UNIT 4

IMAGE FILTERING
OPERATIONS IN SPATIAL
DOMAIN

Structure Page No.

4.1 Introduction 129
Objectives
4.2 Spatial Filtering 130
4.3 Image Smoothing 131
Linear Filters
Non-Linear Filters
4.4 Image Sharpening 140
First Order Filters
Second Order Filters
4.5 Histogram Processing 111
4.6 Histogram Equalization 115
4.7 Histogram Specification
4.8 Summary 146
4.9 Solutions/Answers 147

4.1 INTRODUCTION
In the previous units of this course, we had seen how image transformation
carried out on a specific pixel depends on the gray value of that pixel. In this
unit, we shall consider more powerful transformation where the gray values of
the neighbouring pixels also play a useful part. We have discussed point
processing operations of the type s  T ( r ), where r is the grey level at a single
pixel in the input and s is the new value of that pixel. The capabilities of point
operations are limited as the relation between a pixel and its neighbours is not
exploited.
In this unit, we will be considering a neighbourhood of pixels from the input
image. This is termed as spatial filtering. We introduce image enhancement
with the help of image sharperning (spatial high pass) and image smoothing
(low pass filters). Spatial filtering as generally used in preprocessing
operations for noise removal or for edge detection/enhancement applications.

Subsequently, In this unit, we introduce the concept of histogram and provide

an overview of image enhancement using histogram equalization and
histogram specification. The histogram of an image represents the relative 129
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
frequency of occurrence of the various gray levels in the image. It provides
useful image statistics that help us in analyzing the image. Histograms are the
basis for many spatial domain processing techniques. Histogram-modeling
techniques modify an image so that its histogram has a desired shape.
Histogram manipulation is used for image enhancement.
Now, we shall list the objectives of this unit. After going through the unit,
please read this list again and make sure that you have achieved the
objectives.

Objectives
After studying this unit, you should be able to
 apply spatial filtering
 perform linear filtering in spatial domain
 Use differentiate low pass & high pass filtering in spatial domain
 Use differentiate non-linear filtering and linear filtering
 Perform median, min & max filters in spatial domain

 define histogram;
 generate histogram form given image;
 perform histogram equalization;
 perform histogram specification.

4.2 SPATIAL FILTERING

Filtering is a process that removes some unwanted components or small
details in an image. In digital image processing, filter is basically a subimage
and is known by various names such as mask, kernel, template or window.
Filters can be of two types: Spatial filters and frequency domain filters.
We have discussed point processing operations of the type s  T ( r ) , where r
is the grey level at a single pixel in the input image and s is the new value of
that pixel. The capabilities of point operations are limited as the relation
between a pixel and its neighbors is not exploited. In this section, we will be
considering a neighborhood of pixels from the input image. Spatial filtering is
one of the main tools used in variety of applications such as noise removal,
bridging the gaps in object boundaries, sharpening of edges etc.

(a) Original Image (b) Image after filtering

Fig. 1: Linear Filtering

The idea of spatial filtering is to move a ‘mask’, a square of odd size (such as
3x3, 5x5 etc.) over the whole image pixel by pixel. At each pixel, the
corresponding value of the mask and the image are multiplied and added up to
replace the original grey value of the pixel. This process is known as
“convolution”. 130
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

By this process, we create a new image where grey level values of the pixels
are calculated from the values under the mask. The values under the mask
are modified by a function called ‘filter’. If this filter function is a linear function
of all grey level values in the mask, then filter is called a ‘linear filter’, else it is
called ‘non linear filter’.
Spatial filters can be of various types as shown in Fig. 2.

Spatial Filters

Smoothing [Low Pass] Sharpening [High Pass]

Linear Filters Non-Linear Filters First Derivative Second

(Average Filters) (Median and Minmax) Filters Derivative

Fig. 2: Types of Spatial Filters

In this unit, we will discuss two major types of spatial filters.

i) Smoothing: This type of filter is used to blur (smooth) the image,

therefore called smoothing filter also. These filters are also used for noise
reduction. Noise reduction can be done by blurring with a linear filter, (in
which operation performed on the image pixels is linear) or a non-linear
filter.

ii) Sharpening: This filter is used to highlight transitions in intensity. These

are based on first and second derivatives.

The effect of these filters is shown in Fig. 3.

Fig. 3: Smoothing and Sharpening Filters

In the following section, we discuss image smoothing filters in detail.

