1

INTRODUCTION TO IMAGE-
PROCESSING SYSTEM
Unit structure
1.0 Objectives
1.1 Introduction
1.2 Image Sampling
1.2.1 Theory of 2D Sampling
1.2.2 Retrieving Image From Its Sample
1.2.3 Violation of Sampling criterion
1.3 Quantization

.in
1.4 Resolution
1.5 Human Visual Systems
1.5.1 Anatomy of HVS
es
1.5.2 Scotopic AndPhotopic Vision
1.5.3 Brightness and contrast
ot

1.5.4 Laws Of Physics Associated With Vision

1.5.5 MACH BAND EFFECT
un

1.6 Classification of Digital Images

1.7 Image Types
m

1.8 Elements of an Image-processing System

1.9 Applications of Digital Image Processing
1.10 Summary
1.11 References

1.0 OBJECTIVES
As the name indicates digital image processing helps:
 To study two-dimensional Signals and Systems.
 To understand image fundamentals and transforms necessary for image
processing.
 To study the image enhancement techniques in spatial and frequency
domain.
 To study image segmentation and image compression techniques

1
Digital Image Processing 1.1 INTRODUCTION
In current era digital plays an important role in several areas like
geography, information technology, medicine etc.
Digital image processing(DIP)refers to management of digital images
through a digital computer. It is a subarea of signals and systems but
particularly focusses on images. DIP aims on using a computer system that
is able to perform several processing of an image. The process contains
the input of that system which is a digital image , the system then process
that image using effective algorithms, and gives as an output an image.
The most general example is Matlab, Adobe Photoshop. It is widely used
for processing digital images.
Before proceeding let us define some terms as follows :
1. IMAGE : An image is defined as an two dimensional signal. It is
defined using mathematical function f(x,y) where x and y are the two
co-ordinates horizontally and vertically called as spatial or plane

.in
coordinates.The value of f(x,y) at any point is gives the pixel value at
that point of an image.represents a measure of some characteristic such
as brightness or color of a viewed scene. An image is a projection of a
3- D scene into a 2D projection plane
es
2. ANALOG IMAGE: An analog image cab ne represented
mathematically using continuous range of values which represent
ot

intensity and position. Some of the examples include television

images, photographs, paintings, and medical images etc.
3. DIGITAL IMAGE :It is composed of matrix of small pixels. For
un

operating with the images, there are various software and algorithms.
Some of the examples of digital images are color processing, image
recognition, video processing, etc. Analog signals can be convted into
m

digital using two important operations i.e sampling and quantisation.

Analog Sampling Digital

Image Quantization Image

Figure 1
Advantages of Digital Image Processing

 Fast processing and effectiveness of cost

 Easier Image rebuilding and reformatting
 Speedy image storing and retrieval
 Fast and high-quality image distribution.

Disadvantages of Digital Image Processing

 It is very much time-consuming, needs high speed processor.

 Enlargement of image not possible after certain size.
 It is very much expensive depending on the particular system.
2
4. Digital Image representation : Introduction to Image-
A digital image is an : Processing System
 N x N array of elements
 2-dimensional discrete signal
To convert an image in digital format either scanner or digital camera is
used. These can be created directly on computer screens.

5 Neighbours of Pixel
A pixel will have four neighbpurs if they exist. NORTH, SOUTH, EAST
and WEST.

North

West P East

South

.in
Figure 2

1.2 IMAGE SAMPLING

es
Sampling is the procedure of evaluating at discrete spatial points the
information of brightness . f(x,y) – a continuous image function can be
sampled in the plane with the help of discrete grid of sampling points.
ot

1.2.1 THEORY of 2D SAMPLING

Let the analog image be represented by f(x,y). the corresponding discrete
un

edition of f(x,y) is obtained by defining f(x,y) at particular instances called

as sampling intervals x,y which are real positive constants.
m

Steps in 2D sampling are as follows :

1. Take analog sample f(x,y)
2. Take fourier transform after which we get spectrum of f(x,y) denoted
by F(1,2) as

3
Digital Image Processing 3. Then take inverse Fourire transform after which we get

4. For discrete signal we specify where

and are expressed in radians

5. Differentiating , and substituting in Inverse transform in step 3

we get discretised form of f(x,y) as given below

6. Replacing the integration by summation , changing limits we have

.in
es
An alertnate approach for the above procedure is multiplying the anlaog
image by a 2D comb function.It is a rectangular grid of points as shown in
ot

figure 3 . the spaces in between thr grid points are x,y respectively.
un
m

Figure 3 Figure 4

Figure 4 represents a 3 Dimensional view of comb function . The 2D

view also called as bed-of-nail function defined by :

4
After multiplying the comb function and analog function f(x,y) we get Introduction to Image-
discrete version of analog image given by Processing System

Next fourier transform of above product of functions is computed and

convolution of signals is applied to obtain

Summation being an linear operator , interchanging the order we get :

We can see that the final equations from both the methods are similar.

.in
1.2.2 RETRIEVING IMAGE FROM ITS SAMPLE
Since discreteness in one domain leads to periodicity in other domain,
es
sampling in spatial domain also leads to a periodic spectrum in the
frequency domain as in figure 5
ot
un
m

Figure 5 :Periodic spectrum of sampled image.

In order to retrieve the original image from the sampled spectrum we
apply the condition : sampling frequency should be greater than two
times the maximum signal frequency which is called “The Sampling
Theorem”

5
Digital Image Processing That is (*)

Where 2 are called Nyquist rates .

A low pass filter is used to extract the desired spectrum whose transfer
function is given by :

The region of support is indicated as P in figure 5. The continuous image

can be obtained from sampled spectrum by multiplying the sampled

.in
spectrum with low pass filter as given below :
es
ot

1.2.3 VIOLATION OF SAMPLING CRITERIA

un

Violation of sampling criteria given in (*) leads to aliasing which

fundamentally happens due to under sampling. It also leads to overlapping
of spectrum in the direction as shown in figure 6.
m

Figure 6 : Along the direction of under sampling

6
 Introduction to Image-
Processing System

Figure 7 : Along the direction of under sampling

In the similar manner violation of sampling criteria given in (*) leads to to
overlapping of spectrum in the direction as shown in figure 7.

.in
es
Violation of sampling criteria given in (*) simultaneously leads to to
overlapping of spectrum in both the direction as shown in
figure 8.
ot
un
m

Figure 8 : Along the direction of both and under sampling

7
Digital Image Processing 1.3 QUANTIZATION
Quantization is a process of transforming a real valued sampled image to
one taking only a finite number of distinct values. Under quantization
procedure the amplitude values of the image are digitized. In naïve
words, when you are quantizing an image, you are actually partitioning a
signal into quanta(partitions).
Quantiser design includes :

 Input decision level

 Output representation level

 Number of levels.
Quantisers are classified in two types :
1. Scalar quantisers

.in
2. Vector quantisers es
ot
un

Figure 9 : Classification of quantisers

m

1.4 RESOLUTION
Resolution contributes the degree of distinguishable details. It is classified
as :
1. Spatial resolution :Spatial resolution is defined as the smallest
discernible detail in an image.Spatial resolution states that the clarity
of an image cannot be determined by the pixel resolution. The number
of pixels in an image does not matter.It is the number of independent
pixels values per inch.sampling is the principal factor determining this
resolution .
2. Gray – level resolution : Gray level resolution refers to the predictable
or deterministic change in the shades or levels of gray in an image. In
short gray level resolution is equal to the number of bits per pixel.

8
 Introduction to Image-
1.5 HUMAN VISUAL SYSTEM Processing System
Human Visual System (HVS) is the most complicated system in life.
Many image processing applications are anticipated to produce images
that are to be watched by human observers. In HVS , the eyes act as the
sensor or camera, neurons act as the connecting cable and the brain acts
as the processor. It is hence important to realize the characteristics and
limitations of the human visual system to understand the “receiver” of the
2D signals.

1.5.1 ANATOMY OF THE HVS

.in
es
Figure 10 : Vertical section of the Human Eye
The first part of the visual system is the eye. This is shown in figure. Its
ot

form is nearly spherical and its diameter is approximately 20 mm. Its outer
cover consists of the ‘cornea' and ‘sclera' The cornea is a tough transparent
tissue in the front part of the eye. The sclera is an opaque membrane,
un

which is continuous with cornea and covers the remainder of the eye.
Directly below the sclera lies the “choroids”, which has many blood
vessels. At its anterior extreme lies the iris diaphragm. The light enters in
m

the eye through the central opening of the iris, whose diameter varies from
2mm to 8mm, according to the illumination conditions. Behind the iris is
the “lens” which consists of concentric layers of fibrous cells and contains
up to 60 to 70% of water. Its operation is similar to that of the man made
optical lenses. It focuses the light on the “retina” which is the innermost
membrane of the eye. Retina has two kinds of photoreceptors: cones and
rods. The cones are highly sensitive to color. Their number is 6-7 million
and they are mainly located at the central part of the retina. Each cone is
connected to one nerve end. Cone vision is the photopic or bright light
vision. Rods serve to view the general picture of the vision field. They are
sensitive to low levels of illumination and cannot discriminate colors. This
is the scotopic or dim-light vision. Their number is 75 to 150 million and
they are distributed over the retinal surface.
Several rods are connected to a single nerve end. This fact and their large
spatial distribution explain their low resolution. Both cones and rods
transform light to electric stimulus, which is carried through the optical

9
Digital Image Processing nerve to the human brain for the high level image processing and
perception.

1.5.2 SCOTOPIC AND PHOTOPIC VISION

Scotopic vision is the vision of the eye under low light conditions. Cone
cells do not function as well as rod cells in low level lighting so scotopic
vision occurs completely through rod cells, which are most sensitive to
wavelengths of light on the electromagnetic spectrum of 498nm, which
would be the blue-green bands of colour.
Photopic vision is the vision of the eye under well-lit conditions, generally
usual daylight light intensity. It allows colour perception which is
facilitated by cone cells. Cone cells have a higher visual acuity as well as
providing the eye’s colour sensitivity.
Scotopic vision uses only rods to see, meaning that objects are visible, but
appear in black and white, whereas photopic vision uses cones and
provides color.

.in
1.5.3 BRIGHTNESS AND CONTRAST
The Contrast and Brightness function enhances the appearance of raster
data by modifying the brightness and contrast within the image.
es
Brightness increases the overall lightness of the image—for example,
making dark colors lighter and light colors whiter—while contrast adjusts
the difference between the darkest and lightest colors. Below is an
ot

example of adjustments made to the brightness and contrast of an image.

1.5.4 LAWS OF PHYSICS ASSOCIATED WITH VISION

un

The important laws associated with vision are :

1. Weber's law: It states that the just noticeable difference in stimulus
intensity may affect the production of sensations proportionally.
m

In simple terms, the size of the intensity of stimuli will show a

proportionate change in producing the sense experiences.

 I = kL
Where ΔI(Delta I) represents the difference threshold, I represents the
initial stimulus intensity and k signifies that the proportion on the left side
of the equation remains constant despite variations in the I term.
It is denoted by Delta I.
2. Steven’s Law : If L is physical stimulus and I the perceived sensation
then the mathematical form of steven’s law is
I =  Ln
Where L is the intensity against black background, the value of n is 0.5

10
3. Bloch’s Law : It is given by Introduction to Image-
Processing System

Where C is constant valid for a duration of shorter than 0.1 second

4. Ricco’s law : It is given by

5. Piper’s law : It is given by

.in
1.5.5 MACH BAND EFFECT
Another characteristic of HVS is that it tends to “overshoot” around image
edges (boundaries of regions having different intensity). As a result,
es
regions of constant intensity, which are close to edges, appear to have
varying intensity. Such an example is shown in Figure11 . The stripes
appear to have varying intensity along the horizontal dimension, whereas
their intensity is constant. This effect is called Mach band effect. It
ot

indicates that the human eye is sensitive to edge information and that it
has high-pass characteristics.
un
m

Figure 11 : Mach band effect

1.6 CLASSIFICATION OF DIGITAL IMAGES

Digital images can be classified in two types as follows :
1. Raster or Bitmap Image: Also known as bitmap images are made
of pixels or tiny dots that use color and tone to produce the
image. Pixels appear like little squares on graph paper when the
image is zoomed in or enlarged. These images are created by
digital cameras, by scanning images into a computer or with
raster-based software. Each image can only contain a fixed
number of pixels; the amount of pixels determines the quality of
the image. This is known as resolution. More pixels result in
better quality at the same or larger sizes as the original, but this
11
Digital Image Processing also increases the size of the file and the amount of space it
takes to store the file. The lower the number of pixels, the lower
the resolution. Resolution limits the size the image can be scaled
up without being able to see pixels.
Common raster formats are BMP (windows Bitmap), PCX (Paintbrush),
TIFF (Tag Inter Leave Format) etc.

.in
Figure 12 : Zooming Of Raster Image
es
2. Vector Image :These images consist of anchored dots and are
connected by lines and curves, similar to the connect-the-
dot activities you may have done as a kid. Because these
graphics are not based on pixels, they are known as resolution
ot

independent, which makes them infinitely scalable. Their lines

are sharp, without any loss in quality or detail, no matter what
their size. These graphics are also device-independent, which
un

means their quality doesn't depend on the number of dots

available on a printer or the number of pixels on a screen.
Because they consist of lines and anchor points, the size of the
m

file is relatively small.

Figure 13 : Zooming Of Vector Image

12
 Introduction to Image-
1.7 IMAGE TYPES Processing System
Images are classified into four categories:
1. Binary image : Also called as Black and White image ,the binary image
as it name sys , contain only two pixel values - 0 and 1.Here 0 refers to
black color and 1 refers to white color. It is also know as Monochrome
image . A gray scale image can be converted to black and white image by
using threshold operation.

Figure 14 : Binary image representation

.in
2. Gray scale Image :Grayscale images, a kind of black-and-white or
gray monochrome, are composed exclusively of shades of gray. They
es
conatin only brightness information.The contrast ranges from black at the
weakest intensity to white at the strongest.In a gray scale image a pixel is
represented by a word or 8-bits, where a value of 0 represents black and
255 white.
ot
un
m

Figure 15 Gray scale image representation

3. Colour Image :It has three values per pixel which measure
chrominance and intensity of light.Each pixel is avector of
colourcomponents.Commoncolour spaces are RGB(RegGren Blue),
CMYK(Cyan, Magenta, Yellow, Black) and HSV(Hue, Saturation
,Value).

4. Volume Image: It is three-dimensional image and can be obtained

from some medical imaging euquipmentwhere individual data points
are called voxels- volume pixels. Example: CAT scan

13
Digital Image Processing 5. Range Image :These are special class of digital images .Here each
pixel represents a distance between a reference frame and a visible
point in the screen.Also called as depth images

6. Multispectral Image :These are images of same object taken in

different bands of visible or infrared regions.they contain information
outside the normal perception range of human.

1.8 ELEMENTS OF IMAGE PROCESSING SYSTEM

Different elemenst of image processing system are :
1. Image – auisition elements : Image acquisition can be defined as the
act of procuring an image from sourcesi.e computer and storing them
in devices . This can be done via hardware systems such as cameras,
encoders, sensors, etc. Digital cameras are popular nowadays. Most of
the digital cameras use CCD or CMOS image sensor. Some of
thdefinitios commonly used in solid-state image sensors are:

.in
a. CCD (Charge-Coupled Device): A light-sensitive array of silicon
cells that is commonly used for digital camera image sensors. It
generates electrical current in proportion to light input and allows the
es
simultaneous capture of many pixels with one brief exposure.
b. Dark current:the charge of signal collected by pixel in absence of
light
ot

c. Photo site: it is the portion of silicon that functions as a light sensitive

area.
un

d. Pixel : It is the discrete sensitive cell that collects and holda photo
charge.
e. Fill factor :The fill factor of an imaging sensor is defined as the ratio
m

of a pixel's light sensitive area to its total theoretical area.

f. Quantum efficiency:It is the measure of the effectiveness of an
imaging device to convert incident photons into electrons
2. Image – storage devices : in order to store the image obtained from
camera for processing and future use different storing devices are
used.Basically there are two types of storage devices :
i. Disk system
ii. Flash memory
3. Image – display devices: the final stage of image – processing step is
image display.Usually computer moniors are used for the same.some
of the factors likw size of monitor, number of colours ,spatial
resolution needs to be considered.

