thought of as the image plane in the eye, since on properly
focussing the eye on an object, light from that object is passed
through the lens such that the image is formed on the retina.

The retina consists of two types of cells called rods and cones.
The cones are highly sensitive to colour and are around 6-7
million in a human eye. The cones are located on the focea
which is the central portion of the retina.

However, there are 75-150 million rod cells which are completely
distributed all over the retina. The rod cells give the overall
picture of the object in the scene, and reduce the amount of
detail. Rods are also responsible for low light vision, also known
as SCOTOPIC vision, while cones are responsible for bright
light vision, also known as PHOTOPIC vision.

In the human eye, the distance between the retina and lens, that
is the focal length, varies between 14 and 17 mm as the
refractive power of the lens increases from min to max. For a
nearby object, the lens is most strongly refractive. Moreover, the
lens of the eye is very flexible and is flattened by controlling
muscles to enable the eye to focus on distant objects.

To allow the eye to focus on objects close to the eye, the

controlling muscles allow the lens to become thicker.
75
Block 1
Here, you might be wondering how human eye adapts to different levels
of brightness and how it discriminates various levels of brightness. The
answer to your question is given below.

Brightness Adaptation and Discrimination: Human vision system is

highly complex and can adapt to an enormous range of light intensity
levels-of the order of 10  10. The range starts from the scotopic
threshold and goes upto the glare limit. The subjective brightness, the
perceived intensity by the human eye, has been experimentally found to
be a logarithmic function of the light intensity that falls on the eye.
Since the human eye cannot interpret this dynamic range
simultaneously, brightness adaptation is carried out by the eye. The eye
can discriminate only a small range of distinct intensity levels
simultaneously. Brightness adaption level is the current sensitivity level
of a human eye for a given set of conditions.

Now, try the following exercises.

E1) If an observer is looking at a tree that is 100m far and if h is the

height of the tree in mm in the retinal image, what is h?

So, by now you know the fundamental concepts about human vision
system. In the following section, we are going to highlight various colour
models. You must have heard about some of them in your day-to-day
life.

7.3 COLOUR FUNDAMENTALS

Every colour is defined using three quantities that are independent of
each other, however, taken together they define a particular shade of a
colour. These quantities are:

i) Hue: Hue component of a colour is defined by the dominant

wavelength. The wavelength range on the electromagnetic
spectrum that defines the visible colour spectrum lies between
400nm [nm is nanometre] that represents the violet colour and
700nm that represents the red colour as can be seen in Fig. 2.

Fig. 2: Part of the electromagnetic spectrum that shows the visible spectrum
Source:
76
Unit 4 Colour Image Processing
ii) Saturation: The excitation purity of the colour is determined by the
quantity known as saturation.It is dependent on the amount of
white light that is mixed with hue of that colour. A fully saturated
colour implies that no white light is mixed with that hue.
iii) Chromaticity: The sum of hue and saturation constitutes the
chromaticity of the colour. Therefore, if there is no colour, it is called
achromatic light.
iv) Intensity: The amount of light actually present defines the intensity.
Therefore, intensity is a physical quantity. If more light is present,
the colour is more intense. Achromatic light has only intensity but no
colour. Grayscale images have only intensity
v) Luminance or Brightness: The perception of colour is the quantity
known as luminance or brightness. For example, given two colours
of the same intensity, such as blue and green, it is perceived that
blue is much darker than green.
vi) Reflectance: The ability of an object to reflect light, is the
reflectance property of the object. The reflectance property
determines the colour of the object, since, we see those colours that
are reflected back and not the ones that are absorbed. For example,
an object that reflects green, absorbs all other colours in the white
light spectrum except green.
There are about 6 to 7 million cones in the human eye and they are
responsible for recognising colours. Nearly, 65% of the cones recognise
red, 33% are sensitive to green and about 2% to blue. Red, green and
blue are known as the primary colours and nearly all other colours are
seen as a combination of these primary colours. However, there is a
difference between the primary colours of light and the primary colours
of pigments. The primary colours of light are red, blue and green and
they can be added to produce the secondary colours of light that are
yellow(red plus green), magenta (red+blue) and cyan (green + blue).
Moreover, the primary colours of pigments are said to be those that
absorbs a primary colour of light and reflects the other two. Therefore,
in tHSI case, the primary colours are cyan, magenta and yellow while
the secondary colours are red, blue and green.