131
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

4.3 IMAGE SMOOTHING

We discussed spatial filters in previous section. Spatial filter consists of a
typically small rectangle called neighbourhood and a predefined operation
which is performed on the pixels of the image encompassed by the
neighbourhood. Such filtering creates a new pixel with coordinates equal to
the coordinates of the center of the neighbourhood, and whose value is the
outcome of the filtering operation. Thus, a filtered image is generated as the
center of the filter visits each pixel in the input image. If the operation
performed on the image pixels is linear, then the filter is called a linear spatial
filter. Otherwise, the filter is nonlinear. Here, we shall discuss first linear filters
and then some simple nonlinear filters.

Let us discuss linear spatial filters to start with.

4.3.1 Linear Filters

Linear filtering is a spatial domain process where a filter (mask/ kernel/
template) with some integer coefficient values is applied to input image to
generate the filtered/ output image. Generally, filter size is either 3  3 or
5  5, 7  7 or 21 21 (odd sizes) and filter is centered at a coordinate ( x , y)
called ‘Hot Spot’ as shown in Fig. 4.

Fig. 4

Linear filtering of an image f ( x , y) of size M  N and filter mask of size m  m

is given by

g ( x , y)   f (x  i, y  j) w (i, j) where 0  x  M  1 and 0  y  N  1.

( i , j)R n
(1)

The right hand side of Eqn. (1) denotes the set of coordinates covered by filter.
For a 3 3 filter, with coefficients w (i, j) the output image at coordinate ( x , y) is
calculated as

1 1
g ( x , y)    f ( x  i, y  j) w (i, j) (2)
i  1 j 1

g ( x , y)  f ( x , y) w (0, 0)  f ( x , y  1) w (0,1)  f ( x , y  1) w (0,1)  f ( x  1, y  1)

w (1,1)  f ( x  1, y) w ( 1,0)  f ( x  1, y  1) w (1,1) 
f ( x  1, y  1) w (1,1)  f ( x  1, y) w (1, 0)  f ( x  1, y  1) w (1,1)

We carry out the following steps for linear filtering:

Step 1: Position the mask over the current pixel such that hotspot w(0,0)
coincides with current pixel.

Step 2: Form all products of filter elements with the corresponding elements in
the neighbourhood. 132
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Step 3: Add up all the products and store it at current position in the output
image.

Step 4: Divide it by the scaling constant (sum of all coefficients of the

mask).This must be repeated for every pixel in the image.

Fig 5: Linear filtering

Let us apply these steps in the following example.

Example 1: Apply given 3 3 mask w on the following image f ( x , y).

5 1 2 6 7
4 4 7 5 8 1 1 1
1
f ( x , y)  2 6 20 6 7 , w (i, j)  * 1 1 1
9
3 1 2 4 5 1 1 1
10 2 1 2 3

Solution: First we consider top left 3 3 neighborhood in image f ( x , y) as

shown in Fig. 6 (a) with hot spot coinciding with grey level 4 in f ( x , y) and
calculate output for that pixel position by using equation

1
Output  [5  1  1  1  2  1  4  1  4  1  7  1  2  1  6  1  20  1]
9
49
 5
9

5 1 2 6 7 5 1 2 6 7
4 4 7 5 8 4 4 7 5 8
2 6 20 6 7 2 6 20 6 7
3 1 2 4 5 3 1 2 4 5
10 2 1 1 3 10 2 1 1 3

(a) (b)
Fig. 6

Shift the mask one pixel towards right such that grey level 7 coincides with hot
spot (Fig. 6 (b)). Calculate output for that pixel position and replace in output
image. Shift it right again and locate the mask on pixel with grey value 5, and
repeat the process.

1
Output  [1 1  2  1  6  1  4  1  7  1  5  1  6  1  20  1  6  1]
9 133
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
57
 6
9
Now go back to the first position of the mask and shift it down so that it is now
centered on pixel with grey value 6 (just below 4), and repeat the process.
1
Output  [2  1  6  1  7  1  7  1  5  1  8  1  20  1  6  1  7  1]
9
68
 8
9

Continue doing it over all the possible pixels in the given image.

1 49
Output  [4  4  7  2  6  20  3  1  2]  5
9 9

1 55
Output  [4  7  5  6  20  6  1  2  4]  6
9 9

1 64
Output  [7  5  8  20  6  7  2  4  5]  7
9 9

1 38
Output  [2  6  20  3  1  2  10  2  1]  4
9 9

1 44
Output  [6  20  6  1  2  4  2  1  2]  5
9 9

1 59
Output  [20  6  7  2  4  5  1  2  3]  6
9 9

So, by shifting the mask over the whole image a new image is generated.
These values are rounded off to nearest integer values and put in the output
image as given below at corresponding pixel locations.