14
1.9 APPLICATIONS OF DIGITAL IMAGE Introduction to Image-
Processing System
PROCESSING
Some of the most important fields in which digital image processing is
widely used are as given below :
 Image sharpening and restoration
 Medical field
 Remote sensing
 Transmission and encoding
 Machine/Robot vision
 Color processing
 Pattern recognition

.in
 Video processing
 Microscopic Imaging
es
1.10 SUMMARY
In this chapter we learnt about basics of digital image processing like its
structure components and conversion.
ot

1.11 EXERCISES
un

1. Explain 2D image sampling.

2. Explain anatomy of human vision system.
m

3. Explain sampling and quantization.

4. Explain different element of image processing.
5. Discuss applications of image processing.

1.11 REFERENCES
1) Digital Image Processing, S Jayaraman, S Esakkirajan, T
Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008
3) Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. EsakkirajanAnd T. Veerakumar, 2016


15
 2
CONVOLUTION AND CORRELATION
Unit structure
2.0 Objectives
2.1 Introduction
2.2 2D Signals
2.3 2D Systems
2.4 2D convolution
2.4.1 Graphical Method
2.4.2 Z Transform

.in
2.4.3 Matrix analysis
2.4.4 Circular convolution
es
2.5 2D Correlation
2.6 Summary
ot

2.7 References
un

2.0 OBJECTIVES
In this chapter we are going to learn about theory of signals and some
mathematical concepts required in signal processing .
m

2.1 INTRODUCTION
Signals convey information. Systems transform signals. A signal can be,
for example, a sequence of commands or a list of names. Mathematically,
we model both signals and systems as functions. A signal is a function that
maps a domain, often time or space, into a range, often a physical measure
such as air pressure or light intensity. A system is a function that maps
signals from its domain—its input signals—into signals in its range—its
output signals. Both the domain and the range are sets of signals (signal
spaces). Thus, systems are functions that operate on functions.
An example of one D signal is ECG, 2D signal is a still image.

2.2 2D SIGNALS
2D discrete signals are represented as : x(n1,n2) where x is a real or
complex value and n1,n2a pair of integers.

16
 2D unit impulse sequence: it is given by x(n1,n2) = (n1,n2) Convolution and
Correlation
Which is given by

 Line impulse : different types are :

o Vertical line impulse
o Horizontal line impulse
o Diagonal impulse

.in
Figure 1 : 2D impulse sequence Vertical line impulse Horizontal
Line Impulse
es
ot
un
m

Figure 2 : Diagonal impulse function Diagonal function

representation

 Exponential Sequence : they are defined by :

 Separable Sequence : A signal x(n1,n2) is said to be separable if can be

represented as product of the function n1 alone and a function n2 alone
given as x(n1,n2) =x1(n1)x2(n2)

17
Digital Image Processing  Periodic Sequence : A 2d sequence is periodic if it repeats itself in a
regulary spaced interval.hence a sequence is periodic if
x(n1,n2) = x(n1 + N1,n2) = x(n1,n2 + N2) where N1 and N2 are positive
integers

2.3 2D SYSTEMS
A 2D system is a device or algorithm that carries out some operation on a
2D signal.ifx(n1,n2) is input signal to the system and y(n1,n2) is output of
the system then the relation is : y(n1,n2) = T[x(n1,n2)] where T is the
transformation by the system to produce output from input.
Classification of 2D systems : It can be classified as :
1. Linear Vs Non-Linear systems : A linear system is the one that
follows the laws of superposition. This law is necessary and sufficient
condition to prove the linearity of the system.the Linearity is defines as
:

.in
es
Where a and b are scalar constants.
ot
un
m

Figure 3 : Superposition Principle

2. Shift Variant Vs shift -invariant systems : If the input-output of the
systems does not change with time then the system is said to be shift
invariant. Mathematically it is given by :

3. Static Vs Dynamic system : If the output of the systems at nay instant

depends on most of the input sample but not on past and future
samples of the input then the system is said to be static or memoryless.
In any other case th system is said to be dynamic or of having
memory.

18
4. Stable system : A 2D linear shift invariant system is stable if and Convolution and
only if its impulse response is absolutely summable given by : Correlation

2.4 2D CONVOLUTION
Convolution and correlation are used to extract information from
images.They are linear and shift-invariant operations.Convolution is the
relationship between a system's input signal, output signal, and impulse
response.Shift-invariant means that we perform the same operation at
every point in the image.Linear means that this operation is linear, that is,
we replace every pixel with a linear
combination of its neighbors. These two properties make these operations

.in
very simple;it’s simpler if we do the same thing everywhere, and linear
operations are always thesimplest ones.Correlation is a way to identify a
known waveform in a noisy background. Correlation is a mathematical
operation much very similar to convolution. Just as with convolution,
es
correlation uses two signals to produce a third signal. This third signal is
called the cross-correlation of the two input signals. If a signal is
correlated with itself, the resulting signal is instead called the
ot

autocorrelation.
Convolution has wide range of applications like image fitering ,
enhancement, restoration ,feature extraction and template matching.
un

The 2-dimensional discrete convolution between two signals x[n1,n2] and

h[n1,n2] is given by :
m

Convultions can be performed either in spatial domain or frequency

domain.

2.4.1 2D CONVOLUTION THROUGH GRAPHICAL METHOD

In 2D convolution through graphical method following operations are
performed :

 Folding

 Shifting and

 Addition

19
Digital Image Processing The process is performed as follows :
1. Given two matrices as input :x[n,m] and h[,m] ,first determine the
dimension of the resultant matrix i.e if x[n,m] is matrix of order 2x3
and h[,m] is of order 3x1 then resultant matrix is of order 4x3

2. The resultant matrix is calculated as :

.in
es
ot
un
m

20
3. Determine the values y(0,0), y(0,1) etc.wherey(0,0) is obtained as : Convolution and
Correlation

Similary values for other elements.

4. The resulatant matrix is :

.in
es
ot
un

2.4.2 2D CONVOLUTION THROUGH Z TRANSFORM

The Z-transform is a mathematical tool which is used to convert the
m

difference equations in discrete time domain into the algebraic equations

in z-domain.
Mathematically, if x(n) is a discrete time function, then its Z-transform is
defined as,

The convolution between two signals x(n1,n2) and h(n1,n2) is given by

y(n1,n2) = x(n1,n2) ** h(n1,n2) where ** indicated convolution
taking Z transform on both the sides we get
Y(Z1,Z2) = X(Z1,Z2) x H(Z1,Z2)
Convolution is one domain is equal to multiplication in other domain.
21
Digital Image Processing 2.4.3 2D CONVOLUTION THROUGH MATRIX ANALYSIS
2D convolution can be performed using matrix multiplication with the
help of block Teoplitz matrix of one matrix and column matrix of the
other .
The flowchart for the same is as given below :

.in
es
ot
un
m

The number of block matrices depend on number of rows of x[m,n]

2.4.4 CIRCULAR CONVOLUTION

Circular convolution can be carried out for signals which are periodic in
nature.
The convolution between two signals x(n1,n2) and h(n1,n2) is given by

y(n1,n2) = x(n1,n2)  h(n1,n2)

The flowchart for the same is as given below :

22
 Convolution and
Correlation

.in
es
Circular correlation is widely used in zooming operation of digital
cameras.
ot

2.5 2D CORRELATION
un

Correlation is a technique of signal matching .It is an important

component of radar,sonar, digital communication and other
systems.Mathematically similar to convolution it is of following types :
m

Correlation can be performed using :

 Matrix methods : with the help of block Teoplitz matrix and block
Circulant matrix in terms of convolution

23
Digital Image Processing The flowchart for the same is as given below :

.in
es
ot
un
m

24
Circular correlation can be performed through matrix method as follows : Convolution and
Correlation

.in
es
ot
un
m

25
Digital Image Processing Circular correlation can be performed through transform method as
follows :

.in
es
ot
un
m

2.6 SUMMARY
In this chapter we learnt about convolution that can be carried out for
filtering process, its types and methods to perform.we also learnt about
correlation which is used to find the similarity between images or part of
images.

2.7 EXERCISES
1. Write a note on 2d signals.
2. Explain 2D systems and its classifications.
3. Define convolution . Explain its types of representation.
4. Write a note on cicular convolution.
5. Discuss applications ofCircular Convolution
6. Write a note on Correlation.

26
2.7 REFERENCES Convolution and
Correlation
1) Digital Image Processing, S Jayaraman, S Esakkirajan, T
Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008
3) Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. EsakkirajanAnd T. Veerakumar, 2016




.in
es
ot
un
m

27
 3
IMAGE TRANSFORMS
Unit structure
3.0 Objectives
3.1 Introduction
3.2 Need for transform
3.3 Image transforms
3.4 Fourier transform
3.5 2D Discrete Fourier transform
3.6 Properties of 2D DFT

.in
3.7 Other transforms
3.6 Summary
es
3.7 References

3.0 OBJECTIVES
ot

In this chapter we are going to learn about image transforms which are
widely used in image analysis and image processing .Transform is a
un

mathematical tool basically used to move from one domain to another i.e
from time domain to frequency domain . they are useful for fast
computation of correlation and convolution.
m

3.1 INTRODUCTION
An image transform can be applied to an image to convert it from one
domain to another. Viewing an image in domains such as frequency
or Hough space enables the identification of features that may not be as
easily detected in the spatial domain. Common image transforms include:
 Hough Transform, used to find lines in an image
 Radon Transform, used to reconstruct images from fan-beam and
parallel-beam projection data
 Discrete Cosine Transform, used in image and video compression
 Discrete Fourier Transform, used in filtering and frequency analysis
 Wavelet Transform, used to perform discrete wavelet analysis, denoise,
and fuse images

28
 Image Transforms
3.2 NEED FOR TRANSFORM
The main purpose of the transform can be divided into following 5
groups:
1. Visualization: The objects which are not visible, they are observed.
2. Image sharpening and restoration: It is used for better image
resolution.
3. Image retrieval: An image of interest can be seen
4. Measurement of pattern: In an image, all the objects are measured.
5. Image Recognition: Each object in an image can be distinguished.
Transform is a mathematical toolto transform a signa. Hence it is required
for :

 Mathematical convenience

.in
es
 To extract more information .
ot
un
m

Figure 1 : Transformation Concept

29
Digital Image Processing 3.3 IMAGE TRANSFORMS
Image transfor is represenatation of image . it is done for following two
reasons :
1. To isolate critical component of image so that they are accessible
directly.
2. It will make the image data compact so that they can be efficiently
stored and transmitted.
Different types of image transforms that wil be discussed are :

 Fourier transform

 Walschtransform

 Hadamardtransform

 Discrete Cosine transform

.in
 Wavelet transform etc
Classification of image transforms : they can be classified on the basis of
es
nature of fuctions as follows :
ot
un
m

Figure 2 :Classification of Image transforms

3.4 FOURIER TRANSFORM

The Fourier transform states that that the non periodic signals whose area
under the curve is finite can also be represented into integrals of the sines
and cosines after being multiplied by a certain weight.

30
The Fourier transform is a representation of an image as a sum of complex Image Transforms
exponentials of varying magnitudes, frequencies, and phases. The Fourier
transform plays a critical role in a broad range of image processing
applications, including enhancement, analysis, restoration, and
compression.
The Fourier transform also has many wide applications that include ,
image compression (e.g JPEG compression) , filtrering and image
analysis.

A continuous Time Fourier Transfrom (CTFT) is defined as :

.in
A continuous time signal x(t) is converted to discrete tie signal x(nT)
es
using sampling process where T is the sampling interval.
ot
un

The fourier transform of finite energy Discrete time signal is given by :

m

31
Digital Image Processing After a series of transformations :

3.5 2D DISCRETE FOURIER TRANSFORM

Functioning with the Fourier transform on a computer usually comprises a
form of the transform called as the discrete Fourier transform (DFT). A
discrete transform is a transform whose input and output values are
discrete samples, making it convenient for computer manipulation. There
are two principal reasons for using this form of the transform:
 The input and output of the DFT are both discrete, which makes it
convenient for computer manipulations.

.in
 There is a fast algorithm for computing the DFT known as the fast
Fourier transform (FFT).
es
ot
un
m

where R(k,l) represents the real part of spectrum and l(k,l) the imaginary
part.

32
 Image Transforms
3.6 PROPERTIES OF 2D-DFT
There are many properties of the transformation that give vision into the
content of the frequency domain representation of a signal and allow us to
manipulate singals in one domain or the other.
Following are some of the properties :
1. Separable property : It is computed in two steps by successive 1D
operations on rows and columns of an image .

.in
Figure 3 :Computation of 2D-DFT bySeparable Property
es
2. Spatial Shift property: Since both the space and frequency domains
are considered periodic for the purposes of the transforms, shifting
means rotating around the boundaries.
ot
un
m

3. Periodicity Property :

4. Convolution property : Convolution is a mathematical tool for

merging two signals to produce a third signal. In other words, the
convolution can be defined as a mathematical operation that is used to
express the relation between input and output an LTI
system.convoltion in spatial domain is same as multiplication in
frequency domain .

33
Digital Image Processing 5. Correlation property : It is used to find relative similarity in between
two signals.When we find similarity between of a signal to itself it is
called as autocorrelation . On the other hand when we find similarity
between two signals it is called as cross correlation. The cross
correlation between two signals is same as performing convolution of
one sequence with the folded version of other sequence.Thus this
property says that correlation of two sequneces in time domain is same
as multiplication of DFT of one sequence and time reversal of DFT of
another sequence in frequency domain.
6. Scaling property : Just as in one dimension, shrinking in one domain
causes expansion in the other for the 2D DFT. This means that as an
object grows in an image, the corresponding features in the frequency
domain will expand.Sclaing is used mainly to shrink or expand the size
of an image.

.in
es
7. Conjugate symmetry :
ot
un

8. Orthogonality property :
m

9. RoatationProperty : It states that if a function is rotated y an angle

then its Fourier transform also rotates by ana equal amount.

34
10. Multiplication by Exponent: Image Transforms

All the above properties in nutshell are as follows :

Table 1 : Properties of DFT

.in
es
3.7 OTHER TRANSFORMS
ot

Some other transforms used in image processing are :

1. Walsh Transform : The Walsh-Hadamard transform is a non-
un

sinusoidal, orthogonal transform widely used in signal and image

processing. In this transform, the signal is decomposed into a set of
basis functions (similar to harmonics in Fourier).
m

2. Hadamard Transform: Same as Walsh transform with the difference

that rows of the transform matrix are re-ordered.The elements of
mutually orthogonal basis vectors of a Hadamard transform are either
+1 or -1.

35
Digital Image Processing 3. HAAR transform: Haar wavelet compression is an efficient way to
perform both lossless and lossy image compression. It relies on
averaging and differencing values in an image matrix to produce a
matrix which is sparse or nearly sparse. A sparse matrix is a matrix in
which a large portion of its entries are 0.it is based on class of
orthogonal matrices whose elements are either +1 , -1 or 0 multiplied
by powers of .Algorithm to generate HAAR basis is :

.in
es
ot
un
m

4. Slant Transform: A new unitary transform called the slant transform,

specifically designed for image coding, was developed by Enomoto
and Shibata. The transformation possesses a discrete sawtoothlike
basis vector which efficiently represents linear brightness variations
along an image line. The slant transformation has been utilized in
several transform image-coding systems for monochrome and color
images
5. Discrete Cosine Transform: It is a transform that is mainly used in
compression algorithms. It transforms data points in a spatial domain
into a frequency domain. This makes it easier to find the repetition of
patterns. Like any other transform, it is also invertible. This means we
can return the actual data points if the transforms are given.It was
developed by Ahmad, Natrajan and Rao in 1974.

36
 Image Transforms

Where k varies from 0 to N-1

6. KL Transform: Named after Karim Karhunen and Michel Loeve
Transform (KLT), it is also known as the Hotelling trans- form or
Principal Component Analysis (PCA). It is based on the statistical
properties of the image and has several important properties that make
it useful for image processing particularly for image compression.The
main purpose of image compression is to store the image in fewer bits
as compared to original image, now data from neighboring pixels in an
image are highly correlated.More image compression can be achieved
by de-correlating this data. The KL transform does the task of de-
correlating the data thus facilitating higher degree of compression.It is

.in
used in clustering analysis and image compression.There are four
major steps in order to find the KL transform :-
a. Find the mean vector and covariance matrix of the given image x.
es
b. Find the Eigen values and then the eigen vectors of the covariance
matrix
ot

c. Create the transformation matrix T, such that rows of T are eigen

vectors
d. Find the KL Transform
un

Comparison of various image transforms

m

Table 2 :Comparison of Transforms

3.8 SUMMARY
In this chapter we studied why image transforms are required in digital
image processing , the different types available .

37
Digital Image Processing 3.9 EXERCISES
1. Why we need image transforms
2. What are different types of transforms?
3. Explain properties of DFT.
4. What is main difference between Walsh and Hadamard transform?
5. Explain Fourier transform and its properties .
6. What are advantages of Walsh over Fourier transform?