For standardisation, in 1931, the Commission Internationale de

l'Éclairage (CIE) defined specific wavelengths for the three primary
colours: red = 700nm, blue: 435.8 nm and green = 546.1 nm. The
amounts of red, blue and green required to form a colour are known as
the Tristimulus values and are denoted as X, Y and Z.

Given X, Y and Z the tristimulus coefficients which define a colour. If we

define x, y and z as the relative values of the primary colours, then
these values can be found by

X Y Z
x ,Y and z  ... (1)
XYZ XYZ XYZ

It is obvious that

x  y  z  1. ... (2) 77
Block 1
Thus a 2-D diagram is adequate to show the coordinates x and y.

If we specify colours as a composition as x (red) and y (green). Then,

given the values of x and y, the value of z (blue) can be computed as:

z  1  ( x  y) ... (3)

Here, we can see only two variables are independent. Therefore, we

can show these variables in 2-D coordinate system.

The point on the boundary of the chromaticity chart is fully saturated,

while as a point moves farther from the boundary, more white light is
added and is therefore, less saturated. The saturation is zero at the
point of equal energy. A straightine joining any two points in the
chromaticity diagram, determines all possible colours that can be
obtained by combining the two colours at the endpoints of the segment.
This can be extended to combining three colours. The three line
segments joining the points pairwise form a triangle and various
combinations of the colours at the vertices of this triangle give all
colours inside the triangle or on the boundary of the triangle.

To understand this more clearly, we shall discuss few examples.

Example 1: Consider the coordinates of warm white (0.45, 0.4) and the
coordinates of deep blue (0.15, 0.2). Find the percentage of the
three colours red (X), green (Y) and blue ( Z).

Solution: We first find the trichromatic coefficients x, y and z. At the

point warm white, x  0.45 and y  0.4, therefore

z  1  ( x  y)
 1  (0.45  0.4)
 0.15

Now, we shall find the tristimulus values X, Y and Z.

X
 0.45
XYZ
Y
 0.4
XYZ
Z
 0.15
XYZ

Here, X : Y : Z  0.45 : 0.4 : 0.15

Therefore the percentage of each colour would be as follows:

Percentage of red (X)  45%

Percentage of green (Y)  40%
Percentage of Blue (Z)  15%
78
Unit 4 Colour Image Processing
At the point deep blue, x  0.15, y  0.2, therefore z  0.65.

We can find the percentage of each colour as we found in case of warm

white. We get percentage of red colour as 15%, percentage of green
colour as 20% and the percentage of blue colour as 65%.

We can see the percentage is justified for each colour name.

***
Example 2: Find the relative percentage of colours warm white and
deep blue which mixes the give the colour which lies on the line joining
them. Use the coordinates of these points as given in Example 1.

Solution: Let the colour C lies on the line have the coordinate ( x, y).

The distance of C from the warm white colour  ( x  0.45) 2  ( y  0.4) 2

Similarly, the distance of C from the deep blue colour

 (0.15  x ) 2  (0.2  y) 2

The percentage of warm white in

( x  0.45) 2  ( y  0.4) 2  (0.15  x ) 2  (0.2  y) 2

C  100
(0.45  0.15) 2  (0.4  0.2) 2
This expression can be used to find the percentage of warm white
colour at C by substituting the coordinates of the point C as per the
situation. Also, the percentage of the deep blue colour would be (100 -
percentage of warm white colour).
***
Now try the following exercises.

E2) Derive an expression to find the percentage of each colours

C1 , C 2 , and C 3 at the point C which lies within the triangle having
vertices as C1 , C 2 , and C3 .

In the following section, we shall discuss the most commonly used

colour models such as the RGB (red, green, blue), CMY (Cyan,
magenta, yellow) and HSI (hue, saturation, intensity).