* * * * *
* 5 6 8 *
g ( x , y)  * 5 6 7 *
* 4 5 6 *
* * * * *

***
The filtering effect is visible in the new image where the noisy values 10, 20 in
original image have been replaced by smoothed values.

Remark: There is a small problem in applying a filter at the edge of the image,
where the mask partly falls outside the image. In Example 1, input image size
 5  5, and the mask size  3  3. Therefore, the output image size  3  3.
The size of output image is smaller than input image size as the mask does
not overlap fully in the 1st row, 1st column last column and last row as shown
in image g ( x , y).

However, it is not a serious problem as in practical situations the image sizes

are much larger such as 200  200 and loss of few values at the boundary of
the image is not even visible to us.
134
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Now, try the following exercises.

E1) For a 3 3 mask, explain the mechanics of linear filtering with necessary
equations.

E2) List various applications of linear filtering.

Now, we discuss one of the commonly used linear filter known as a mean
filter.

The aim of image smoothing is to diminish the effect of camera noise, spurious
pixel values, missing pixel values etc. This is done by a neighborhood
averaging filter. Each pixel in the smoothened image g ( x , y) is obtained from
the average pixel value in the neighborhood of ( x , y) in input image. Such a
mask is also known as a Mean filter.
1
g ( x , y)   f ( m, n ) w ( m , n )
M ( m ,n )Sxy
(3)

Where Sxy  neighbourhood, M  Number of pixels in Sxy and w is the mask

for averaging.

Two masks averaging and weighted averaging masks are shown in Fig. 7. The
output is average of pixels contained in the neighbourhood of filter mask.

Therefore, smoothing filters are also called averaging filters or low pass filters.
Fig. 7 (a) and Fig. 7 (b) shows 3 3 smoothing filter masks. Spatial filter, where
all coefficients are equal, is called box filter. However, to give maximum
importance to pixel at the centre of the mask, it is given maximum weight, and
the weights of the neighnouring pixels are progressively reduced.

(a) Box Filter ( 3 3 Averaging Mask)

(b) 3 3 Weighted Averaging Mask

Fig. 7
In this case centre pixel is given the most importance and other pixels are
inversely weighted as a function of their distance from the centre of the mask.
135
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
This filter reduces blurring in the smoothing process as the centre pixel is
weighted highest.
Smoothing filter is used for the following purposes:

i) It reduces ‘sharp’ transitions in grey levels.

ii) It reduces noise.
iii) It blurs edges. This is a side effect.
iv) It helps in smoothing false colours.
v) It reduces ‘irrelevant’ details in an image.

‘Blurring’ (a side effect of smoothing filter) is used in preprocessing steps for

following applications:

i) Removal of small details from an image prior to object extraction. Intensity

of smaller objects blends with the background and larger object become
‘blob like’ and easy to detect.

ii) Bridging small gaps in lines or curves of the boundaries of objects and
text. Small gaps in boundary can lead to enormous results in object
extraction. Smoothing fills up all smaller gaps and helps in producing
correct result.

(a) noisy image (b) output of 3 3 averaging filter

Fig. 8
See this example.

Example 2: Apply averaging filter to input image f ( x , y) as given in Fig. 10 to

produce output image g ( x , y) in Fig. 10.

5 1 2 6 7
5 6 8
4 4 7 5 8
5 6 7
2 6 20 6 7
4 5 6
3 1 2 4 5
g ( x , y)
10 2 1 2 3
f ( x , y)
Fig. 9

Solution: Smoothing filter replaces every pixel of the input image by the
average of the grey levels in the neighborhood. It reduces sharp transitions in
grey levels. Note that pixel value 20 in f ( x , y) has been changed to 6 in the
output image.

Sharp transitions can be due to 136

UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

• Random noise in the image

• Edge of objects in the image

Smoothing filter reduces noise (desirable) and also blurs edges (undesirable).
In the image f ( x , y), pixel value 20(high value as compared to neighborhood)
becomes 6, pixel value 1 (low value as compared to neighborhood) becomes
4, whereas pixel value 7(similar value as compared to neighborhood)
becomes 6.

Thus, noisy pixel or edges (20, 1: arbitrarily high/ low values) are reduced to
average value, closer to neighborhood values. Note that smoothing effect will
be more if instead of 3x3 mask, a large sized mask such as 7  7 or 11 11
mask is used. The size of the mask will depend on amount of noise and size of
the image.
***
Try these exercises.

E3) With various spatial masks and equations, explain spatial averaging
technique for enhancement.