3.10 REFERENCES
1) Digital Image Processing, S Jayaraman, S Esakkirajan, T Veerakumar,
Tata McGraw-Hill Education Pvt. Ltd., 2009
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E

.in
Woods, Pearson, 2008
3) Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. EsakkirajanAnd T. Veerakumar, 2016
es

ot
un
m

38
 4
IMAGE ENHANCEMENT
Unit Structure
4.0 Objectives
4.1 Introduction
4.2 Image Enhancement in spatial domain
4.2.1 Point operation
4.2.2 Mask operation
4.2.3 Global operation
4.3 Enhancement through Point operations

.in
4.3.1 Types of point operations
4.4 Histogram manipulation
es
4.5 LinearGray Level Transformation
4.6 Nonlinear Gray Level Transformation
ot

4.6.1 Thresholding
4.6.2 Gray-level slicing
un

4.6.3 Logarithmic transformation

4.6.4 Exponential transformation
m

4.6.5 Power law transformation

4.7 Local or neighborhood operation
4.7.1 Spatial filtering
4.7.2 Linear filtering
4.7.3 Mean filter/ Average filter/ Low pass filter
4.7.4 Weighted average filter
4.7.5 Bartlett filter
4.7.6 Gaussian filter
4.8 Median Filter
4.9 Spatial domain High pass filtering

39
Digital Image Processing 4.9.1 High boost filtering
4.9.2 Unsharp masking
4.10 Bit-plane slicing
4.11 Image Enhancement in frequency domain
4.12 Homomorphic filter
4.13 Zooming operation
4.14 Image Arithmetic
4.14.1 Image addition
4.14.2 Image subtraction
4.14.3 Image multiplication
4.14.4 Image division

.in
4.15 Summary
4.16 List of References
es
4.17 Unit End Exercises

4.0 OBJECTIVES
ot

 To understand the image enhancement in spatial and frequency

domain
un

 To get familiar with point and mask operations in spatial domain

 To learn different types of transformations on an image and the
filtering operations
m

4.1 INTRODUCTION
The goal of image enhancement is to make it easier for viewers to
understand the information contained in images. A better-quality image is
produced by an enhancement algorithm when it is used for a certain
application. This can be accomplished by either reducing noise or boosting
image contrast.
For presentation and analysis, image-enhancement techniques are used to
highlight, sharpen, or smooth visual characteristics. Application-specific
enhancement techniques are frequently created via empirical research.
Image enhancement techniques highlight particular image elements to
enhance how an image is perceived visually.
Image-enhancement methods can be divided into two groups in general,
including
(1) A method in the spatial domain; and (2) A method in the transform
domain.
40
Whereas the transform domain approach works with an image's Fourier Image Enhancement
transform before transforming it back into the spatial domain, the spatial
domain method operates directly on pixels. Because they are quick and
easy to use, histogram-based basic improvement approaches can produce
results that are acceptable for some applications. By removing a portion of
a filtered component from the original image, unsharp masking sharpens
the edges. Unsharp masking is a technique that has gained popularity as a
diagnostic aid.

4.2 IMAGE ENHANCEMENT IN SPATIAL DOMAIN

The alteration of pixel values is the focus of the spatial domain approach.
The three basic categories of the spatial domain technique are i) point
operation, (ii) mask operation, and (iii) global operation.

4.2.1 Point operation

In point operation, each pixel is modified by an equation that is not
dependent on other pixel values. The point operation is illustrated in

.in
Figure 1. The point operation is represented by
es
In point operation, T operates on one pixel, or there exists a one-to-one
mapping between the input image f (m, n) and the output image g (m, n).
ot
un
m

Figure 1: Point operation

4.2.2 Mask operation

As shown in Figure 2, each pixel is altered during the mask operation in
accordance with the values in a small neighbourhood. Spatial low-pass
filtering with a box filter or median filter are examples of mask operations.
In \smask operation, the operator T operates on the neighbourhood of
pixels.Here, mask is a little matrix whose elements are frequently referred
to as weights. Every mask has a history. A symmetric mask's roots are
typically found in its centre pixel position. Depending on the desired
usage, any pixel location may be used as the origin for non-symmetric
masks.

41
Digital Image Processing

Figure 2: Mask processing

4.2.3 Global operation

All of the image's pixel values are taken into account during global
operation. Frequency domain operations are often global operations.

4.3 ENHANCEMENT THROUGH POINT OPERATIONS

Each pixel value is mapped to a new pixel value in point operation. Point
operations essentially have no memory. The enhancement at any position

.in
in a point operation depends only on the image value there.The point
operation, which is shown in Figure 3, converts the input picture f (m, n)
into the output image g (m, n).It is clear from the figure that every f (m, n)
es
pixel with the same grey level maps to a single grey value in the final
image.
ot
un
m

Figure 3: Illustration of point operation

4.3.1 Types of point operations

Some of the examples of point operation include (i) brightness
modification, (ii) contrast manipulation, and (iii) histogram manipulation

1] Brightness modification
The value assigned to each pixel in an image determines how bright it is.
A constant is either added to or subtracted from the luminance of each
sample value to alter the brightness of the image. By adding a constant
value to each and every pixel in the image, the brightness can be

42
raised.The brightness can also be reduced by deducting a fixed amount Image Enhancement
from each and every pixel in the image.

 Increasing the brightness of an image: A simple method to increase the

brightness value of an image is to add a constant value to each and
every pixel of the image. If f [m, n] represents the original image then
a new image g [m, n] is obtained by adding a constant k to each pixel
of f [m, n].This is represented by
g[m, n] = f [m, n] + k

Figure 4: Brightness enhancement

 Decreasing the brightness of an image: The brightness of an image can

.in
be decreased by subtracting a constant k from all the pixels of the
input image f [m,n]. This is represented by
g [m, n] = f [m, n] – k
es
ot
un

Figure 5: Brightness suppression

2] Contrast adjustment
Contrast adjustment is done by scaling all the pixels of the image by a
m

constant k. It is given by

g [m, n] = f [m, n] ∗ k
Changing the contrast of an image, changes the range of luminance values
present in the image
Specifying a value above 1 will increase the contrast by making bright
samples brighter and dark samples darker, thus expanding on the range
used. A value below 1 will do the opposite and reduce a smaller range of
sample values. An original image and its contrast-manipulated images are
illustrated in Figure 6.

43
Digital Image Processing

Figure 6: Contrast manipulation

4.4 HISTOGRAM MANIPULATION

Histogram manipulation generally involves changing the histogram of an
input image to enhance the image's visual appeal. One must have a
foundational understanding of the histogram of the image in order to
comprehend histogram manipulation.

.in
A] Histogram: The gray-level values of an image are plotted against the
number of instances of each type of grey in the image to create the
histogram. A useful summary of an image's intensities is provided by the
histogram, but it is unable to transmit any knowledge of the spatial
es
connections between pixels. Further information on image contrast and
brightness is given by the histogram.
ot

1. The lowest grey levels will be concentrated in the histogram of a dark

image.
2. A bright image will have a greater grey level cluster in the histogram.
un

3. The histogram won't be evenly distributed for a low-contrast image;

instead, it will be narrow.
m

4. The histogram will have an equal spread in the grey level for a high-
contrast image.
Image brightness may be improved by modifying the histogram of the
image.
B] Histogram equalisation: In order to distribute the grey levels of an
image uniformly across its range, a procedure called equalisation is used.
Based on the image histogram, histogram equalisation reassigns the
brightness values of pixels. The goal of histogram equalisation is to
produce a picture with a histogram that is as flat as feasible. More
aesthetically acceptable outcomes are produced by histogram equalisation
across a wider spectrum of photos.
C] Procedure to perform Histogram equalization : Histogram
equalisation is done by performing the following steps:
1. Find the running sum of the histogram values.

44
2. Normalise the values from Step (1) by dividing by the total number of Image Enhancement
pixels.
3. Multiply the values from Step (2) by the maximum gray-level value and
round.
4. Map the gray level values to the results from Step (3) using a one-to-
one correspondence.
Example: Perform histogram equalisation on the following image:

Solution: The maximum value is found to be 5. We need a minimum of 3

bits to represent the number. There are eight possible gray levels from 0 to

.in
7. The histogram of the input image is given below:
es
Step 1:Compute the running sum of histogram values. The running sum of
ot

histogram values is otherwise known as cumulative frequency distribution

un
m

Step 2: Divide the running sum obtained in Step 1 by the total number of
pixels. In this case, the total number of pixels is 25.

Step 3:Multiply the result obtained in Step 2 by the maximum gray-level

value, which is 7 in this case

45
Digital Image Processing

The result is then rounded to the closest integer to get the following table:

.in
Step 4: Mapping of gray level by a one-to-one correspondence:
es
ot
un
m

The original image and the histogram equalised image are shown side by
side.

4.5 LINEAR GRAY LEVEL TRANSFORMATION

An image's linear transformation is a function that converts each pixel's
grayscale value into a different grayscale at the same location using a
linear function. The linear transformation is represented by

46
 Image Enhancement

The linear transformation is given by g(m, n) = T [ f (m, n)].

Image negative or inverse transformation:

Light and dark are turned around in the inverse transformation. An image
negative is an illustration of an inverse transformation. Each pixel is
subtracted from the highest pixel value to create a negative image. The
negative picture for an 8-bit image can be created by reversing the scaling
of the grey levels according to the transformation
g (m, n) = 255 − f (m, n)
The graphical representation of negation is shown figure 7.

.in
es
ot

Figure 7: Graphical representation of negations

un

Negative images are useful in the display of medical images and

producing negative prints of images.
m

4.6 NONLINEAR GRAY LEVEL TRANSFORMATION

Non-linear transformation maps small equal intervals into non-equal
intervals. Following are few of the non-linear transformations:

4.6.1 Thresholding
To extract a portion of an image that contains all the information,
thresholding is necessary. The problem of segmentation in general
includes thresholding. Hard thresholding and soft thresholding are two
major categories for thresholding.The procedure of hard thresholding
entails reducing coefficients with absolute values below the threshold to
zero. Another technique is soft thresholding, which shrinks the nonzero
coefficients towards zero by first setting to zero coefficients whose
absolute values are below the threshold.

47
Digital Image Processing 4.6.2 Gray-level slicing
The purpose of gray-level slicing is to highlight a specific range of gray
values. Two different approaches can be adopted for gray-level slicing
(a) Gray-level Slicing without Preserving Background This displays
high values for a range of interest and low values in other areas. The main
drawback of this approach is that the background information is discarded.
(b) Gray-level Slicing with Background In gray-level slicing with
background, the objective is to display high values for the range of interest
and original gray level values in other areas. This approach preserves the
background of the image.

4.6.3 Logarithmic transformation

The logarithmic transformation is given by
g (m, n) = clog2( f (m, n) + 1)

.in
This type of mapping spreads out the lower gray levels. For an 8-bit
image, the lower gray level is zero and the higher gray level is 255. It is
desirable to map 0 to 0 and 255 to 255. The function g (m, n) = clog( f (m,
n) + 1) spreads out the lower gray levels.
es
4.6.4 Exponential transformation
In multiplicative filtering processes, exponential transformation has
ot

primarily been utilized in conjunction with logarithmic transformation. An

exponential transfer function has the effect of extending high-contrast
edges while contracting low-contrast edges in an image. This is not a
un

desired enhancement transformation because it typically results in images

with less visible detail than the original.

4.6.5 Power law transformation

m

A physical device like a CRT does not produce light with a linear
relationship to the input signal.The intensity generated at the display's
surface is about the applied voltage raised by a factor of 2.5. Gamma is the
name given to the numerical value of this power function's exponent. To
obtain accurate intensity reproduction, this non-linearity must be adjusted.
The formula for the power law transformation is

where f(m, n) is the input image and g(m, n) is the output image. Gamma
(γ) can take either integer or fraction values.

4.7 LOCAL OR NEIGHBORHOOD OPERATION

Neighborhood operation modifies an image's pixels based on how the
pixels in their immediate vicinity work. The term "convoluting a mask
with an image" is frequently used to describe linear spatial filtering.
48
Convolution masks or convolution kernels are other names for the filter Image Enhancement
masks.

4.7.1 Spatial filtering

Using spatial filtering, the original picture pixel value that corresponds to
the kernel's centre is replaced with the sum of the original pixel values in
the region that the kernel corresponds to, multiplied by the kernel weights.

4.7.2 Linear filtering

Each pixel in the input image is replaced with a linear combination of
brightness of nearby pixels in linear filtering. In other words, each pixel
value in the output image represents a weighted sum of the pixels nearby
the corresponding pixel in the input image. A picture can be both
sharpened and smoothed via linear filtering. Using a convolution mask,
one can create a spatially invariant linear filter.The linear filter is spatially
variable if various filter weights are applied to various portions of the
image.

.in
4.7.3 Mean filter/ Average filter/ Low pass filter
The average of all the values in the immediate area is used to replace each
pixel with the mean filter. The degree of filtration is determined by the
es
size of the neighbourhood. Each pixel in a spatial averaging operation is
replaced by the weighted average of the pixels in its immediate vicinity.
The low-pass filter removes the sharp changes that cause blurring and
ot

keeps the smooth area of the image. The following is the 3 by 3 spatial
mask that may execute the averaging operation:
un
m

It should be noted that for a low-pass spatial mask, the elements' sum is
equal to 1. The larger the mask becomes, the more blurring there will be.
In order to precisely find the core pixel, the mask's size is typically
unusual.

49
Digital Image Processing 4.7.4 Weighted average filter
The mask of a weighted average filter is given by

The weighted average filter gets its name because it is clear from the mask
that the closest pixels to the centre are weighted more heavily than the
farthest pixels.The pixel that has to be updated is changed by multiplying
the value of the neighbouring pixel by the matrix's weights and dividing
the result by the sum of the matrix's coefficients.

4.7.5 Bartlett filter

In the spatial domain, the bartlett filter is a triangle-shaped filter. A bartlett
filter can be created by taking the product of two box filters in the
frequency domain or by convolving two box filters in the spatial domain.

.in
es
ot

4.7.6 Gaussian filter

un

A class of linear smoothing filters known as "Gaussian filters" select their

weights based on the characteristics of a Gaussian function. The Gaussian
kernel is frequently employed for smoothing. The formula for the
Gaussian filter in continuous space is
m

The expression above demonstrates the separability of a Gaussian filter. A

extremely effective filter for eliminating noise derived from a normal
distribution is the Gaussian smoothing filter. A particular type of
averaging called Gaussian smoothing uses a 2D Gaussian as its kernel.

4.8 MEDIAN FILTER

Similar to the mean filter, the median filter in image processing is
typically used to minimize noise in an image. To preserve relevant detail
in the image, it frequently performs better than the mean filter.

50
Median Filter is a simple and powerful non-linear filter. Image Enhancement

 It is employed to lessen the degree of intensity difference between

adjacent pixels.
 With this filter, the pixel value is swapped out for the median value.
 In order to calculate the median, all the pixel values are sorted first
into ascending order, and then the middle pixel value is substituted for
the pixel being calculated.
In order to determine if a pixel is typical of its surroundings, the median
filter examines each pixel in turn and looks at its close neighbours. The
median of those values is used to replace the pixel value rather than just
the mean of its neighbours' pixel values. The median is derived by placing
the pixel under consideration in the middle of the neighborhood's pixel
values, which have been sorted into numerical order. The average of the
two middle pixel values is used (if the neighbourhood under consideration
has an even number of pixels).
Example: Compute the median value of the marked pixel using 3x3 mask

.in
es
ot
un
m

4.9 SPATIAL DOMAIN HIGH PASS FILTERING

Image sharpening's primary goal is to draw attention to the image's minute
features. The high-frequency components are enhanced through image
sharpening. The spatial filter or spatial mask which performs image
sharpening is given below:

The sum of all the weights is zero; this implies that the resulting signal
will have zero dc value.

51
Digital Image Processing 4.9.1 High boost filtering
High-frequency emphasis filter is another name for a high-boost filter. To
help with visual interpretation, a high-boost filter is utilized to keep some
of the low-frequency components. Before deleting the low-pass image in
high-boost filtering, the input image f(m, n) is amplified by an
amplification factor A.
The expression for the high-boost filter then becomes
High boost = A x f(m,n) - low pass
Adding and subtracting 1 with the gain factor, we get

Thus,

4.9.2 Unsharp masking

.in
es
One of the methods frequently used for edge enhancement is unsharp
masking. This method tips the balance of the image towards the sharper
material by subtracting the smoothed version of the image from the
ot

original image. Below is a description of how to accomplish unsharp

masking:
1. Apply an image blur filter.
un

2. Subtract from the original image the result from Step 1.

3. Add a weighted fraction to the result from Step 2 to get the final result.
m

4. Join the original image to the outcome from Step 3.

The equation for the unsharp masking operation is

4.10 BIT-PLANE SLICING

The gray level of each pixel in a digital image is stored as one or more
bytes in a computer. For an 8-bit image, 0 is encoded as 0 0 0 0 0 0 0 0,
and 255 is encoded as 1 1 1 1 1 1 1 1. Any number between 0 and 255 is
encoded as one byte. The bit in the far-left side is referred as the Most
Significant Bit (MSB), because a change in that bit would significantly
52
change the value encoded by the byte. The bit in the far right is referred as Image Enhancement
the Least Significant Bit (LSB), because a change in this bit does not
change the encoded gray value much. The bit-plane representation of an
eight-bit digital image is shown in figure 8.