7.4 COLOUR MODELS

Colour models or Colour spaces or Colour systems have been
introduced so as to be able to specify each colour in a generally
accepted manner. There are various colour models or colour spaces.
Each colour space specifies a particular colour in a standard manner,
by specifying a 3-D coordinate system and a subspace that contains all
possible colours in that colour model. Then, each colour in that colour
space is represented as a point in that subspace, given by three
coordinates ( x, y, z). These colour models are either oriented towards
79
Block 1
specific hardware or image processing applications. In this section, we
shall discuss three important colour models and the conversion of one
colour model into other.

Before we discuss each colour model, let us discuss the principles of

absorption of colours of any model by human eye.

i) The human eye has absorption characteristics of colours and

recognises them as variables. Thus, the colours red (R), green (G)
and blue (B) are called primary colours of light.
ii) Secondary colours of light are produced by adding primary colours.
For example red and blue produces magenta, red and green
produces yellow, green and blue produces cyan, etc.
iii) Proportion of primary and secondary colours in appropriate amount
produces white light.

Now, let us discuss each model separately.

7.4.1 The RGB Model

The RGB colour is based on a cartesian coordinate system, where the

colour subspace is a cube with axes representing red, green and blue.
A colour in the RGB model is therefore, specified as a 3-tuple (R , G, B)
where, R , G and B represent the amount of red, green and blue,
respectively, present in that colour. The geometry of the RGB colour
model is a cube as shown in Fig. 3. It is assumed that all colour values
are normalised, that is, it is a unit cube and all values of R, G and B lie
between 0 and 1. It is clear from Fig. 3, that it is a model of a cube with
the three axes for Red, Green and blue. The grayscale values lies on
the diagonal of the cube, which joins the black (0, 0, 0) and white (1,1,1)
vertices of the cube.

Fig. 3: The RGB colour (Image taken from [1])

A colour image in the RGB model consists of three images

corresponding to each of the three colours: Red (R), Green (G) and
Blue (B) colours. These three images combine to form one composite
colour image on a monitor. To covert an RGB image to a grayscale
80
Unit 4 Colour Image Processing
image, the intensity of the gray-pixel is given by the average of R, G
and B values. The RGB colour model is mainly used for colour monitors
and screens.

Now the question arises how do we find the composite colour in RGB
colour model at any point. For this we follow the following steps:

Step 1: Pixel depth is the number of bits used to represent each pixel. If
an image in RGB model has 8-bit image in each of its three colours,
then each RGB pixel has a depth of 3 image planes  8-bit per plane
that is 24 bits. This gives rise to 2 24 colour shades.

Step 2: We fix one of the three colours and let the other two colours to
vary. Suppose we fix R  127 and let G and B to vary. Then the colour at
any point on the plane parallel to GB plane would be (127, G, B), where,
G, B  0, 1, ...., 255.

Example 3: In a RGB image, the R and B components are at mid and

the G component is at 1, then which colour would be seen by a person?

Solution: At the given point, we have

R B 1 G
  G  (R  G  B) 
2 2 2 2
1
 mid grey  G.
2
 Pure green with some grey component.
***
Now, try an exercise.

E3) How many different shades of grey are there in a colour RGB
system if each RGB image is an 8 bit image?

After discussing RGB model, we discuss CMY and CMYK colour

models.

7.4.2 The CMY and CMYK Colour Model

Cyan (C), Magenta (M) and Yellow (Y) are the primary colours of
pigments and the secondary colours of light. The CMY model is a
subtractive model, implying that it subtracts a colour from the white light
and reflects the rest. For example, when white light is reflected on cyan,
it subtracts Red and reflects the rest. While RGB is an additive model,
where something is added to black (0,0,0) to get the desired colour
CMY is a subtractive model. The conversion between CMY and RGB
model is given by the Equation below.

 C  1  R 
M   1  G  ... (4)
    
 Y  1  B 
81
Block 1
where, the RGB values have been normalised. THSI also gives a
method to convert from RGB to CMY to enable printing hardcopy, since
the CMY model is used by printers and plotters.

(a) (b)

Fig. 4: (a) RGB and (b) CMY

Fig. 4(a) shows the RGB model, in which the white colour is produced
by adding the three primary colours Red, Green and Blue. Fig. 4 (b)
shows the CMY model, where black is obtained as the sum of Cyan,
Magenta and Yellow. The inverse relation between the RGB and CMY
models are also shown by these two images.