E4) Discuss image smoothing filter in the spatial domain.

4.3.2 Nonlinear Filters

Smoothing linear filters have an important disadvantage, image structures
such as individual points, edges and lines are blurred. Overall quality of the
image reduces in some applications. These side effects are not tolerated
which limits the usage of linear filters. Order statistics filters are non-linear
filters whose response is based on ordering (ranking) the pixels contained in
the image area encountered by the filter. Like all other spatial filters, non linear
filters compute the result at some position ( x , y) from the pixels inside the
moving region S of the original image. These filters are called non-linear
because source pixels are processed by some non-linear function.

Here, we discuss some simple non linear filters such as median filter and
Min/Max filters.

Median Filters

Median filters are edge preserving smoothing filters, where the level is set to
the median of pixel values in the neighborhood of that pixel. It is impossible to
design a filter that removes only noise and retains all the important image
structures intact, because no filter can discriminate which image content is
important to the viewer and which is not. Median filter replaces every image
pixel by median of the pixels in the corresponding filter region Sxy.

g(x, y)  median{f (x  i, y  j) (i, j)  Sxy} (5)

Median of 2k  1 pixel values is defined as median

(p 0 , p1 , p 2 ,..., p k , p k 1 ,....., p 2 k ) ˆ p k . Median is the centre value p k if the
sequence ( p 0 , p1 ,..., p 2 k ) is sorted.

137
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Fig 10: Median Filtering

If the number of elements are even (2k) the median of the sorted sequence
(p 0 , p1,...., p k 1 , p k ,...., p 2 k 1 ) is defined as arithmetic of two middle values
p k 1  p k
Thus, median ( p 0 , p1 ,...., p k 1 , p k ,..., p 2 k 1 ) 
2
Fig. 10 shows median filtering process and generating output image. Median
filter is very popular because for impulse noise (salt and pepper noise,
randomly placed white and black dots), it provides excellent noise reduction
capabilities with considerably less blurring than linear smoothing filter of same
size. Median filters force the points with distinct levels to be more like their
neighbors. It eliminates isolated clusters of pixels that are dark or light with
respect to their neighbors and whose area is less than k 2 / 2 (one half of filter
area).

Advantages of Median Filters

• It has excellent noise reduction capability.

• It preserves edges.
• It introduces less blurring than smoothing linear filters.
• It is very effective in the presence of impulse noise.

Disadvantages

• It performs poorly if number of noise pixels in S

• is greater than half the number of pixels in the windowxy
• It performs poorly in the presence of gaussian noise.

Steps to Implement Median Filter as shown in Fig. 10.

Step 1: Sort the values of the pixel in question and its neighbours.
Step 2: Determine their median value.
Step 3: Assign median value to that pixel.

Example 3: Compute the median value of the marked pixels show below
using a 3 3 mask
18 22 33 25 32 24
34 128 24 172 26 33
 
22 19 32 31 28 26

Solution: i) Median (18, 22, 33, 34, 128, 24, 22, 19, 32)
 Median (18, 19, 22, 22, 24, 32, 33, 34, 128)  24
ii) Median (22, 33, 25, 128, 24, 172,19, 32, 31)
 median (19, 22, 24, 25, 31, 32, 33,128,172)  31
iii) Median (33, 25, 32, 34,172, 26, 32, 31, 28)
138
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
 median (24, 25, 26, 28, 31, 32, 32, 33,172)  31
iv) Median (25, 32, 24, 172, 26, 23, 31, 28, 26)
 median (23, 24, 25, 26, 26, 28, 31, 32,172)  26
***
Minimum and Maximum Filter

Max and min filters are defined as:

g(x, y)  min{f (x  i, y  j) (i, j)  Sxy } (6)
g(x, y)  max{f (x  i, y  j) (i, j)  Sxy }, (7)
where S xy denotes the filter region, usually a size of 3 3 pixels. Min filter
removes salt noise (white dots with large grey level values) because any large
grey level with in a 3 3 filter region is replaced by one of its surrounding pixels
with smallest value. As a side effect, min filter introduces dark structures in the
image. The reverse effect is expected from a max filter. It removes pepper
noise (black dots with small grey level values) because any black dot within
3 3 filter region is replaced by one of its surrounding pixels with the largest
value. White dots/bright structures are widened as a side effect and black dots
(pepper noise) will disappear. Fig. 12 shows the process of min/ max filter.

Fig. 11: Min, max filtering

Example 4: For the image segment f ( x , y) given below.