Figure 8: Bit plane representation of a digital image

One or more bits of the byte are used for each pixel when an image is
represented using the bit-plane slicing technique. The original grey level is
converted to a binary representation because one can only use the MSB to
represent a pixel.

.in
Bit plane slicing has three major purposes:
i) converting a grayscale image to a binary image;
es
ii) representing an image with fewer bits and compressing it to a smaller
size; and
iii) enhancing the image by sharpening certain areas.
ot

4.11 IMAGE ENHANCEMENT IN FREQUENCY

DOMAIN
un

Frequency is the rate at which a particular periodic event repeats itself.

Spatial frequency in image processing is the fluctuation of image
brightness with image position in space. A signal that changes over time
can be converted into a sequence of easy periodic variations. A signal is
m

broken down into a collection of sine waves with varying frequency and
phase using the Fourier transform. Because a Fourier transform is entirely
reversible, if we compute a picture's Fourier transform and then
immediately inverse convert the result, we can recover the original image.
On the other hand, we can emphasize some frequency components and
attenuate others by multiplying each element of the Fourier coefficient by
an appropriately selected weighting function.After computing an inverse
transform, the corresponding changes in the spatial form can be observed.
Frequency domain filtering, often known as Fourier filtering, is the
process of selectively enhancing or suppressing frequency components.
The adjacency relationship between the pixels is described by the spatial
representation of image data. The frequency domain representation, on the
other hand, groups the visual data based on the frequency distribution. In
frequency domain filtering, the picture data is divided into different
spectral bands, each of which represents a particular range of image
information. Frequency domain filtering is the method of selective
frequency inclusion or exclusion.
53
Digital Image Processing It is possible to perform filtering in the frequency domain by specifying
the frequencies that should be kept and the frequencies that should be
discarded. Spatial domain filtering is accomplished by convolving the
image with a filter kernel. We know that
Convolution in spatial domain = Multiplication in the frequency domain
If filtering in spatial domain is done by convolving the input image f (m,
n) with the filter kernel h(m, n)

In the frequency domain, filtering corresponds to the multiplication of the

image spectrum by the Fourier transform of the filter kernel, which is
referred to as the frequency response of the filter

Here, F (k, l) is the spectrum of the input image and H (k, l) is the

.in
spectrum of the filter kernel. Thus, frequency domain filtering is
accomplished by taking the Fourier transform of the image and the Fourier
transform of the kernel, multiplying the two Fourier transforms, and taking
the inverse Fourier transform of the result. The multiplication of the
es
Fourier transforms needs to be carried out point by point. This point-by-
point multiplication requires that the Fourier transforms of the image and
the kernel themselves have the same dimensions. As convolution kernels
ot

are commonly much smaller than the images they are used to filter, it is
required to zero pad out the kernel to the size of the image to accomplish
this process.
un

4.12 HOMOMORPHIC FILTER

An image can be modeled as the product of an illumination function and
m

the reflectance function at every point. Based on this fact, the simple
model for an image is given by

This model is known as illumination-reflectance model. The illumination-

reflectance model can be used to address the problem of improving the
quality of an image that has been acquired under poor illumination
conditions.
In the above equation, f(n1,n2) represents the image, i(n1,n2) represents the
illumination component and r (n1,n2) represents the reflectance
component. For many images, the illumination is the primary contributor
to the dynamic range and varies slowly in space, while the reflectance
component r (n1,n2) represents the details of the object and varies rapidly
in space. If the illumination and the reflectance components have to be
handled separately, the logarithm of the input function f (n1,n2) is taken.

54
Because f(n1,n2) is the product of i (n1,n2) with r (n1,n2), the log of f (n1,n2) Image Enhancement
separates the two components as illustrated below:

where FI (k, l) and FR (k, l) are the Fourier transform of the illumination
and reflectance components respectively. Then, the desired filter function
H(k,l) can be applied separately to the illumination and the reflectance
component separately as shown below:

In order to visualize the image, inverse Fourier transform followed by

exponential function is applied. First, the inverse Fourier transform is
applied as shown below:

.in
The desired enhanced image is obtained by taking the exponential
es
operation as given below:
ot

Here, g (n1,n2) represents the enhanced version of the original image f

(n1,n2). The sequence of operation can be represented by a block diagram
un

as shown in figure 9
m

Figure 9: Homomorphic filtering block diagram

4.13 ZOOMING OPERATION

When an image has fine features, it could be difficult to see those elements
well in a monitor's normal display. Under these circumstances, zooming in
on the image enables viewing of fine features. The zooming function
allows for the expansion of an image. Holding a magnifying glass in front
of the screen is the same as zooming. The simplest method of zooming is
the "replication" of pixels. "Put the same value of the pixel in a grid of
NxN pixels for every pixel of the image," is the replication principle.
Diagrammatic representation of this is in figure 10.

55
Digital Image Processing

Figure 10: Zooming illustration

4.14 IMAGE ARITHMETIC

Several arithmetic processes, including image addition, subtraction,
multiplication and division are taken into account in image arithmetic.
Objects can be added to and subtracted from images using addition and
subtraction operations.

4.14.1 Image addition

Image addition is used to create double exposure. If f (m, n) and g(m, n)
represent two images then the addition of these two images to get the
resultant image is given by

.in
If multiple images of a given region are available for approximately the
es
same date and if a part of one of the images has some noise, then that part
can be compensated from other images available through image addition.
The concept of image addition is illustrated in figure 11.
ot
un
m

Figure 11: Image addition

In figure 11 there are three images (a), (b) and (c). Here, image (c) is
obtained by addition of images in (a) and (b).

4.14.2 Image subtraction

Image subtraction is used to find the changes between two images of a
same scene. The mathematical representation of image subtraction is given
by

To assess the degree of change in an area, two dates of co-registered

images can be used with the subtraction operation. Image subtraction can

56
be used to remove certain features in the image. The image subtraction is Image Enhancement
illustrated in the figure 12.

Figure 12: Image subtraction

The image subtraction is illustrated in figure 12.The images (a) and (b) are
subtracted to get the image (c).
4.14.3 Image multiplication
Image multiplication is basically used for masking. If the analyst is

.in
interested in a part of an image then extracting that area can be done by
multiplying the area by one and the rest by zero. The operation is depicted
in the figure 13.
es
ot
un
m

Figure 13: Image multiplication

From the figure, it is obvious that the multiplication operation is used to
extract specific information from the image. From the figure, it is obvious
that the coin of interest is highlighted.

4.14.4 Image division

Dividing the pixels in one image by the corresponding pixels in a second
image is commonly used in transformation.The operation is depicted in
the figure 14.From the figure it is obvious that the result of image division
is just opposite to that of image multiplication.

57
Digital Image Processing

Figure 14: Image division

4.15 SUMMARY
 By modifying intensity functions, picture spectral content, or a
combination of these functions, image enhancement aims to alter how
the image is perceived.

 Images can be enhanced by removing blurriness and noise, boosting

contrast, and bringing out features, for instance.

.in
 Either the spatial domain or the frequency domain can be used for
image enhancement. Application determines which image-
enhancement method should be used.
es
 The gray-level changes used in the spatial-domain picture
enhancement technique include image negative, image slicing, image
ot

thresholding, gamma correction, etc.

 Each pixel's value is recalculated in a point operation in accordance

un

with a specific transformation. A common aspect of point operations is

image brightness and contrast correction.

 The frequency of occurrence of the grey levels is indicated by the

m

histogram of an image. The histogram of two photos can be identical.

A histogram is rotation-invariant.

4.16 LIST OF REFERENCES

1. Digital Image Processing, S Jayaraman, S Esakkirajan, T Veerakumar,
Tata McGraw-Hill Education Pvt. Ltd., 2009.
2. Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008.
3. Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. Esakkirajan and T. Veerakumar, 2016
(https://scilab.in/textbook_companion/generate_book/125).

58
4.17 UNIT END EXERCISES Image Enhancement

1) Explain the image enhancement in spatial domain.

2) Describe the enhancement process through point operations.
3) Write a note on Histogram manipulation.
4) Explain the Linear Gray Level Transformation.
5) Write a note on Nonlinear Gray Level Transformation.
6) Describe Median Filter.
7) Write a note on Spatial domain High pass filtering.
8) Explain the concept of Bit-plane slicing.
9) Describe the process of Image Enhancement in frequency domain.
10) Write a note on Homomorphic filter.
11) Describe various zooming operations.

.in
12) State and explain various arithmetic operations performed on an
image.


es

ot
un
m

59
 5
BINARY IMAGE PROCESSING
Unit Structure
5.0 Objectives
5.1 Introduction
5.2 Mathematical morphology
5.3 Structuring elements
5.4 Morphological image processing
5.5 Logical operations
5.6 Morphological operations

.in
5.6.1Dilation
5.6.2 Erosion
es
5.6.3 Dilation and Erosion-based operations
5.7 Distance Transform
ot

5.7.1 Euclidean distance

5.7.2 City-Block distance
un

5.7.3 Chessboard distance

5.7.4 Quasi- Euclidean distance transform
m

5.8 Summary
5.9 List of References
5.10 Unit End Exercises

5.0 OBJECTIVES
 To get familiar with the concept of binary image processing and
various morphological operations involved

 To understand several binary image operations

 To get familiar with the fundamentals of distance transform

60
5.1 INTRODUCTION Binary Image Processing

The study of form, organization, and appearance is known as morphology.

A group of non-linear techniques known as mathematical morphology can
be used to an image to eliminate details that are smaller than a particular
reference shape. The initial definition of mathematical morphology's
operations as set operations demonstrated its use for handling sets of 2D
point sets. The morphological operations can be used to skelatonize, filter,
and extract the edges of a picture. Dilation, erosion, opening, and closing
are some of the fundamental morphological operations covered in this
chapter. They are followed by the characteristics of morphological
processes.Although binary, grayscale, and colour pictures can all be
subjected to morphological processes, our focus in this chapter is on using
various morphological procedures on binary images.An important subclass
of digital images are binary images, which have only two grey levels. The
final product of picture segmentation is typically a binary image.

5.2 MATHEMATICAL MORPHOLOGY

.in
A highly fruitful area of image processing is mathematical morphology. A
method for removing visual elements that are helpful for representation
and description is mathematical morphology. The technique was originally
es
developed by Matheron and Serra at the Ecole des Mines in Paris. The
collection of structural data on the visual domain serves as the inspiration.
Set theory serves as the sole foundation for the subject matter of
mathematical morphology. Several useful operators described in
ot

mathematical morphology can be used by employing set operations. In

mathematical morphology, sets stand in for the image's objects.For
un

instance, a binary image's black or white pixels collectively represent a

morphological description of the image. The sets in a binary image are the
3-D image domain's individuals with their integer elements. A 3-D tuple
with the elements x and y coordinates is used to represent each element in
m

the image. The development of image-segmentation techniques with a

variety of applications can be based on mathematical morphology, and it
also plays a significant part in techniques for picture description.

5.3 STRUCTURING ELEMENTS

When applied to an image, mathematical morphology is a group of non-
linear procedures that can be used to eliminate details that are smaller than
a particular reference shape, known as a structuring element. With its
various shapes and sizes, the structural element in a morphological
operation plays a significant role. A number of 0s and 1s in the structural
elements determine shape and size. The resultant value is applied to the
circle in figure 1 known as the centre pixel. Depending on the user's
perspective, this circle can be located wherever in the structuring element.
Figure 1 displays some examples of structural element references.

61
Digital Image Processing

.in
Figure 1: Example of structuring element reference

5.4 MORPHOLOGICAL IMAGE PROCESSING

es
By shifting a structuring element so that it is centred over each pixel of the
binary picture to be edited at some time, morphological operations are
defined. A logical operation is carried out on the pixels covered by the
ot

structuring element when it is centred over a specific area of the image,

producing a binary output. As seen in figure 2, the morphological image
un

processing resembles a convolution method. The structural element, like

the convolution kernel, can be any size and can include complements of 1s
and 0s.A specific logical action is carried out between the structuring
element and the underlying binary image at each pixel point. At that pixel
m

position in the output image, the binary outcome of that logical operation
is saved. The structuring element's size, content, and type of logical
operation all influence the impact that is produced. The binary picture and
structural element sets could be defined in 1, 2, or even higher dimensions
rather than being limited to sets on the 2D plane. Do the logical operation
if the structuring element is precisely positioned on the binary image; else,
leave the generated binary image pixel alone.

Figure 2: Morphological image processing

62
5.5 LOGICAL OPERATIONS Binary Image Processing

Logical operations includes AND, OR, NOT, EXOR operations, etc.

1] AND operation: The AND logic operation is similar to the intersection
operation of set theory and the AND operation of two images is shown in
Fig. 3. The AND logic operation computes the common area of images A
and B as a resultant image.

Figure 3: AND operation of images A and B

2] OR operation: The OR operation is similar to the union operation and

.in
the OR operation result of images A and B is shown in figure 4.
es
ot

Figure 4: OR operation of images A and B

un

3] NOT operation: The NOT operation is similar to the complement

function of set theory. In the NOT operation, the image pixel 1 is changed
to 0 and vice versa. The NOT operation of Image A is shown in figure 5.
m

Figure 5: NOT operation of image A

4] EXOR operation: In the XOR operation, if the pixels of Image A and
Image B are complementary to each other than the resultant image pixel is
black, otherwise the resultant image pixel is white. The XOR operation of
images A and B as shown in figure 6.

63
Digital Image Processing

Figure 6: EXOR operation of images A and B

5.6 MORPHOLOGICAL OPERATIONS

Dilatation and erosion are the two primary morphological processes. They
can be stated using a kernel that operates on a binary input image called X,
where white pixels signify uniform regions and black pixels signify region
boundaries.By translating a structural element, B, over the image points
and looking at the intersection of the translated kernel coordinates and
image coordinates, erosion and dilation operate conceptually.

5.6.1 Dilation

.in
The binary picture is expanded from its original shape during the dilation
process. The structural element determines how the binary picture is
enlarged. Compared to the image itself, this structuring element is smaller
es
in size; typically, the structuring element is 3 × 3. The structuring element
is reflected and shifted from left to right and from top to bottom
throughout the dilation process, similar to the convolution process. During
each shift, the procedure searches for any overlapping identical pixels
ot

between the structuring element and that of the binary picture. The pixels
under the centre location of the structuring element will be set to 1 or
black if there is an overlap.
un

Let us define X as the reference image and B as the structuring element.

The dilation operation is defined by
m

where Bˆ is the image B rotated about the origin. Equation 10.7 states that
when the image X is dilated by the structuring element B, the outcome
element z would be that there will be at least one element in B that
intersects with an element in X. If this is the case, the position where the
structuring element is being centred on the image will be ‘ON’. This
process is illustrated in figure 7. The black square represents 1 and the
white square represents 0.
Initially, the centre of the structuring element is aligned at position *. At
this point, there is no overlapping between the black squares of B and the
black squares of X; hence at position * the square will remain white. This
structuring element will then be shifted towards right. At position **, we
find that one of the black squares of B is overlapping or intersecting with
the black square of X. Thus, at position ** the square will be changed to
black. Similarly, the structuring element B is shifted from left to right and

64
from top to bottom on the image X to yield the dilated image as shown in Binary Image Processing
figure 7.
The dilation is an expansion operator that enlarges binary objects. Dilation
has many uses, but the major one is bridging gaps in an image, due to the
fact that B is expanding the features of X.

Figure 7: Dilation process

5.6.2 Erosion

.in
Dilation's opposite process is erosion. Erosion decreases an image if
dilation enlarges it. The structural element determines how the image is
reduced. With a 3x3 size, the structural element is typically smaller than
es
the image. When compared to greater structuring-element sizes, this will
guarantee faster computation times. The erosion process will shift the
structural element from left to right and top to bottom, almost identical to
the dilatation process. The method will check whether there is a complete
ot

overlap with the structuring element at the centre location, denoted by the
centre of the structuring element, or not.If there is no complete
un

overlapping then the centre pixel indicated by the centre of the structuring
element will be set white or 0.
Let us define X as the reference binary image and B as the structuring
element. Erosion is defined by the equation
m

The above equation states that the outcome element z is considered only
when the structuring element is a subset or equal to the binary image X.
This process is depicted in figure 8. Again, the white square indicates 0
and the black square indicates 1.
The erosion process starts at position *. Here, there is no complete
overlapping, and so the pixel at the position * will remain white. The
structuring element is then shifted to the right and the same condition is
observed. At position **, complete overlapping is not present; thus, the
black square marked with ** will be turned to white. The structuring
element is then shifted further until its centre reaches the position marked
by ***. Here, we see that the overlapping is complete, that is, all the black
squares in the structuring element overlap with the black squares in the

65
Digital Image Processing image. Hence, the centre of the structuring element corresponding to the
image will be black.
Erosion is a thinning operator that shrinks an image. By applying erosion
to an image, narrow regions can be eliminated, while wider ones are
thinned.