In practice, black is obtained by combining cyan, yellow and magenta,

however, this leads to a muddy looking black. For publishing, the black
colour plays an important role, therefore, in the CMYK colour model,
black is added as the fourth colour, where K stands for black.

Now check by doing the following exercise, what have you understood.

E4) Why do we get green coloured paint on mixing blue and yellow
coloured paints?

Now, let us discuss HSI model.

7.4.3 The HSI Model

This colour model is very close to human colour perception which uses
the hue, saturation and intensity components of a colour, when we see
a colour, we cannot describe it in terms of the amount of cyan, magenta
and yellow that the colour contains. Therefore, the HSI colour model
was introduced to enable describing a colour by its hue, saturation and
intensity/ brightness. Hue describes the pure colour, saturation
describes the degree of purity of that colour while intensity describes
the brightness or colour, sensation. In a grayscale image, intensity
defines the graylevel. Fig. 6 shows the HSI colour model and the way
colours may be specified by this colour model.

82
Unit 4 Colour Image Processing

(a) (b)
Fig. 5: (a) HSI and (b) RGB

In Fig.5, the HSI colour model is represented and its relation to RGB
model is shown in the Fig. 5 (b). The HSI triangle in Fig. 5 (b) shows a
slice from the HSI solid at a particular intensity as shown in Fig. 5 (a).

You may notice that in Fig. 5, the hue, saturation and intensity values
required to form the HSI colour space can be computed using the RGB
values.

Try the following exercises.

E6) Write the full form of HSI and define each of the components.

E7) What is colour space? Mention its classification.

Now we indicate how a RGB colour model is converted to a HSI colour

model.

To convert an image in RGB format to HSI colour space, the RGB value
of each pixel in the image, is converted to the corresponding HSI value
in the following manner. Hue, H is given by

  if B  G
H ... (5)
360   if B  G

where,
 1 
 [(R  G )  (R  B)] 
  cos 1  2
1/ 2 
[(R  G )  (R  B)(G  B)] 
2

 
Saturation, S is given by

3
S  1 [min (R , G, B)]
(R  G  B) 83
Block 1
And, intensity, I is given by

1
I  (R  G  B)
3

Where, the RGB values have been normalised in the range [0,1] and
the angle  is measured with respect to the red axis in the HSI space.

Now we would convert HSI colour model to RGB colour space.

Given pixel values in the HSI colour space in the interval [0,1] , the RGB
values can be computed in the same range. However, depending on the
H value, the RGB values are computed in different sectors, based on
the separation of the RGB colours by 120° intervals.

First, multiply the H value by 360°, to get the hue value in the interval
[0°, 360°].

In the RG sector, H takes the value in the interval [0°, 120°[,

that is, 0 o  H  120o and the RGB values are given by

B  1(1  S) ... (8)

 S cos H 
R  11   ... (9)
 cos(60  H ) 
o

G  31  (R  B) ... (10)

In the GB Sector, 120o  H  240o then in this case, we first convert the
value of H as

Then, the RGB values are computed as

R  I(1  S) ... (11)

 S cos H 
G  I 1   ... (12)
 cos(60  H ) 
o

B  31  (R  B) ... (13)

In BR sector, when, 240o  H  360o , we first convert H as H  H  240o

Then, the RGB values are

G  I(1  S) ... (14)

 S cos H 
B  I 1   ... (15)
 cos(60  H) 
o

84
Unit 4 Colour Image Processing
R  3I  (R  B) ... (16)

Example 4: Consider the image with different colours as given in Fig. 6

Write the RGB colours which would appear on monochrome display.
You may assume that all colours are at maximum intensity and
saturation. Also show each of the colour in black and white considering
them as 0 and 255 respectively.