Apply

a) Smoothing filter
b) Weighted average filter
c) Median filter
d) Min filter
e) Max filter

Solution: (a) Smoothing filter of size 3 3

Smoothing filter of size 5 5

1 31
g1 (2, 2)  [3  1  2  2  7  5  0  6  5]   3.44  3
9 9
139
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
1
g 2 (2, 2)  [0  1  0  6  5  2  3  1  2  5  1  2  7  5  4  1  0  6  5 
25
81
2  2  3  5  5  7  6]  9
9

b) 3  3 weighted average mask is

1 2 1
1
2 4 2
16
1 2 1

Thus for weighted average filter,

1
g 3 (2, 2)  [3  1  1  2  2  1  2  2  7  4  5  2  0  1  2  6  5  1]
16

66
  4.125  4
16

c) Median filter of 3  3 , sort the values in ascending order

g 4 (2, 2)  median {0,1, 2, 2, 3, 5, 5, 6, 7}  3

Median filter of size 5  5 , sort the values in ascending order

g 5 (2, 2)  median [0, 0, 0,1,1,1,1, 2, 2, 2, 2, 2, 3, 3, 4, 5, 5, 5, 5, 5, 6, 6, 6, 7, 7]  3

d) For min filter of size 3  3

g 6 (2, 2)  0
g 7 (2, 2)  0
g 8 (2, 2)  7

For min filter of size 5  5

e) For max filter of size 3  3 and 5  5

***
Try these exercises.

E5) What is a Median filter?

E6) What is maximum filter and minimum filter?

E7) Differentiate linear spatial filter and non-linear spatial filter.

Now, in the following section, we shall discuss sharpening spatial filters.

4.4 IMAGE SHARPENING

140
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
Image sharpening is opposite of image smoothing. This is done to highlight
fine details and edges in an image. Applications of image sharpening are
outlining object boundaries in industrial applications, identifying tumor
boundaries, or bone structures in medical imaging etc. Smoothing is achieved
by pixel averaging which is analogous to integration. Image sharpening is just
the reverse process, and is achieved by differentiation. The derivative
operation enhances the degree of discontinuity in an image.

The advantages of image sharpening are the following.

• It enhances edges and other discontinuities (noise) in an image.

• It de-emphasizes area with slowly varying grey levels (background) in an

image. First order derivative of a one-dimensional signal f(x) is defined as

f
 f ( x  1)  f ( x ) (8)
x

Similarly 2nd order derivative f ( x , y) is given by

 2f
 f ( x  1)  f ( x  1)  2f ( x ) (9)
x 2

Example 5: Find first and second derivative of given data

f (x) : f (x)  [4 3 2 5 9] .

f
Solution:  [3  4, 2  3, 5  2, 9  5]
x
 [1,  1,  3, 4]
 f
2
 [ 4  2  6, 3  5  4, 9  2  10]
2x
 [0, 4, 1]
***
With respect to the first and second derivatives, following things are observed

i) In flat regions (region of constant intensity levels) both derivatives produce

zero.
ii) In ramp (constant slope), first derivatives gives a non-zero value, whereas
second derivative gives non zero values (with different sign) only at the
beginning and end of ramp and zero along the ramp. First derivative
produces a thick edge called ‘double edge effect’ along the ramp as the
output is non zero. Second derivative produces a double edge one pixel
thick separated by zeros.
iii) For a noisy pixel, first derivative produces a positive or negative value
leading to one zero crossing. Whereas second derivative produces two
zero crossing making it easier to identify.
iv) In case of an edge (step change in intensity values), first derivative
produces a sign change leading to a zero crossing. This makes edge
detection easier by second derivative. Second derivative enhances the
details and edges much better as compared to first derivative. Thus for
image sharpening applications, second derivative is most suitable.

141
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
Here, we will discuss some sharpening filters base on first derivatives and
second derivatives.

4.4.1 First Derivatives Filters

Gradient Operator

An edge is the boundary between two regions with distinct grey level
properties. Derivative operators are used for most edge detection techniques.
The magnitude of first derivative calculated within a neighborhood around the
pixel of interest is used to detect presence of edge in an image. First
derivatives are directional operators and is defined as

 f 
g x   x 
f      f  (10)
g y   
 y 

Magnitude of this vector is given by

 f  2  f  2 
f  mag(f ( x , y))  g  g       
2
x
2
y (11)
 x   y  
In common practice, gradient with absolute values are simpler to implement.
Thus
f f
f  g x  g y   (12)
x y
Robert Operator

Consider the a pixel of interest f ( x , y)  z 5 and a rectangular neighborhood of

size 3 × 3 in Fig. 12 (a).