Figure 8: Erosion process

5.6.3 Dilation and erosion-based operations

.in
Erosion and dilation can be combined to solve specific filtering tasks. The
most widely used combinations are opening and closing.
1] Opening:Opening is based on the morphological operations, erosion
es
and dilation. Opening breaks up thin strips, smoothes the interior of object
contours, and removes thin image areas. It involves applying erosion and
dilation procedures to the image in that order.
ot

The opening procedure is used to clean up CCD flaws and image noise.
By rounding corners from inside the object where the kernel utilizes fits,
un

the opening filters details and simplifies images. The mathematical

representation of the opening procedure is
m

where X is an input image and B is a structuring element.

2] Closing:The opposite of the opening process is the shutting operation. It
is a process of first dilatation and then erosion. A single-pixel object's
closure process fills in the tiny imperfections and spaces. It has the same
result as an opening operation in that it retains object forms and sizes and
smooths out features.
The mathematical representation of the closing procedure is

where X is an input image and B is the structuring element. Closing

protects coarse structures, closes small gaps and rounds off concave
corners.

66
5.7 DISTANCE TRANSFORM Binary Image Processing

An operator called the distance transform is often exclusively used with

binary pictures. Each feature pixel in a binary image is given a value by
the distance transform that is equal to its distance from the closest non-
feature pixels. It can be used to gather details on the form and placement
of the foreground pixels in relation to one another. It has been used in
numerous scientific areas, including surface reconstruction, pattern
recognition, and medical imaging. When comparing binary images, the
distance transforms are crucial, especially for images produced by local
feature-detection methods like edge or corner detection.
Depending on how the distance is calculated, there are various types of
distance transforms. The Gaussian distance transform, city block,
chessboard, and Euclidean distance can all be used to calculate the
distance between pixels. The following is a description of the distance
metrics utilized.

5.7.1 Euclidean distance

.in
The value at a pixel in the Euclidean distance transform is directly
proportional to the distance in Euclidean terms between that pixel and the
object pixel that is closest to it. The procedure is noise-sensitive because it
es
determines the value at the pixel of interest using the value at a single
object pixel.
ot

The straight-line distance between two pixels is given by

un

The input image and its Euclidean distance transform are shown in the
figure 9.
m

Figure 9: Euclidean distance transform

5.7.2 City-Block distance

The city-block distance of two pixels

This metric measures the path between the pixels based on a four-
connected neighbourhood, as shown in figure 10. For a pixel ‘p’ with the
coordinates (x, y), the set of pixels given by,
67
Digital Image Processing

Above equation is called its four neighbours, as shown in Fig. 10. The
input and its city block distance are shown in Fig. 11.

Figure 10: Four-connected neighbours

.in
Figure 11: City block distance transform
es
5.7.3 Chessboard distance
The chessboard distance of two pixels is given by
ot

The chessboard distance metric measures the path between the pixels
un

based on an eight-connected neighbourhood, as shown in figure 13. The

set of eight neighbour pixels is defined as
m

The above equation is called the eight connected neighbours as shown in

Fig. 12. The chessboard distance of the image and the original image is
shown in Fig. 13.

Figure 12: Eight connected neighbours

68
 Binary Image Processing

Figure 13: Chessboard distance transform

5.7.4 Quasi- Euclidean distance transform

The quasi-Euclidean metric measures the total Euclidean distance along a
set of horizontal, vertical and diagonal line segments. The quasi-Euclidean
distance of two pixels is given by

.in
The quasi-Euclidean distance transform of the image is shown in Figure
14
es
ot
un

Figure 14: Quasi-Euclidean distance transform

m

5.8 SUMMARY
 The principle behind morphological image processing is to probe an
image with a tiny form or pattern called a structuring element. To
carry out a certain activity, structural elements of various shapes may
be used.

 Dilatation and erosion are the two primary morphological processes.

 By successively moving the structuring element's origin to all feasible

pixel locations throughout the dilation procedure, the image is probed
with the structuring element; the result is 1 if the structuring element
and the picture have a non-zero intersection and 0 otherwise.

 The erosion procedure involves probing the image with the structuring
element by successively moving the structuring element's origin to all

69
Digital Image Processing potential pixel locations. If the structuring element is entirely
contained in the image, the result is 1, otherwise it is 0.

 Opening and closing morphological processes are morphological

filters that eliminate undesirable elements from an image.

5.9 LIST OF REFERENCES

1] Digital Image Processing, S Jayaraman, S Esakkirajan, T
Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009.
2] Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008.
3] Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. Esakkirajanand T. Veerakumar, 2016
(https://scilab.in/textbook_companion/generate_book/125).

5.10 UNIT END EXERCISES

.in
1] Explain the mathematical morphology.
2] What do you mean bystructuring elements?
es
3] Write a note on Morphological image processing.
4] Explain thelogical operations.
ot

5] State and describe the Morphological operations.

un

6] What is Dilation?
7]What is Erosion?
m

8] Describe the Dilation and Erosion-based operations.

9] Write a note on Distance Transform.
10] Explain Euclidean distance.
11] Describe City-Block distance.
12] What is Chessboard distance?
13] Explain Quasi- Euclidean distance transform.




70
 6
COLOUR IMAGE PROCESSING
Unit Structure
6.0 Objectives
6.1 Introduction
6.2 Formation of colour
6.3 Human perception of colour
6.4 Colour Model
6.4.1 RGB colour model
6.4.2 CMY colour model

.in
6.4.3 HIS colour model
6.4.4 YIQ colour model
es
6.4.5 YCbCr colour coordinate
6.5 Colour image quantization
ot

6.5.1 Classification of colour quantisation technique

6.6 Histogram of a colour image
un

6.6.1 Histogram equalization of a colour image

6.7 Summary
m

6.8 List of References

6.9 Unit End Exercises

6.0 OBJECTIVES
 To understand the human perception and concepts related to colour
image processing
 To learn different colour models
 To get familiar with colour histogram equalization

6.1 INTRODUCTION
The perception of colour is the result of incoming visible light on the
retina. The most visually arresting aspect of every image is its colour,
which also has a big impact on how scenically beautiful it is.

71
Digital Image Processing Understanding the nature of light is required in order to comprehend
colour. Light has two distinct natures. It exhibits wave and particle
characteristics. After striking the retina, photons cause electric impulses
that, once they reach the brain, are interpreted into colour. Light waves
with various wavelengths are seen as having various colours.The human
eye cannot, however, see every wavelength. The visible spectrum is made
up of wavelengths between 380 and 780 nanometers.

6.2 FORMATION OF COLOUR

A variety of physiochemical processes result in the colours that people see
every day.Some things act as light sources, while others do little more than
reflect incident light. Fundamentally, there are three different ways that
colours are formed: three processes: additive processing, subtractive
processing andpigmentation.

1] Additive colour formation

The spectral distributions for two or more light rays are added in additive

.in
colour creation.The total amount of photons in the same range that are
present in all of the component colours makes up the final colour.TV
monitors use the additive color-formation principle.
es
2] Subtractive colour formation
When light is transmitted through a light filter, subtractive colour creation
happens. A light filter partially transmits and partially absorbs the light
ot

that it receives. For instance, a green filter allows radiation from the green
portion of the spectrum to pass while blocking radiation from other
wavelengths. The resultant colour is made up of the wavelengths that can
un

pass through each filter when used in sequence. The projection of colour
slides onto a screen result in subtractive colour generation.

3] Colour formation by pigments

m

Colored particles suspended in a liquid make up a pigment. The light that

strikes these particles can be absorbed or reflected by them. A light beam
is scattered by the particles as it encounters a surface covered in pigment,
with consecutive and simultaneous actions of reflection, transmission, and
absorption. These occurrences define the type (colour) of light that the
surface reflects. One can see colours in a painting thanks to colour
pigmentation.

6.3 HUMAN PERCEPTION OF COLOUR

The 400 to 700 nm range of the light spectrum is detectable by the human
visual system. The two most crucial senses through which humans
perceive their environment are sight and hearing. According to estimates,
humans get 75% of their information visually. Perception, sight, or
comprehension are terms used to describe the use of visual information.
The idea for creating analogous paradigms for computers came from an
understanding of human perceptual processing capabilities. A brief
72
explanation of the anatomy of the human visual system is the best way to Colour Image Processing
introduce the intricate aspects of visual perception.The Human Visual
System (HVS) is a system for processing information that receives
spatially sampled images from the cones and rods in the eye and infers the
characteristics of the objects it sees. Cones are the photoreceptors in
charge of colour vision. The retina contains three different kinds of cones.
They are
i) L-receptors which are sensitive to light of long wavelengths
ii) M-receptors which are sensitive to light of middle wavelengths
iii) S-receptors which are sensitive to light of short wavelengths
The RGB sensors are often identified by the Greek letters rho (red),
gamma (green), and beta (blue). Fig. 1 displays the sensitivity curves for
rho, gamma, and beta.

.in
es
ot

Figure 1: Sensitivity of rho, gamma and beta curves

un

The strength of the colours one receives for each of the wavelengths in the
visual spectrum is determined by the sensitivity curves of the rho, gamma,
and beta sensors in the human eye. The visual system can create the
perception of colour by selective sensing of various light wavelengths.The
m

sensitivity of light or the lighting of things affects colour perception as

well. Cones are not activated at very low illumination levels. Yet, even
with very little illumination, rods are triggered.

6.4 COLOUR MODEL

By specifying a 3D coordinate system and a subspace that encompasses all
constructible colours within a specific model, colour models offer a
uniform approach to identify a specific colour. A model can be used to
specify any colour. Each colour model is tailored for a particular piece of
software or image-processing tool.

6.4.1 RGB colour model

In an RGB colour model, the three primary colours red, green and blue
form the axis of the cube which is illustrated in Fig. 2. Each point in the
cube represents a specific colour. This model is good for setting the
electron gun for a CRT.

73
Digital Image Processing

Figure 2: RGB colour cube

RGB is an additive colour model. From Fig. 3, it is obvious that
Magenta = Red + Blue
Yellow = Red + Green
Cyan = Blue + Green

.in
es
ot
un
m

Figure 3: RGB colour model

Limitations of RGB model:
(i) The RGB colour coordinates depend on the device. This suggests
that the RGB model will generally fail to accurately replicate the
same colour on different displays.
(ii) The RGB model is not consistently perceived. The implication is that
in all regions of the colour space, a unit of coordinate distance does
not equate to the same perceived colour difference.
(iii) Because this model is based on device signals rather than display
brightness values, it is challenging to relate it to the look of colours.

74
6.4.2 CMY colour model Colour Image Processing

Printers produce an image by reflective light, which is basically a

subtractive process. Printers commonly employ the CMY model. The
CMY model cube is illustrated in Fig. 4. The conversion between RGB
and CMY is done as follows
C = 1-R
M = 1-G
Y = 1-B

.in
es
Figure 4: CMY colour cube
CMY is a subtractive colour model. From Fig. 5, it is obvious that
ot

Magenta = White – Green

Cyan = White – Red
un

Yellow = White –Blue

m

Figure 5: CMY colour model

75
Digital Image Processing 6.4.3 HIS colour model
Hue, Saturation, and Intensity is referred to as HSI. The observer's
perceived main colour is represented by hue. It is a characteristic
connected to the predominant wavelength. The term "saturation" describes
the degree of purity or the quantity of white light combined with a hue.
Brightness is reflected in intensity. As hue and saturation correspond to
how colours are seen by humans, HSI decouples colour information from
intensity information, making this representation particularly helpful for
creating image-processing algorithms. HIS colour space is well-liked
because it is based on how people see colour.
The conversion from RGB space to HSI space is given below:

.in
es
The HIS colour model can be considered as a cylinder, where the
ot

coordinates r, θ, and z are saturation, hue and intensity respectively. The

cylindrical representation of the HIS colour model is illustrated in Fig 6.
un
m

Figure 6: Cylindrical representation of HIS colour model

6.4.4 YIQ colour model

The YIQ colour model is defined by the National Television System
Committee (NTSC). In this model, Y represents the luminance; and I and
Q describe the chrominance. Conversion between RGB and YIQ is given
below
76
 Colour Image Processing

6.4.5 YCbCr colour coordinate

The YCbCr colour coordinate was developed as part of the ITU-R BT.601
during the establishment of a worldwide digital video component standard.
The YCbCr signals are scaled and offset versions of the YIQ format. Y is
defined to have a nominal range of 16 to 235; Cb and Cr are defined to
have a range of 16 to 240, with zero signal corresponding to level 128.
There are several YCbCr formats such as 4:4:4, 4:2:2 and 4:1:1. The
sampling format implies that the sampling rates of Cb and Cr are one-half
of that of Y.

6.5 COLOUR IMAGE QUANTISATION

.in
The process of dividing the original colour space into larger cells in order
to reduce the size of a colour space is known as colour quantization. The
method of color-image quantization involves lowering the number of
colours in a digital colour image. A lossy image compression operation is
es
essentially what color-image quantization is. There are two main processes
in the process of quantizing colour images:
ot

(i) Choose a small number of colours from the available combinations of

red, green, and blue to create a colour map.
(ii) Assigning a colour from the colour map to each colour pixel in the
un

colour image.
The basic goal of colour image quantization is to transfer the original
colour image's set of colours to the quantized image's significantly smaller
m

range of colours. A colour palette is a more condensed collection of

exemplary hues. The discrepancy between the original and the quantized
images should be as little as possible after the mapping.

6.5.1 Classification of colour quantisation technique

Colour quantisation, in general, can be classified into (i) uniform colour
quantisation, and (ii) non-uniform colour quantisation
1] Uniform colour quantisation
The simplest colour quantization technique is uniform quantization. Each
colour cell in uniform colour quantization represents a cube whose edges
are parallel to the colour axes and each cell is the same size. Fig. 7
provides a visual picture of uniform colour quantization.

77
Digital Image Processing The formula for the uniform colour quantization scheme is

.in
Figure 7: Representation of uniform quantisation in two dimensions
es
2] Non-uniform colour quantisation
By utilizing non-uniform quantization, different thresholds are used on
ot

each colour axis individually to split the colour space. Based on an

investigation of the initial distribution in the colour space, the thresholds
are determined. Fig. 8 provides a visual illustration of non-uniform
un

quantization.
Non-uniform quantization over various thresholds is mathematically
represented as follows
m

78
 Colour Image Processing

Figure 8:Representation of non-uniform quantisation in two dimensions

6.6 HISTOGRAM OF A COLOUR IMAGE

The histogram of the image displays the grey level's occurrence frequency.
It provides the frequency with which a specific grey level appears in the
image. The histogram is anticipated to provide the frequency with which a

.in
specific colour appears in a colour image.

6.6.1 Histogram equalization of a colour image

es
A colour image's histogram equalisation is done by first converting the
RGB image to the YIQ(NTSC) format. The Y component alone is then
subjected to histogram equalisation. The I and Q parts are unaltered. The
equalised Y, unmodified I, and Q components are then converted to the
ot

RGB format in the histogram.

6.7 SUMMARY
un

 Between 400 nm (blue) and 700 nm is the range of visible light (red).

 Three different types of color-sensitive pigments, each responding to a

m

different spectrum of wavelengths, are present in cones.

 The tristimulus model of colour vision, which is commonly used to

describe how the responses of these many sensors combine to give us
colour perception.

 Red, green, and blue are replaced by the three primaries X, Y, and Z in
CIE.

 The Y component was selected to match luminous efficacy;Y=

luminance X, Z = chromaticity

 The colour space spanned by a set of primary colours is called a colour

gamut.

 The different colour models are RGB, CMY, YCbCr, YIQ, etc.

79
Digital Image Processing  The RGB model is used most often for additive models. In an RGB
model, adding a colour and its complement create white.

 The most common subtractive model is the CMY model. In a CMY

model, adding a subtractive colour and its complement create black.

 Each pixel of the colour image will have three values, one each for the
red, green and blue components.

 The histogram of a colour image can be considered to consist of three

separate one-dimensional histograms, one for each tricolour
component.

6.8 LIST OF REFERENCES

1) Digital Image Processing, S Jayaraman, S Esakkirajan,
T Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009.
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E

.in
Woods, Pearson, 2008.
3) Scilab Textbook Companion for Digital Image Processing, S.
Jayaraman, S. Esakkirajan and T. Veerakumar, 2016
es
(https://scilab.in/textbook_companion/generate_book/125).