Fig. 6

Solution: It is given here that the intensity and saturation are maximum,
therefore the value of each of the components RGB would be 0 or 1.
Let us check each colour of the image one by one by starting from the
left most colour.
Colour RGB Intensity/Saturation Monochrome
combination colours
R G B R G B
White R+G+B 1 1 1 255 255 255
Magenta R+B 1 0 1 255 0 255
Blue B 0 0 1 0 0 255
Cyan G+B 0 1 1 0 255 255
Green G 0 1 0 0 255 0
Yellow R+G 1 1 0 255 255 0
Red R 1 0 0 255 0 0
Black NIL 0 0 0 0 0 0
Now hence forth we shall follow the conversion that 0 represents black
and 255 represents white. Also, the grey is represented by 128. You
see that the table has R colour series as 255, 255, 0, 0, 255, 255, 0.
Thus, it would show W, W, B, B, B, W, W, B in monochrome display,
which is shown in Fig. 7 (a).
Similarly monochrome display of green colour would be shown by the
series W,B,B,W,W,W,B,B and blue would be shown as W, W, W, W, B,
B, B, B, B as shown in Fig. 7 (b) and Fig. 7 (c).

(a) (b) 85
Block 1

(c)

Fig. 7
***
Example 5: Let us sketch the HSI components of the image considered
in Fig. 6 [Given in Example 4] on a monochrome display.
Solution: We tr
transform
ansform HSI by computing values of H, S and I for each
colour.
For white R  1, G  1, B  1

Using Eqn. (5), we can say that H does not exist as denominator is zero
while computing .

1
Using Eqn. (7), we get I  (1  1  1)  1 and using Eqn. (6), we get
3
3
S  1 [min(1,1,1)]  0.
111
Similarly we can find H, S, I for each of the colours as shown in the
following table.

Colour R G B H S I Monochromatic
H S I
White 1 1 1 Cannot 0 1 ̶ 0 255
be
computed
Magenta 1 0 1 5 1 2 213 255 170
6 3
Blue 0 0 1 2 1 1 170 255 85
3 3
Cyan 0 1 1 1 1 2 128 255 170
2 3
Green 0 1 0 1 1 1 85 255 85
3 3
Yellow 1 1 0 1 1 2 43 255 170
6 3
Red 1 0 0 0 1 1 0 255 85
3
Black 0 0 0 ̶ 0 0 ̶ ̶ 0
86
Unit 4 Colour Image Processing

(a) (b)

(c)
Fig. 8

The output is given in Fig. 8 (a), Fig. 8 (b) and Fig. 8 (c) for each of the
attribute H,S and I.
***
Now try the following exercises.

E8) Describe how the grey levels vary in RGB primary images that
make up the font face of the colour cube.
E9) Transform the RGB cube by its CMY cube. Lable all the vertices.
Also, interpret the colours at the edges with respect to saturation.

In the following section, we discuss pseudocolour image processing.

7.5 PSEUDOCOLOUR IMAGE PROCESSING

Pseudocolour image processing is the process of assigning colour to
each pixel of a grayscale image based on specific conditions. As
mentioned above, colour carries with it a large amount of information
regarding the objects that we are viewing and therefore, for better
visualisation, converting a grayscale image to a colour image helps in
improved interpretation of the image.
Intensity slicing or density slicing is one of the simplest forms of
pseudocolour image processing technique. In this technique, the image
is interpreted as a 3D function and can be imagined as a set of 2D grid
which are parallel to the coordinate planes and placed at each intensity
value. Each plane can then be thought of as a slice of the image
function in the area of intersection. For example, the plane at f ( x , y)  I i
slices the image function into two parts. Then, any pixel whose
graylevel is on or above the plane can be coded in one colour and

87
Block 1
whose graylevel is below the plane can be coded in another colour,
thereby converting the grayscale image into a two colour image.

Fig. 9: Intensity slicing technique represented as slicing planes perpendicular to

the intensity axis at various intensity levels.
In general, intensity slicing technique is given as follows:
Let [0, L  1] be the existing grayscale values such that black is
represented by l0 , that is, f ( x, y)  0 and white is represented by l L1 ,
that is, f ( x, y)  L  1. Then, we define P planes such that they are
perpendicular to the intensity axis at levels l1 , l2 ,..., l P , where 0  P  L  1.
These P planes are parallel to the coordinate plane. The grayscale
intensity levels are partitioned into P  1 intervals, namely, V1 , V2 ,...VP1 .
Then, assign a colour to each pixel in the following manner
f ( x , y)  c k if f ( x , y)  Vk
where, Vk is the k th intensity interval defined by planes at levels lk  1
and lk and c k is the colour associated with Vk .
You may try an exercise:

E10) Define an application of intensity level slicing.