1 0 1 1
1 0 0 0

(a) Robert Operator (b)

1 0 0 1
0 1 1 0

(c)

Fig. 12

Equation can be approximated at point z 5 in a number of ways. Simplest is

g x  z 5  z8 (13)

142
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
g y  z5  z6 (14)

f  z 5  z8  z5  z6 (15)

Another approach is to use cross difference to realized

f ~
 z 5  z 9  z 6  z8 (16)
Equations can be implemented by the masks in Fig. 12 (a) and Fig. 12 (c). The
original image is convolved with both the masks separately and the absolute
values of the two outputs of convolution are added.

Although Roberts operator illustrates the derivative process, there is a difficulty

in implementation. Since only two neighboring pixels are used, it is not clear
as to exactly where the derivative value should be stored in the output image.
This difficulty is taken care of in other filters like Prewitt and Sobel filters which
are symmetric in nature.

Prewitt Operator

In the Prewitt Operator, the edge value at f ( x , y) is calculated by taking the -

difference of f ( x  1, y) and f ( x  1, y) . To make sure that the difference is not
coming from two arbitrary points, three pairs of points are taken together to
result in one edge value. This creates a 3  3 mask.

Fig. 13: Perwitt Operator

The above mask in Fig. 13 can be used to identify the presence of a vertical
edge in the following image:

30 30 30 80 80 80
30 30 30 80 80 80
30 30 30 80 80 80
30 30 30 80 80 80
30 30 30 80 80 80
30 30 30 80 80 80

f ~
 (z 7  z 8  z 9 )  (z1  z 2  z 3 )  (z 3  z 6  z 9 )  (z1  z 4  z7 )

Sobel Operator
Another sharpening first order filter is the Sobel Operator. It is slightly different
from the Prewitt filter, in that the middle pair of pixels are given higher weight
compared to other two pairs of pixels. Both the filters are also called EDGE
FILTERS. Sobel is the most popular 3  3 Edge operator. It is described by

f  (z 7  2z 8  z 9 )  (z1  2z 2  z 3 )  (z 3  2z 6  z 9 )  (z1  2z 4  z7 )

Sobel mask is shown in Fig. 14. Notice that all mask coefficients sum up to
zero, thus giving no response in the area of constant intensity.
143
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Fig. 14: Sobel Operator

Sobel operator and Prewitt operator are similar to each other. Sobel operator
is also used to detect two kinds of edges in an image; vertical and horizontal.

You can observe from the masks given in Fig. 14, that first masks has only
one difference in the members of first and third rows that is of sign, that is it
has “2” and “–2” values in center of first and third rows. When applied on an
image this mask will highlight the horizontal edges. Similarly, the second mask
in Fig. 14 shows the vertical edges.

When we apply this mask on the image it prominent vertical edges. It simply
works like as first order derivative and calculates the difference of pixel
intensities in an edge region.

As the center column is of zero so it does not include the original values of an
image but rather it calculates the difference of right

Now, let us discuss the filters based on second derivatives.

4.4.2 Laplacian Operator

For a 2D function f ( x , y), the gradient (first derivative) is defined as

f ( x , y) f f
f   
x y x y

Laplacian (second derivative) is a rotation invariant and linear operator and

it is defined as
 2 f ( x , y)  2 f f 2
2 f   2 2 (17)
x 2 y 2 x y
 2f
 f ( x  1, y)  2f ( x , y)  f ( x  1, y)  [1  2 1]
x 2
1
 2f
 f ( x , y  1)  2f ( x , y)  f ( x , y  1)   2
y 2
 1 
Thus 2f  [f ( x  1, y)  f ( x  1, y)  2f ( x , y)]  [f ( x , y  1), f ( x , y  1)  2f ( x , y)
2f  f ( x  1, y)  f ( x  1, y)  f ( x , y  1)  f ( x , y  1)  4f ( x , y)
  # *

144
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
(a) (b) (c) (d)

1 2 1 0 1 0 1 1 1
2 4 2 1 4 1 1 8 1
1 2 1 0 1 0 1 1 1

(e) (f) (g)

Fig. 15: Masks for laplacian operator

Fig. 15 (b) shows a 3 3 sharpening filter mask. Fig. 15 (a) shows the
coordinate notations and Fig. 15 (d) shows the location of non zero elements.
Notice that addition of all the coefficients of the mask produces zero. In
this mask, diagonal elements are not included. Fig. 15 (c) shows a mask with
diagonal elements also. Fig. 15 (e) shows a weighted Laplacian mask. Notice
that the centre element is negative. Fig. 15 (f) and Fig. 15 (g) show two
Laplacian masks with centre element positive which gives the same results.
Laplacian operation is a derivative operation which highlights the intensity
discontinuities of an image and de-emphasize region with slowly varying
intensity values. This tends to produce image that have greyish edge lines with
dark featureless background. (Fig.16). Background features can be
‘recovered’ while ‘preserving’ the sharpened effect of Laplacian by
subtracting Laplacian from the image. Thus, sharpened imaged is generated
by the following operation.