6.9 UNIT END EXERCISES

ot

1) Explain the formation of colour.

2) Describe the human perception of colour.
un

3) Discuss different colour models.

4) Write a note on colour image quantization.
m

5) Explain the classification of colour quantisation technique.

6) Write a note on Histogram of a colour image.
7) Describe the Histogram equalization of a colour image.



80
 7
IMAGE SEGMENTATION
Unit Structure
7.0 Objectives
7.1 Introduction
7.2 Image segmentation techniques
7.3 Region approach
7.3.1 Region growing
7.3.2 Region splitting
7.3.3 Region splitting and merging

.in
7.4 Clustering techniques
7.4.1 Hierarchical clustering
es
7.4.2 Partitional clustering
7.4.3 k-means clustering
ot

7.4.4 Fuzzy clustering

7.5 Thresholding
un

7.5.1 Global Thresholding

7.5.2 Adaptive Thresholding
m

7.5.3 Histogram-based threshold selection

7.6 Edge-based segmentation
7.7 Edge detection
7.7.1 Gradient operator
7.7.2 Edge detection using first order derivates
7.7.3 Roberts kernel
7.7.4 Prewitt kernel
7.7.5 Sobel kernel
7.7.6 Frei-Chen edge detector
7.7.7 Second derivative method of detecting edges in an image

81
Digital Image Processing 7.7.8 Laplacian of gaussian (LOG)
7.7.9 Difference of gaussian filters (DoG)
7.7.10 Canny edge detectors
7.8 Edge Linking
7.9 Hough Transform
7.10 Summary
7.11 List of References
7.12 Unit End Exercises

7.0 OBJECTIVES
 To separate or combine different image elements that can be used to
develop objects of interest that can be the subject of investigation and
interpretation.

.in
 To get familiar with different segmentation techniques along with their
applicability
es
7.1 INTRODUCTION
Image segmentation is the division of a picture into groups of pixels that
ot

are homogeneous in terms of a particular criterion. Adjacent groupings

must be diversified and different groups must not intersect. Instead,then
being pixel-oriented, segmentation algorithms are area-oriented. The
un

division of the image into linked regions is the outcome of segmentation.

Therefore, segmentation is the process of splitting an image into useful
sections.
m

7.2 IMAGE SEGMENTATION TECHNIQUES

Image segmentation is classified into two categories:Local segmentation
and Global segmentation
i] Localsegmentation: Local segmentation divides sub-images, which are
tiny windows on a larger image, into separate segments. Local
segmentation has a lot fewer pixels accessible than global segmentation.
Local segmentation needs to use pixel data sparingly.
ii] Global segmentation:Segmenting an entire image is what global
segmentation is all about. The majority of global segmentation's work
involves segments with a sizable number of pixels. This strengthens the
reliability of estimated parameter values for global segments.

82
7.3 REGION APPROACH Image Segmentation

A region in an image is a collection of related pixels that are joined

together. The region technique assigns each pixel to a specific object or
region. The goal of the boundary technique is to identify and pinpoint the
actual regional borders. The edges are first found in the edge approach,
and then they are connected to create the necessary boundaries.

7.3.1 Region growing

An approach to picture segmentation known as "region growing" involves
looking at nearby pixels and adding them to a region class if no edges are
found. Each border pixel in the area goes through this process one more
time. If neighbouring regions are discovered, a region-merging technique
that preserves strong edges while dissolving weak edges is employed.
Starting with a seed is necessary for region growing. The seed could be a
single pixel, but ideally it would be a region. By including as many nearby
pixels as feasible that match the homogeneity requirement, a new segment

.in
is formed from the seed. After then, the resulting section is dropped from
the procedure. The remaining pixels are used to choose a new seed. This
keeps happening until every pixel has been assigned to a segment. The
settings for each segment must be updated when pixels are gathered. The
es
initial seed selected and the sequence in which neighbouring pixels are
evaluated may have a significant impact on the segmentation that results.
The choice of homogeneity criteria in image growth depends not only on
ot

the issue at hand but also on the kind of image that needs to be divided
into segments. The criteria used to determine whether a pixel should be
included in the area or not, the connectivity type used to identify
un

neighbours, and the approach taken to visit neighbouring pixels all affect
how regions evolve.
Comparing region-growing to traditional segmentation methods has
m

various benefits. The boundaries of the growing zones are incredibly

intertwined and narrow. In terms of noise, the algorithm is likewise highly
steady.

7.3.2 Region splitting

Top-down strategies involve region division. Starting with a complete
image, it is divided up so that the divided portions are more homogeneous
than the entire. Since splitting substantially restricts the forms of
segments, splitting alone is insufficient for acceptable segmentation.
Therefore, the split-and-merge algorithm, splitting followed by a merging
phaseis always preferred.

7.3.3 Region splitting and merging

An image-segmentation method that takes spatial information into account
is region splitting and merging. The following is the region-splitting-and-
merging method:

83
Digital Image Processing  Splitting
1]Let R stand for the full image in split 1.
2] Split or partition the image into successively smaller quadrant portions
using the predicate P.
As seen in figure 1, a quadtree structure provides a practical way to
visualize the splitting method. The root of a quadtree represents the
complete image, and each node represents a subdivision.

.in
Figure 1: (a) Splitting of an image (b) Representation by a quadtree
es
It is probable that contiguous zones with the same characteristics will be
found in the final partition. Applying merging and only merging nearby
regions whose combined pixels fulfil the predicate P will address this
flaw.
ot

 Merging
un

Any adjacent regions that are similar enough should be merged. The split-
and-merge algorithm's process is described below:
1. Commence with the entire picture.
m

2. Divide the variation into quadrants if it is too large.

3. Combine any nearby locations that are sufficiently comparable.
4. Continue iteratively repeating procedures (2) and (3) until no more
splitting or merging takes place.
The image that has been separated and combined is shown in Figure 1(a).
Figure 1(b) depicts the image's quadtree representation.

7.4 CLUSTERING TECHNIQUES

By grouping the patterns into clusters or groups so that they are more
similar to one another than to patterns belonging to different clusters, the
clustering technique aims to access the links between patterns of the data
set. In other words, clustering is the division of things into groups based
on shared characteristics. The clustering technique makes an effort to
extract a feature vector from the image's localised regions. Each pixel is
84
given the class of the closest cluster mean as part of the conventional Image Segmentation
clustering process. The two types of clustering techniques are partitional
and hierarchical. There are numerous types of clustering methods within
each category.

7.4.1 Hierarchical clustering

A proximity matrix is used in hierarchical clustering algorithms to show
how similar each pair of data points is to the others. The outcome is a tree
of clusters that displays the nested group of patterns and the similarity
thresholds at which groups shift. While the root node is set aside for the
full dataset and the leaf nodes are for individual data samples, the resulting
clusters are always formed as the internal nodes of the tree. The
procedures used to combine two small clusters or divide a large cluster
vary depending on the clustering techniques.
The agglomerative and divisive algorithm categories are the two main
types employed in the hierarchical clustering framework. Agglomerative
algorithms start with N singlepoint clusters and attempt to combine them

.in
into larger and larger clusters. There are three categories into which the
algorithm can be subdivided: (1) single-link algorithm, (2) complete-link
algorithm, and (3) minimum-variance algorithm. According to the shortest
distance between data samples from two clusters, the single-link algorithm
es
combines two clusters. In light of this, the algorithm permits a propensity
to create clusters with elongated shapes. The complete-link approach,
however, consistently yields compact clusters despite incorporating the
ot

maximum distance between data samples in clusters.The definition of the

dissimilarity measurement between two groups affects the quality of
hierarchical clustering. By minimising the cost function and creating a
un

new cluster with the least amount of cost function rise, the minimum-
variance algorithm joins two clusters. This approach, known as the
pairwise-nearest-neighbor algorithm, has generated a lot of attention in
vector quantization.
m

Divide and conquer clustering starts with the entire dataset in one cluster
and splits it repeatedly until single-point clusters are found on leaf nodes.
In order to combat agglomerative clustering, it employs a reverse
clustering approach. The divisive algorithm performs an exhaustive search
for all pairings of clusters for data samples on each node.
Hierarchical algorithms include COBWEB, CURE, and CHAMELEON.

7.4.2 Partitional clustering

The iterative optimization process used in partition-based clustering tries
to minimize an objective function f that gauges the effectiveness of
clustering. The division of each pattern to the nearest cluster and the
computing of the cluster centroids are the two learning processes that
make up partition-based clustering.Partition-based clustering frequently
begin with an initial solution that contains a predetermined number of
clusters. The optimality criterion is typically used to compute the cluster
centroids so that the objective function is minimized. The two types of
85
Digital Image Processing partitional algorithms are density-based partitioning and partitioning
relocation techniques.The first category of algorithms is further divided
into probabilistic clustering, K-medoids, and K-means. Density-based
partitioning, the second category of partitional algorithms, includes
DBSCAN, OPTICS DBCLAS, and DENCLUE.Compared to hierarchical
clustering techniques, partitional clustering techniques, such K-means
clustering and ISODATA, have the advantage of maximizing some
criterion function.

7.4.3 k-means clustering

The simplest method for unsupervised classification is the K-means
approach. There is no need for training data for the clustering methods.
Iterative processes such as K-means clustering are used. The K-means
clustering technique segments the image by categorizing each pixel in the
class with the nearest mean, iteratively computing a mean intensity for
each class, and clustering the data. The K-means clustering algorithm's
steps are listed below:

.in
1] Choose K initial clusters z1(l), z2(l), ………. zk(l).
2]Distribute the samples x among the K clusters at the kth iterative step
using the connection
es
ot

For i = 1, 2, …, K, i ≠ j, where Cj(k) denotes the set of samples whose

cluster centre is zj(k).
un

3] Determine the new cluster centreszj(k + 1), j = 1, 2,..., K, with the goal
of minimizing the sum of the squared distances between all locations in Cj
(k) and the new cluster. The sample mean of Cj (k) is the measurement
that minimizes this. Consequently, the new cluster centre is provided by
m

where Nj is the number of samples in Cj (k)

4] If zj (k + 1), j = 1, 2, ……, K, the algorithm has converged and the
procedure is terminated. Otherwise go to Step 2
The K-means algorithm's disadvantage is that, once K is selected, the
number of clusters is fixed and it always returns K cluster centres.

7.4.4 Fuzzy clustering

Depending on whether a pattern data belongs only to one cluster or to
numerous clusters with different degrees, clustering methods can be
categorized as either hard or fuzzy. In contrast to fuzzy clustering, which
assigns a value between zero and one to each pattern, hard clustering
86
assigns a membership value of zero or one to each pattern's data. Since Image Segmentation
they can more accurately depict the relationship between the input pattern
data and clusters, fuzzy clustering algorithms can generally be regarded as
being superior to those of their hard counterparts. Fuzzy clustering takes
advantage of the fact that each pattern has some graded membership in
each cluster in order to reduce a heuristic global cost function. The
clustering criterion permits numerous cluster assignments for each pattern.
The fuzzy K-means algorithm employs a gradient descent method to
estimate the class membership function and iteratively update the cluster
centroid.

Dissimilarity function
The dissimilarity function is a crucial parameter for determining how
different data patterns are from one another. The diversity and size of the
characteristics contained in patterns necessitates careful selection of the
dissimilarity functions. Different clustering outcomes are produced by
using various clustering dissimilarities. The Minkowski metricis a popular
way to express how different two data vectors are which is given by,

.in
es
p =1 means the L1 distance or Manhattan distance
p =2 means the Euclidean distance or L2 distance
ot

Below are the pre requisite of a distance function that must be fulfilled by
a metric
un
m

The triangle inequality, a stricter requirement for the definition of metric,

is defined as follows:

The case with l = ∞ in Minkowski metric comes to the maximum metric:

These measurements are translation-invariant but scaling-variant. The

Canberra metric is one of the scale-invariant differences given by

87
Digital Image Processing The cosine similarity is given by

The dot in this instance is the L2 norm in Euclidean space and indicates
the inner product of two vectors. The Mahalanobis distanceis another
method for generalizing the Euclidean distance to the scale invariant
situation and is expressed as

Where Σ: covariance matrix with respect to the entire data set

When using a model-based clustering technique like the EM clustering

.in
algorithm, the Mahalanobis distance is used as a measure of dissimilarity.
When used within the context of K-means clustering, the Mahalanobis
distance application is always hampered by the singularity problem and a
high computational cost.
es
7.5 THRESHOLDING
Segments made using thresholding techniques have pixels with
ot

comparable intensity. Setting limits in photos with solid objects resting on

a contrasting background using thresholding is a valuable method. There
are many segmentation techniques based on grey levels that use either
un

global or local image data. The thresholding approach requires a

background with a varied intensity level and an object with uniform
intensity. Such a picture can be divided into two areas using
m

straightforward thresholding.

7.5.1 Global Thresholding

The simplest and most used segmentation technique is global thresholding.
In global thresholding, the following stipulation is imposed along with a
threshold value of θ:

The equation mentioned above provides a detailed explanation of a

binarization technique but makes no mention of how to choose the
threshold parameter's value. It is necessary to choose the value of in the
best possible way. When pixels from distinct segments employ intensities
that overlap, global thresholding will suffer. If the cause is noise, a method
like the minimum-error method can be used to determine the parameters
of the underlying cluster and select the thresholds that will minimize the

88
classification error. Variable thresholding could be utilized if the overlap Image Segmentation
is brought on by differences in illumination across the image. One way to
see this is as a type of local segmentation.

7.5.2 Adaptive Thresholding

Global thresholding, also known as fixed thresholding, functions well
when the background is uneven but has a generally uniform grey level and
the objects of interest have an inner grey level that is reasonably
consistent. In many instances, the object contrast within an image varies
while the backdrop grey level is not consistent. In these circumstances, a
threshold that performs well in one section of the image could perform
poorly in other areas. In these situations, it is practical to employ a
threshold grey level that progressively changes depending on where it is in
the image.

7.5.3 Histogram-based threshold selection

The gray-level histogram of an image with an object against a contrasting
background is bimodal. The comparatively large number of points inside

.in
and outside the object are represented by the two peaks. The
comparatively few spots along the object's edge correlate to the dip
between the peaks. It is usual practice to determine the threshold grey
es
level using this dip.
The area function derivative for an object whose border is determined via
thresholding is the histogram
ot
un

where D is the gray level, A(D) is the area of the object obtained by
thresholding at gray level D, and H(D) is the histogram
m

If D is close to the histogram dip, raising the threshold from D to D +ΔD

just slightly reduces the area. As a result, setting the threshold near the
histogram's dip reduces the area measurement's susceptibility to minute
threshold selection errors.
The histogram will be noisy if the image that contains the object is noisy
and small. Unless the dip is exceptionally sharp, the noise can obscure or
at least make the placement unpredictable from one image to the next. By
applying a convolution filter or the curve-fitting method to smooth the
histogram, this effect can be somewhat mitigated.

7.6 EDGE-BASED SEGMENTATION

Edge-based segmentation uses the edges of a picture to extract spatial
information. Edges are when the homogeneity criteria for segments
discontinues. Local linear gradient operators like Prewitt, Sobel, and
Laplacian filters are typically used for edge detection. These operators
perform best with images that have few noise sources and sharp edges.
Some edge connecting may be necessary since the boundaries identified
89
Digital Image Processing by these operators may not always consist of a collection of closed linked
curves.

7.7 EDGE DETECTION

Edge detection is a technique for spotting significant changes in an image.
One of the main functions of the lowest levels of image processing is edge
detection. The boundaries between various objects are often formed by the
locations where there are abrupt changes in brightness. By calculating
intensity differences in local picture regions, these locations can be
located. That is, an area with high signals of change should be sought after
by the edge-detection algorithm. The majority of edge detectors measure
the intensity gradient at a specific location in the image.

7.7.1 Gradient operator

A gradient is a two-dimensional vector that points to the direction in
which the image intensity grows fastest. The gradient operator ∇ is given
by

.in
es
If the operator ∇ is applied to the function f then
ot
un
m

The gradient magnitude and gradient orientation are the two functions that
can be represented in terms of the directional derivatives. It is feasible to
determine the gradient's magnitude ||∇f|| and direction (∇f ).
The strength of the edge is determined by the gradient's magnitude, which
reveals how much the neighborhood's pixels differ from one another. The
gradient's size is given by

The greatest rate of growth of f (x, y) per unit distance in the gradient
orientation of |∇f| is determined by the gradient's size.

90
The gradient orientation indicates the direction of the biggest shift, which Image Segmentation
is probably the edge-crossing direction. The gradient's direction is
determined by

where the angle is measured with respect to the x-axis.