Now let us, summarise what we have discussed in this unit.

7.6 SUMMARY
In this unit, we discussed the following points:
1. The need for colour image processing. Since the human eye has the
wonderful capability of seeing millions of colour, we realise that
colour gives a large amount of information about the objects and
scene in the images.
2. We first discussed the structure of the human eye and then the
tristimulus theory that connects the perception of colour with the
various colour models that exist.
3. We then discussed the main colour models or colour spaces that are
mainly used in both TV and print.

7.7 SOLUTIONS/ANSWERS
E1) Since when object is far, the focal length is 17 mm for the human
eye, therefore, 15 / 100  h / 17  h (17 *15) / 100  2.55 m
88
Unit 4 Colour Image Processing
E2) THSI problem is the extention of the problem solved in Example 2.
Here, we consider two possibilities.
i) When the point C at which percentage of colours C1 , C 2 and
C 3 to be found is on the sides of triangle. In this case the
percentage is found by considering the point on the line
joining the corresponding vertices as we solved in Example
2. There would be 0% from the vertex which does not lie on
the line. For example, if the point lies on the line joining C1
and C 2 , then the percentage of C1 and C 2 can be found as
given below.
Let the coordinates of C1 be ( x 1 , y1 ), C 2 be ( x 2 , y 2 ), C 3 be
( x 3 , y 3 ) and C be ( x, y).
Percentage of C1 in
( x  x 1 ) 2  ( y  y1 ) 2  ( x 2  x ) 2  ( y 2  y) 2
C 100
( x 2  x 1 ) 2  ( y 2  y1 ) 2
Percentage of C 2 in C  (100  percentage of C1 ) %
Percentage of C 3 in C  0 % .
ii) When the point C does not lie on the sides of the triangle
with vertices C1 , C 2 and C3 .
In HSI case we join the point C with any of the vertices say
C 3 . We follow the following steps.
iii) Join the points C and C 3 and extend the line towards the
side C1C 2 . Suppose it intersects C1C 2 at C 4 .
iv) Find the percentage of C1 and C 2 at C 4 .
v) Use the concept that the ratio of C1 and C 2 will remain same
at each of the points on the line C 3C 4 .
vi) Now, we can easily find the coordinates of the point C 4 by
writing equation of the lines C1C 2 and C 3C. C 4 is the point of
intersection of C1C 2 and C 3 C.
vii) Finally, we can find the percentage of C 4 and C 3 for the
colour C.
E3) For an 8-bit image, there are 28  256 possible values. A colour
will be grey if each of the colour in RGB is same. Therefore, there
can be 256 shades of grey.
E4) You can see in Fig. 5, yellow paint is made by combining green
and red while imperfections in blue leads to reflection of some
amount of green from blue paint also. Therefore, when both blue
and yellow are mixed, both reflect the green colour, while all
other colours are absorbed. Therefore, green coloured paint
results from mixing of blue and yellow paints.
E5) H stands for Hue, which represents dominant colour as observed
by an observer and the corresponding wavelength is also
dominant. S stands for Saturation, which is the amount of white
89
Block 1
light mixed with a hue. I stands for intensity which reflects the
brightness.
E7) A colour space allows one to represent all the colour perceived
by human eye. The colour space can be broadly classified into (i)
RGB, (ii) GMY and (iii) HSI colour space.
E8) Each of the components in RGB model would vary from 0 to 255.
Here, we are discussing the front face. So, we fix all pixel values
in the Red image as 255 and let the columns to vary from 0 to
255 in the green image and rows to vary from 255 to 0 in the blue
image.
E9) The vertices would be as given below:
White  (0, 0, 0)
Cyan  (1, 0, 0)
Magenta  (0,1, 0)
Blue  (1,1, 0)
Green  (1, 0,1)
Red  (0,1,1)
Black  (1,1,1) .
The edges which are free from black or white pixels are fully
saturated. The saturation decreases towards the ends having
black or white pixel.

E10) Detecting blockages in medical image processing is an

application of intensity level slicing.

References
[1] R.C. Gonzoles and R.E. Woods, Digital Image Processing, Addision-
wesley, 1992.

[2] A.K. Jain, Fundamentals of Digital Image Processing, PHI.

90