Fig 16: Output after laplacian operator

g ( x, y)  f ( x , y)  2f ( x , y)
 f ( x , y)  [f ( x  1, y)  f ( x  1, y)  f ( x , y  1)  f ( x, y  1)  4f ( x, y)]
 5f ( x , y)  f ( x  1, y)  f ( x  1, y)  f ( x , y  1)  f ( x, y  1)

A mask is generated for sharpening of images.

0 1 0 0 0 0 0 1 0
1 5 1  0 1 0  1  4 1
0 1 0 0 0 0 0 1 0

Similarly

1 1 1 0 0 0 1 1 1
1 9 1  0 1 0  1  8 1
1 1 1 0 0 0 1 1 1

These are high frequency enhancement filters that pass all frequencies and
enhances high frequencies . Notice that all the coefficients in the laplacian
mask add up to one, so convolution with this mask will not change the pixel 145
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
values in the part of the image that has constant grey levels. But if the pixels
have higher or low grey level values than its neighbours, this contrast will be
enlarged by the filter.

Example 5: For the image segment f ( x , y) given below. Apply laplacian filter
of size 3  3 and 5  5 on the centre pixel.

Solution: Laplacian 4-connectivity mask is

0 1 0
1 4 1
0 1 0

Thus, for Laplacian filter,

g 9 (2, 2)  [1  2  5  6  4  7]  14

Laplacian 8-connectivity mask is

1 1 1
1 8 1
1 1 1

Thus for Laplacian filter

g10 (2, 2)  [3  1  2  2  5  0  6  5  8  7]  22

***
Try these exercises.

E8) How are various filter mask generated to sharpen images in spatial
filtering?
E9) Explain spatial filtering in image enhancement.

E10) What is meant by Laplacian filter? Using the second derivative,

develop a Laplacian mask for image sharpening

E11) What are image sharpening filters? Explain the various types of q
sharpening filters.

E12) Name the different types of derivative filters.

E13) Suggest typical derivative masks for Image enhancement. a) Roberts

b) Prewitt c) Sobel.

E14) Write the applications of sharpening filters.

4.5 HISTOGRAM PROCESSING

Before we begin with histogram processing, let us define what is the histogram
of an image?
Histogram
146
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
Histogram of an image represents the number of times a particular grey level
has occurred in an image. It is a graph in which the x-axis represents gray
levels, and y -axis represents number of pixels for each grey level.
Histogram of an image is defined as h (rk )  n k , where rk  k th grey level
n k  number of pixels with grey level rk , and k takes on the values
0,1, 2, ..., L  1.
Example 1: Find histogram of image in Fig. 1.

0 0 0 0
0 1 2 3
0 2 4 6

Fig. 1: Image

Solution: A table is generated with rk and n k , where rk contains all possible

grey level values 0,1, 2, 3, 4, 5, 6 and n k consists the number of times that
grey value has occurred in the Fig. 1. 0 has occurred 6 times, 1 has occurred
1 time etc. Then histogram is drawn in Fig. 2.

rk 0 1 2 3 4 5 6
nk 6 1 2 1 1 0 1

Fig. 2: Histogram
***
Normalized histogram is obtained by dividing the frequency of occurrence of
each gray level rk by the total number of pixels in an image.
nk
p(rk )  , where k  0........ L  1 , where n  total number of pixels in the
n
image p( rk ) gives the probability of occurrence of grey level rk . The sum of all
components of a normalized histogram is equal to 1.
Example 2: Find normalized histogram of the image

0 0 0 0
0 1 2 3
0 2 4 6

Solution: Fig. 2 image is similar to Example 1. A table is made with rk , n k and

p(rk ) and normalized histogram is drawn in Fig. 3.
rk 1 1 2 3 4 5 6
nk 6 1 2 1 1 0 1
P(rk ) 6 1 2 1 1 0 1
12 12 12 12 12 12

147
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Fig. 3: Normalized Histogram