7.7.2 Edge detection using first order derivates

A digital pixel grid's derivative can be described in terms of differences.
The first derivative of an image with gray-value pixels must meet the
following requirements: it must be non-zero at the start of a gray-level step
or ramp, and it must be non-zero along the ramp (constant change in grey
values). It must also be zero in flat segments, or in areas of constant gray-
level values. Using a one-dimensional function f (x), the first-order
derivative can be found by

.in
A function of two variables, f (x, y), is an image. Only the partial
es
derivative along the x-axis is mentioned in the previous equation. It is
possible to identify pixel discontinuity in eight different directions,
including up, down, left, right, and along the four diagonals.
ot

By estimating the finite difference, the second technique of obtaining the

first-order derivative is provided:
un
m

The finite difference can be approximated as

Using the pixel coordinate notation and considering that j corresponds to

the direction of x, and i corresponds to the y direction, we have

91
Digital Image Processing

The majority of edge-detecting algorithms can be compared to gradient

estimators. The gradient computations must be estimated since the
gradient is a continuous function notion and the input signal (the image) is
a finite signal. Convolution is most frequently used in gradient calculation
because derivatives are shift-invariant and linear. For identifying edges,
numerous kernels have been presented.

7.7.3 Roberts kernel

The main goal is to calculate the differences between adjacent pixels. To
directly utilize{+ 1, -1} which calculates these differences, is one
technique to detect an edge. These are referred to as forward differences in
mathematics. In reality, the Roberts kernels are too small to accurately

.in
locate edges when there is noise. The Roberts operator for a cross-
gradient method is the easiest approach to implement the first-order partial
derivative.
es
ot
un

Implementing the partial derivatives shown above involves roughly

converting them to two 2x2 masks. The providers of the Roberts operator
masks are
m

These filters have the shortest aid, which results in a more precise
positioning of the edges, but the short support of the filters has the
drawback of being susceptible to noise.

7.7.4 Prewitt kernel

Judy Prewitt is honoured by the name of the Prewitt kernels. The concept
of central difference is the foundation of Prewitt kernels. In comparison to
the Roberts operator, the Prewitt edge detector is a considerably better
operator. Take a look at how the pixels are arranged around the core pixel
[i, j] in the illustration below:

92
 Image Segmentation

The partial derivates of the Prewitt operator are calculated as

In the above equations c is the constant thatsignifies the emphasis given to

pixels closer to the centre of the mask. Gx and Gy are the approximations
at [i, j].
Setting c = 1, the Prewitt operator mask is obtained as

.in
es
The support is longer on the Prewitt masks. The edge detector is less
susceptible to noise because the Prewitt mask differentiates in one
direction and averages in the other.
ot

7.7.5 Sobel kernel

un

The Sobel kernels bear Irwin Sobel's name. The Sobel kernel is based on
central differences, however when averaging, it gives more weight to the
central pixels. The Sobel kernels can be viewed as 3x3 approximations to
Gaussian kernels' first derivatives. The Sobel operator's partial derivates
m

are calculated as

The Sobel masks in matrix form are given as

A Sobel mask has superior noise-suppression qualities compared to a

Prewitt mask.

93
Digital Image Processing 7.7.6 Frei-Chen edge detector
By mapping the intensity vector with a linear transformation and then
identifying edges based on the angle between the intensity vector and its
projection onto the edge subspace, Frei-Chen mask edge detection is
performed. The normalized weights allow for the realization of Frei-Chen
edge detection. Unique Frei-Chen masks are masks that include all of the
basis vectors. This suggests that the weighted sum of nine Frei-Chen
masks is used to represent a 3x3 image area. The image is initially
confounded with each of the nine masks. The convolution outcomes of
each mask are then combined to produce an innerproduct. Following are
the nine Frei-Chen masks:

.in
es
ot

The first four Frei-Chen masks are used to represent edges, the next four
to represent lines, and the last mask to represent averages. The image is
un

projected onto the required masks for edge detection.

7.7.7 Second derivative method of detecting edges in an image

m

Finding the ideal edge is comparable to determining the largest or least

derivative point. It is possible to determine a function's maximum or
minimum value by differentiating the provided function and looking for
locations where the derivative is zero. The second derivative is obtained
by differentiating the first derivative. It is analogous to looking for
locations where the second derivative is zero to locate the best edges.
Images can be processed using differential operators, although zeros rarely
land perfectly on a pixel. They typically lie between pixels. By locating
the zero crossings, the zeros can be isolated. When one pixel is positive
and its neighbour is negative, this is known as a zero crossover. The
following are issues with zero-crossing methods:

 Zero crossing methods produce two-pixel thick edges.

 Zero crossing methods are extremely sensitive to noise.
For images, there is a single measure, similar to the gradient magnitude,
that measures the second derivative, which is obtained by taking the dot
product of ∇ with itself
94
 Image Segmentation

The operator ∇⋅∇ = ∇2 is called the Laplacian operator.

The Laplacian operator when applied to the function f, we get

The following difference equations can be used to express the Laplacian

.in
operation:
es
ot
un
m

The brightness values of each of the surrounding pixels are subtracted

from the central pixel via the Laplacian operator. The Laplacian's output is
non-zero when the neighbourhood has a discontinuity in the form of a
point, line, or edge. Depending on how close to the edge the central point
is, it could be either positive or negative.
Rotational invariance characterises the Laplacian operator. As long as they
are orthogonal, the Laplacian operator does not depend on the direction.

7.7.8 Laplacian of gaussian (LOG)

Noise in the input image is a significant factor in the Laplacian operator's
performance degradation. Prior performing edge enhancement, the image
95
Digital Image Processing might be smoothed to reduce the effects of noise. The Laplacian-of-
Gaussian (LOG) operator smoothes the image via convolution with a
kernel that has a Gaussian form and is then applied after.

7.7.9 Difference of gaussian filters (DoG)

The DoG filter is obtained by taking the difference of two Gaussian
functions. The expression of a DoG filter is given by

7.7.10 Canny edge detectors

.in
The fact that a Laplacian zero-crossing merely combines the principal
curvatures together makes it difficult to use as an edge detector. In other
words, it doesn't actually determine the gradient's maximum magnitude.
es
Edges are defined by the Canny edge detector as second derivative zero-
crossings in the direction of the largest first derivative. The Canny
operator performs their work in stages. The image is initially smoothed
ot

using a Gaussian convolution. Then, to highlight the areas of the image

with high spatial derivatives, a 2D first derivative operator is used to the
smoothed image. In the image of the gradient magnitude, edges give birth
un

to ridges. Non-maximal suppression is the method by which the algorithm

tracks along the top of these ridges and sets to zero all pixels that are not
actually on the ridge top in order to produce a thin line in the output. The
two thresholds T1 and T2, with T1 > T2, regulate the hysteresis of the
m

tracking process. Only at a position on a ridge higher than T1 may

tracking start. From there, tracking continues in both directions until the
height of the ridge drops below T2. Hysteresis like this prevents noisy
edges from becoming many edge fragments. The three parameters that
make up a Canny edge detector are (i) the width of the Gaussian kernel,
(ii) the higher threshold, and (iii) lower threshold used by the tracker
The detector's sensitivity to noise is decreased by widening the Gaussian
kernel, but some of the image's finer details are lost in the process. As the
Gaussian width is raised, the localization inaccuracy in the identified
edges also significantly rises. The Canny edge detector uses gaussian
smoothing for two reasons. In addition to noise reduction, it can be used to
first regulate how much detail emerges in the edge image.
For optimal results, the upper tracking threshold is often set relatively high
and the lower threshold value is set quite low. If the lower threshold is set
too high, noisy edges will fragment. If the higher threshold is set too low,
more false and undesired edge fragments will be produced.

96
7.8 EDGE LINKING Image Segmentation

One can threshold an edge-magnitude image and thin the resulting binary
image down to single pixel-wide closed, linked boundaries if the edges are
reasonably strong and the noise level is low. However, such an edge
picture will contain gaps that need to be filled under less-than-ideal
circumstances. Small gaps can be filled by looking for other end points in
a neighbourhood of 5 by 5 pixels or greater, centred on the end point, and
then filling in border pixels as necessary to join them. This, however, has
the potential to oversegment images with several edge points.

 Edge linking using Heuristic Search Algorithm

The heuristic search strategy to connect two edge points P and Q is
presented below if there is a gap between them:
A. Determine if P's neighbours qualify as the first step towards Q.
1. Name three P neighbours that are about in the same general area as Q.

.in
2. For each of the points from Step 1, compute the edge quality function
from P.
3. Decide on the option that improves edge quality from P to those
es
points.
4. Begin the subsequent iteration from the location established in Step 2.
ot

5. Continue in this manner until you reach point Q.

B. Establish and qualify the Path.
un

1. Evaluate the newly formed path's edge quality function against a

threshold.
m

2. Discard the freshly formed edge if it is insufficiently robust.

3. Use the resulting graph as the boundary if the newly formed edge is
robust enough.

7.9 HOUGH TRANSFORM

The straight-line y = mx + b isrepresented in polar coordinate as
ρ = x cos (θ) + y sin (θ)
where ρ, θ defines a vector from the origin to the nearest point on the
straight line. This vector will be perpendicular from the origin to the
nearest point to the line as shown in figure:

97
Digital Image Processing

Figure 2: Straight line

A point in the 2D space determined by the parameters and corresponds to
any line in the x, y plane. So, a single point in the ρ, θ space is the Hough
transform of a straight line in the x, y space. Each straight line that
traverses a certain point in the x, y plane, xi and yi maps to a point in the ρ,
θ space, and these points should meet the aforementioned equation with x1
and y1 as constants. Any point in the x, y plane corresponds to a certain
sinusoidal curve in the, space, and as a result, the locus of all such lines in
the x, y space is a sinusoid in parameter ρ, θ space.

.in
Suppose we have a set of edge points xi, yi that lie along a straight-line
having parameters ρ0, θ0. Each edge point plots to a sinusoidal curve in the
ρ, θ space, but these curves must intersect at a point ρ0, θ0since this is a
es
line, they all have common.

7.10 SUMMARY
ot

 The goal of image segmentation is to divide an image into discernible,

essentially homogeneous sections.
un

 Global segmentation deals with segmenting an entire image, whereas

local segmentation focuses on segmenting sub-images, which are tiny
windows on a larger image.
m

 There are three methods for performing image segmentation: the

region approach, boundary approach, and edge approach.

 An approach to picture segmentation known as "region growing"

involves looking at nearby pixels and adding them to a region class if
no edges are found. With a top-down strategy known as region
splitting, an entire image is split into portions that are more
homogeneous than the whole.

 By grouping the patterns into clusters or groups so that they are more
similar to one another than to patterns belonging to different clusters,
the clustering technique aims to access the links between patterns of
the data set.

 By using a thresholding process, a picture can be segmented, and the

threshold value can be chosen using the image's histogram as a tool.

98
 Edges are essentially discontinuities in the intensity of the image Image Segmentation
brought on by adjustments to the image structure.

 It is possible to identify the edges in a picture using the Robert,

Prewitt, Sobel, and Frei-Chen kernels.

7.11 LIST OF REFERENCES

1) Digital Image Processing, S Jayaraman, S Esakkirajan,
T Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009.
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008.
3) Scilab Textbook Companion for Digital Image Processing,
S. Jayaraman, S. Esakkirajan and T. Veerakumar, 2016
(https://scilab.in/textbook_companion/generate_book/125).

.in
7.12 UNIT END EXERCISES
1) What is content-based image retrieval?
2) What is an ‘edge’ in an image? On what mathematical operation are
es
the two basic approaches for edge detection based on?
3) What are the three stages of the Canny edge detector? Briefly explain
each phase.
ot

4) Compare the Canny edge detector with the Laplacian of Gaussian edge
detector
un

5) Describe various edge detection methods.

6) What is edge linking?
m

7) Describe Hough transform.

8) Discuss on image segmentation techniques.
9) Describe the region-based approach.



99
 8
IMAGE COMPRESSION
Unit Structure
8.0 Objectives
8.1 Introduction
8.2 Need for image compression
8.3 Redundancy in images
8.4 Image-compression scheme
8.5 Fundamentals of Information Theory
8.5.1 Entropy and mutual information

.in
8.5.2 Shannon’s source coding theorem
8.5.3 Rate-distortion theory
es
8.6 Run-length coding
8.6.1 1-DRun-length coding
ot

8.6.2 2-DRun-length coding

8.7 Shannon-Fano coding
un

8.8 Huffman Coding

8.9 Arithmetic Coding
m

8.10 Transform-based compression

8.11 Summary
8.12 List of References
8.13 Unit End Exercises

8.0 OBJECTIVES
To understand and get familiar with the following concepts:

 need for image compression along with their metrics and techniques
 lossless and lossy image compression
 spatial domain and frequency domain

100
8.1 INTRODUCTION Image Compression

The demand for digital information has rapidly increased as a result of the
development of multimedia technologies over the past few decades. The
widespread use of digital photographs is largely due to technological
advancements. Applications like medical and satellite imagery frequently
use still photos. Digital photos contain a tremendous quantity of data. As
digital images find more applications, image data size reduction for both
storage and transmission are becoming more and more crucial. A mapping
from a higher dimensional space to a lower dimensional space is what
image compression is. Many multimedia applications, including image
storage and transmission, heavily rely on image compression.The
fundamental objective of image compression is to represent an image with
the fewest possible bits while maintaining acceptable image quality. All
image-compression methods aim to reduce the amount of data as much as
possible while removing statistical redundancy and taking advantage of
perceptual irrelevancy.

8.2 NEED FOR IMAGE COMPRESSION

.in
The amount of information that computers handle has increased
tremendously over the past few decades thanks to advancements in
es
Internet, teleconferencing, multimedia, and high-definition television
technology. Consequently, a significant issue with multimedia systems is
the storage and transmission of the digital image component. To exhibit
photographs with an acceptable level of quality, a tremendous amount of
ot

data is needed. Large amounts of storage space and transmission

bandwidth are needed for high-quality image data, which the existing
un

technology cannot provide both technically and economically.

Compressing the data to save on storage space and transmission time is
one method that could work for this issue.
m

For instance, the amount of storage space needed to hold a 1600 x 1200
colour photograph is
1200 ×1600 × 8 × 3 = 46,080,000 bits
= 5,760,000 bytes
= 5.76 Mbytes
1.44 Mb is the maximum amount of space on a floppy disc. The maximum
space available if we have three floppies is 1.44 x 4 = 5.76 Mbytes. In
other words, a 1600x1200 RGB image requires at least four floppies to
store.
Every year, the amount of data carried across the Internet doubles, and a
sizable chunk of that data is made up of photographs. Any given device
will experience significant cost savings and become more accessible to use
if its bandwidth requirements are reduced. photos can be represented more
compactly through the use of image compression, allowing for speedier
transmission and storage of photos.
101
Digital Image Processing 8.3 REDUNDANCY IN IMAGES
Because images are typically highly coherent and contain redundant
information, image compression is achievable. Redundancy and
irrelevancy reduction help in compression. Duplication is referred to as
redundancy, while irrelevancy is the portion of an image's information that
the human visual system will not pick up on. Figure 1 depicts the
classification of redundancy

.in
Figure 1: Classification of redundancy

 Statistical redundancy
es
As previously mentioned, there are two kinds of statistical redundancy:
interpixel redundancy and coding redundancy.
When adjacent pixels in an image are correlated, interpixel redundancy
ot

results. The implication is that the pixels next to each other are not
statistically independent. Interpixel redundancy is the term used to
describe the interpixel correlation.
un

Information representation is connected to redundancy in coding. Codes

are used to represent the information. Examples of codes include the
arithmetic codes and the Huffman code. The efficiency of different codes
m

may vary. To properly compress the image, codes must be effective.

 Spatial redundancy
In an image, the statistical relationship between adjacent pixels is
represented by spatial redundancy. The term "spatial redundancy" refers to
the relationship between adjacent pixels in an image. Each pixel in an
image need not be individually represented. Instead, a pixel's neighbours
can be used to forecast it. The fundamental idea behind differential
coding, which is commonly used in picture and video reduction, is to
eliminate spatial redundancy by prediction.

 Temporal redundancy
The statistical relationship between pixels from subsequent frames in a
video clip is known as temporal redundancy.Interframe redundancy is
another name for the temporal redundancy. To lessen temporal
redundancy, motion compensated predictive coding is used. Effective

102
video compression is achieved by removing a significant amount of Image Compression
temporal redundancy.

 Psychovisual redundancy
The traits of the human visual system (HVS) are connected to
psychovisual redundancy. Visual information is not perceived similarly in
the HVS. It's possible that certain information is more crucial than other
information. Perception won't be impacted if less data is used to represent
less significant visual information. This suggests that visual information is
redundant in a psychovisual sense. Effective compression results from
removing the psychovisual redundant information.

8.4 IMAGE-COMPRESSION SCHEME

Figure 2 depicts the general block diagram of an image-compression
technique. While the channel encoder and decoder pair is often referred to
as a channel codec module, the source encoder and decoder pair is
frequently referred to as a source codec module. Over the years, Shannon's

.in
well-known separation theorem has been used to support the two coding
modules' independent designs.
es
ot
un

Figure 2: Image compression scheme

 Source Coding: The effective translation of the source data—the input

m

image data—into a series of bits is the aim of source coding. By

reducing entropy, the source coder lowers the average number of bits
needed to represent the image. In essence, source decoding is an
inverse process.

 Channel: The mathematical representation of the medium used for

communication is the "channel."

 Channel Coding: The source encoder compresses the output and adds
controlled redundancy using the channel encoder. The channel
encoder's job is to shield the communication system from channel
noise and other transmission mistakes. To recreate the compressed
bits, the channel decoder makes use of the redundant bits in the bit
sequence.

103
Digital Image Processing 8.5 FUNDAMENTALS OF INFORMATION THEORY
Both lossless and lossy compression are mathematically based on
information theory. Entropy, average information, and rate-distortion
principles are covered in this section.

8.5.1 Entropy and mutual information

For a random variable X generated by a discrete memoryless source from
a finite alphabet set X = {x1, x2, ………… xN} with px being the
probability density function, Shannon defined a measure of the
information content of X, called entropy, as

The entropy H(X) is the lower bound of the bit rate that can represent the
source without distortion. The conditional entropy of a random variable Y
with a finite alphabet set y is defined as

.in
es
where pxy(x, y) is the joint probability density function of X and Y and px|
y(x, y) is the conditional probability density function of X given Y. The
mutual information I (X; Y) between two random variables X and Y is a
measurement of the reduction in uncertainty due to conditioning of Y,
ot

which is defined as
un

The above equation implies that mutual information is symmetric.

m

8.5.2 Shannon’s source coding theorem

The aim of source coding is to effectivelyexpress a random experiment’s
sequence of outcomes. According to Shannon’s source-coding theorem
“Let X be the ensemble of letters from a discrete memoryless source with
finite entropy H (X). Blocks of J symbols from the source are encoded into
code words of length N from a binary alphabet. For any ε > 0, the
probability Pe of a block-decoding failure can be made arbitrarily small if

and J is sufficiently large. Conversely, if R ≤ H (X) − ε then Pe becomes

arbitrarily close to 1 as J is made sufficiently large. The source-coding
theorem demonstrates that the source entropy H (X) is the lower bound of
the average number of bits/symbols required to encode the output of a
discrete memoryless source without distortion.

104
8.5.3 Rate-distortion theory Image Compression

The task of reducing redundancy from a source while adhering to a

predetermined fidelity requirement is the focus of the rate distortion
theory. The maximum distortion of any source-coding scheme that
satisfies the rate constraint R is the definition of the rate-distortion
function represented by R(D) of the source X. The rate-distortion function
for a discrete memoryless source using the mean square error criterionis
given by

where PY|X implies source-coding rule and MSE the MSE of the encoding
rule.

8.6 RUN-LENGTH CODING

Long sequences of the same symbol can be effectively encoded using run-

.in
length coding (RLC). By encoding the total amount of symbols in a run,
run-length coding takes advantage of the spatial redundancy. The word
"run" denotes the repetition of a symbol, while the word "run-length"
es
denotes the quantity of repeated symbols. A series of numbers is converted
into a series of symbol pairs (run, value) using run-length coding. For this
type of compression, images with sizable regions of consistent shade make
suitable candidates. The bitmap file format used by Windows employs
ot

run-length coding.There are two types of run-length coding: (i) 1D run-

length coding (ii) 2D run-length coding. As opposed to 2D RLC, which
uses both horizontal and vertical correlation between pixels, 1D RLC only
un

uses the horizontal correlation between pixels on the same scan line.

8.6.1 1-DRun-length coding

m

Each scan line in 1D run-length coding is separately encoded. A series of

alternating, independent white runs and black runs can be thought of as
each scan line. The first run in each scan line is taken into consideration to
be a white run as per an agreement between the encoder and decoder. The
run-length of the initial white run is set to zero if the first actual pixel is
black. There is end-of-line (EOL) at the conclusion of each scan line.
When an EOL codeword is encountered, the decoder is aware that the scan
line has ended.
Example: Consider figure 3 for an example of a binary image and its
associated binary representation. Using run-length coding, the image must
be encoded.Run-length coding transmits two values: the first value
represents the quantity of times a specific symbol has appeared, and the
second value is the sign itself. It is clear from Fig. 3 that each row is
successively scanned, and the associated run-length code is delivered. An
illustration of run-length coding is provided for the fourth row of the input
image. The bit stream for the fourth row in the example is 4, 0, 10, 1, 4, 0.

105
Digital Image Processing

Figure 3: Example ofrun-length coding

8.6.2 2-DRun-length coding

The correlation between pixels in a scanline is used in 1D run-length
coding. 2D run-length coding was created to take use of the correlation
between pixels in nearby scan lines to increase coding effectiveness.

8.7 SHANNON-FANO CODING

.in
A top-down binary tree approach, Shannon-Fano coding uses binary trees.
The Shannon-Fano coding algorithm is provided below:
es
1. The collection of source symbols is arranged in reverse probability
order.
2. The interpretation of each symbol is that of a tree's root.
ot

3. The list is split into two groups with roughly equal overall probabilities
for each category.
4. A 0 is added to the first group's code phrase.
un

5. The number one is added to the second group's code phrase.

6. For each group, steps (3) to (4) are repeated until there is only one node
m

in each subset.
Example: Construct the Shanon–Fano code for the word MUMMY
Solution: The given word is MUMMY. The number of alphabets in the
word MUMMY (N) is five. Step 1 Determining the probability of
occurance of each character Probability of occurrence of each character in
MUMMY is given by

Step 2 Determining the number of steps Now, the number of steps to be

involved in this problem is number of steps = number of symbols − 1
=3−1=2
Step 3 Computation of the code word for each symbol the steps involved
in the computation of the code words is shown in the table given below

106
 Image Compression

Step 4 Determination of entropy

After determining the probability and the code for each character, the next
step is to compute the entropy. The entropy is given by

.in
es
Step 5 Computation of average length
ot

The average length is given by

un
m

Step 6 Computation of efficiency

The efficiency of the code η is given by

8.8 HUFFMAN CODING

In 1952, D A Huffman created the Huffman coding system. The algorithm
was created by David Huffman in 1950 while he was a student at MIT
taking an information theory course. One symbol is mapped to one code
word in Huffman codes, which are the best codes. It is assumed in
Huffman coding that each pixel's intensity has a specific probability of
occurrence, and that this probability is spatially invariant. Each intensity
value is given a binary code through Huffman coding, with shorter codes
107
Digital Image Processing going to intensities with higher probabilities. The Huffman codes table can
be fixed in both the encoder and decoder if the probabilities can be
calculated a priori. If not, the compressed image data must also be
transmitted with the coding table to the decoder. Entropy, Average length,
Efficiency, Variance are some of the parameters involved in Huffman
coding.
In the first step of Huffman's method, the probabilities of the symbols
under consideration are ranked, after which the lowest probability symbols
are combined into a single symbol and substituted in the following source
reduction. This binary coding technique is seen in figure 4. A hypothetical
collection of source symbols and their probabilities are arranged from top
to bottom in decreasing order of probability values at the far left. The
lowest two probabilities, 0.06 and 0.04, are combined to create a
"compound symbol" with probability 0.1 to create the initial source
reduction.In order to rank the probabilities of the reduced source from
most to least likely, this compound symbol and its corresponding
probability are placed in the first source reduction column. Then, this
process is continued until the reduced source (at the far right) has two

.in
symbols. es
ot
un

Figure 4: Huffman source reductions

m

The coding of each reduced source, proceeding backwards from the

smallest to the largest, is the second stage in Huffman's method. The
symbols 0 and 1 are, of course, the smallest length binary code for a two-
symbol source. These symbols are given to the two symbols on the right,
as seen in Fig. 5 (The assignment is arbitrary; it would also be effective to
reverse the order of the 0 and 1). The 0 used to code the reduced source
symbol with probability 0.6 is now assigned to each of these symbols, and
a 0 and 1 are arbitrarily appended to each to distinguish them from one
another. This is because the reduced source symbol with probability 0.6
was created by combining two symbols in the reduced source to its
left.This operation is then repeated for each reduced source until the
original source is reached. The final code appears at the far left in Fig. 5.
The average length of this code is

and the entropy of the source is 2.14 bits/symbol.

108
 Image Compression

Figure 5: Huffman code assignment procedure

When symbols are coded one at a time, Huffman's method finds the best
possible code for a collection of symbols and probabilities. Once the code
has been written, it may be easily decoded without error using a simple
search table. The code is a block code that can be instantly and exclusively
decoded. Because each source symbol is translated into a predetermined
order of code symbols, it is known as a block code. Because each code
word in a string of code symbols may be deciphered without referring to

.in
subsequent symbols, it is instantaneous.It can only be decoded in one
method for each given string of symbols, making it uniquely decodable.
As a result, any string of Huffman-encoded symbols may be decoded by
looking at each symbol in the string from left to right.For the binary code
es
of Fig. 5, a left-to-right scan of the encoded string 010100111100 reveals
that the first valid code word is 01010, which is the code for symbol a3.
The next valid code is 011, which corresponds to symbol a1. Continuing
ot

in this manner reveals the completely decoded message to be a3 a1 a2 a2 a6.

8.9 ARITHMETIC CODING

un

In arithmetic coding, which dates back to Elias's work (Abramson [1963]),

source symbols and code words do not always correlate exactly. Instead, a
single arithmetic code word is given to an entire series of source symbols
m

(or a message). By itself, the code word designates a range of real values
between 0 and 1. The interval used to represent the message's increasing
number of symbols gets smaller as well as the quantity of information
units (let's say, bits) needed to do so.The size of the gap is decreased for
each symbol in the message in line with the likelihood that it will appear.
The technique achieves (but only in theory) the bound set by Shannon's
first theorem since it does not require, as does Huffman's approach, that
each source symbol translate into an integral number of code symbols (i.e.,
that the symbols be coded one at a time).
The procedure of fundamental arithmetic coding is shown in Figure 6.
Here, a1a2a3a3a4, a five-symbol sequence or message, is coded from a four-
symbol source. The message is first believed to take up the entire half-
open interval [0, 1) before the coding process even begins. This interval is
originally partitioned into four sections depending on the probabilities of
each source symbol, as shown in Table 1. For instance, the symbol a1 is
related to the subinterval [0, 0.2).The message interval is initially reduced
to [0, 0.2) because it is the first symbol of the message that is being coded.
109
Digital Image Processing As a result, in Fig. 6, the range [0, 0.2) is expanded to fill the entire height
of the figure, with the values of the restricted range designating its end
points. The next message symbol is then added to the narrower range, and
the procedure is repeated according to the probability of the origin source
symbol.In this manner, symbol a2 narrows the subinterval to [0.04, 0.08),
a3 further narrows it to [0.056, 0.072), and so on. The range is reduced to
[0.06752, 0.0688) by the last message symbol, which must be reserved as
a unique end-of-message signal. Of course, the message can be
represented by any integer falling inside this subinterval, such 0.068. The
five symbols of the arithmetically coded message shown in Fig. are
represented by three decimal digits. In comparison to the source's entropy,
which is 0.58 decimal digits per source symbol, this equates to 0.6 decimal
digits per source symbol.The resulting arithmetic code approaches the
bound defined by Shannon's first theorem as the length of the sequence
being coded grows. The use of finite precision arithmetic and the addition
of the end-of-message indicator, which is required to distinguish one
message from another, are the two variables that cause coding
performance to fall short of the bound in practise. The latter issue is

.in
addressed by scaling and rounding strategies introduced in practical
arithmetic coding systems. The scaling approach divides each subinterval
in accordance with the symbol probabilities after renormalizing it to the
[0, 1) range. The rounding technique ensures that the coding subintervals
es
are faithfully represented despite the truncations brought on by finite
precision arithmetic.
Table 1:Arithmetic coding example
ot
un
m

Figure 6:Arithmetic coding procedure

110
8.10 TRANSFORM-BASED COMPRESSION Image Compression

The method most frequently employed in image-coding applications is

image-transform coding. A set of linear transform coefficients, which are
typically scalar-quantised and entropycoded for transmission, are created
by linearly transforming an image. Transform coding, then, is a
mathematical process that divides a large number of highly correlated
pixels into a smaller set of uncorrelated coefficients. The foundation of
transform coding is the existence of a specific degree of connection
between pixels in an image. When compressing images, transformation is
a very helpful technique. It is used to convert time-domain image data to
frequency-domain data. The spatial redundancy in the time domain can be
reduced by converting the data into frequency domain.The advantage of
using transformation is that the energy of the transformed data is mainly
condensed in the lowfrequency region, and is represented by a few
transform coefficients. Thus, most of these coefficients can be discarded
without significantly affecting the reconstructed image quality. In a
transform coder, the discrete data signal is segmented into non-

.in
overlapping blocks, and each block is expressed as a weighted sum of
discrete basis functions. The purpose of transform coding is to decompose
the correlated signal samples into a set of uncorrelated transform
coefficients, such that the energy is concentrated into as few coefficients
es
as possible. The block diagram of transform-based image coding scheme
is shown in Fig. 7.
ot
un
m

Figure 7:Transform-based image-coding scheme

Transformation: The image's correlation qualities are altered as a result of
the transformation's "reorganization" of the grey values. The process of
transformation condenses the image data into a tiny set of coefficients.
The transform should have the following characteristics for effective
compression:

 Decorrelation: The transform should generate less correlated or

uncorrelated transform coefficients to achieve high compression ratio.
 Linearity: Linearity principle allows one-to-one mapping between
pixel values and transform coefficients.
 Orthogonality: Orthogonal transforms have the feature of eliminating
redundancy in the transformed image
111
Digital Image Processing Quantization: It is the process of limiting a quantity's range of possible
values in order to minimize the number of bits required to represent it. By
itself, the picture transformations don't lead to any compression. Instead,
they represent the image in a distinct domain where the image data is
divided into parts with varied degrees of significance. In essence,
quantization is an irreversible process.
Entropy encoder: The reduction of the number of bits needed to represent
each symbol at the quantizer output is the goal of an entropy encoder.
Techniques for entropy-coding that are frequently employed include run-
length coding, arithmetic coding, and Huffman coding. A lossless coding
system is essentially what entropy coding is.
Essentially, the encoder's output is a bit stream that is sent over the
channel to the decoder. If one is just focused on source code, the channel
is typically taken to be lossless. If one is interested in joint source and
channel coding, one should be aware that the mistake happens when the
bitstream traverses the channel. In essence, the decoder reverses the
encoder's process to produce the reconstructed image.

.in
8.11 SUMMARY
 Compact representation is compression. The representation of a
es
picture with the fewest possible bits is known as image compression.

 Redundancies in an image are reduced to accomplish compression.

ot

Spatial redundancy, temporal redundancy, and psychovisual

redundancy are three different types of redundancies in an image.
un

 Image compression can be divided into two categories: lossless

compression and lossy compression. The rebuilt image closely
resembles the original image when using lossless compression.
Information is lost during lossy compression.
m

 These methods are frequently used in lossless picture compression:

Run-length coding, Huffman coding, arithmetic coding, Shannon-Fano
coding, and dictionary-based methods like LZW are only a few
examples.

 Quantization is basically approximation. The process of quantization is

non-linear and irreversible. Scalar quantization and vector quantization
are the two basic categories under which quantization may be
classified.

8.12 LIST OF REFERENCES

1) Digital Image Processing, S Jayaraman, S Esakkirajan,
T Veerakumar,Tata McGraw-Hill Education Pvt. Ltd., 2009.
2) Digital Image Processing 3rd Edition, Rafael C Gonzalez, Richard E
Woods, Pearson, 2008.
112
3) Scilab Textbook Companion for Digital Image Processing, Image Compression
S. Jayaraman, S. Esakkirajan and T. Veerakumar, 2016
(https://scilab.in/textbook_companion/generate_book/125).

8.13 UNIT END EXERCISES

1) Explain the need for image compression.
2) Describe Redundancy in images.
3) Explain Image-compression scheme.
4) What is Entropy and mutual information?
5) Explain Shannon’s source coding theorem.
6) Describe Rate-distortion theory.
7) What is Run-length coding? Explain along with its types.
8) Write a note on Shannon-Fano coding.

.in
9) Describe Huffman Coding.
10) Explain Arithmetic Coding.
es
11) Explain Transform-based compression.


ot


un
m

113