Now, you might be thinking that why histogram is needed. We shall answer
this now.
Histogram gives an insight about the contrast in an image. It tells us the
difference between average grey level of an object and that of the
surroundings. It also provides a useful image statistics which are helpful in
various image processing applications, for example, thresholding, intensity
level slicing, segmentation etc. Besides it is very simple to calculate histogram
of an image. Intuitively, it tells how vivid or washed out an image appears.
Based on histogram, we can categorize images in the four categories, which
are defined below.
i) Under exposed image: Fig. 4 (a) shows an under exposed image, and
its histogram is shown in Fig. 4 (b). It may be noted that almost all the
pixels are concentrated at the lower end of the histogram. The maximum
value of grey is less than 50. No details are seen in the image.
ii) Over-exposed image: Fig. 4 (c) shows an over-exposed image and its
histogram is shown in Fig. 4 (d). The histogram is prominent towards the
higher side of the grey values with the lowest grey level value around 150.
This image has a washed-out look because of the absence of darker
shades of grey.
iii) Low contrast image: Fig. 4 (e) shows a low contrast image and its
histogram in Fig. 4 (f). The shape of the histogram is narrow and centered
towards the middle of the grey scale. Very few grey levels participate in
the image formation as neither the lower grey levels nor the higher grey
levels are present. Thus, details are not visible in the image.
iv) High contrast image: Fig. 4 (g) shows a high contrast image and its
histogram in Fig. 4 (h). This histogram covers a broad range of grey
levels, but distribution of pixels is not uniform. The number of some grey
levels is very high as compared to the number of other grey levels.

You can visualize histograms of different images when you will be doing your
scilab sessions as given in Block 5 (Lab Manual). Now let us discuss various
applications of histogram.

Applications of Histogram
Histogram is widely used in image processing applications.

i) Histogram gives a quick indication as to whether or not we have used the

entire dynamic range of digitizer. Information is lost if dynamic range is not
used fully.
148
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

(a) under exposed image (b) histogram

(c) over-exposed image (d) histogram

(e) low contrast image (f) histogram

(g) high contrast image (h) histogram

Fig. 4
ii) Histogram indicates if clipping is a problem in the image.
iii) It is extensively used in thresholding to separate contrasting objects from
background.
iv) It tells us about image contrast.
149
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
Now, try the following exercises.

E15) What is a histogram?

E16) Highlight the importance of histograms in image processing.

In the following section, we shall discuss histogram equalization.

4.6 HISTOGRAM EQUALIZATION

The histogram of a poor quality image would show presence of only few grey
levels, while most would be missing. The number of pixels for various grey
levels vary widely. Histogram equalization is a point operation that maps an
input image onto a new output image such that there are (almost) equal
numbers of pixels at each grey level in output. Thus it can be used for contrast
enhancement.

Given information: Input image f ( x , y) from which we can calculate its

histogram h ( rk ) .

Goal: To obtain a uniform histogram for the output image. To devise a point
operation s  T ( r ) that maps input image f ( x , y) to an output image g( x , y)
that has a flat histogram as shown in Fig. 5.

Fig. 5: Histogram Equalization

This is done by adjusting the probability density function (PDF) of the
histogram of original image so that probability density function (PDF) spreads
equally. We want to find a transformation of the image given in Fig. 6 (a) by
the following function:

s  T (r ), 0  r  1 … (1)

which should satisfy following conditions:

1) T (r ) is a single valued function (one to one relationship) as shown in Fig.

6 (b). For each value of r there should be a single value of s. This is
needed so that inverse transform exists.

2) T (r ) is a monotonically increasing function in the interval 0  r  1 ,

Amonotonically increasing function is one which keeps on growing ( once
it goes up it does not) comedown. This preserves intensity level order in
the output image from black to white. This transformation doesn’t cause a
150
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
negative image. Fig. 6 (c) shows a non-monotonical function and Fig. 6
(d) shows a monotonically increasing function.

3) The range of r and s is same, that is for 0  r  1, 0  T ( r )  1 . This

guarantees that the output grey levels are in the same range as input grey
levels.

(a) Transfer Function (b) Single Valued Function

(c) Non Monotonical Function (d) Monotonical Increasing Function

Fig. 6

Inverse transformation of the function given in Eqn. (1) is defined as

r  T 1 (s) 0  s  1 .

In practice, for integer values of images, many times condition 1 is violated

and non-unique inverse transformation is generated.

Probability Density Function (PDF)

The grey levels in an image may be viewed as random variables in the interval
[0,1] . Probability density function (PDF) is one of the fundamental descriptors
of a random variable. PDF of a random variable x is defined as the derivative
of cumulative distribution function (CDF)

dF( x )
p( x ) 
dx
F( x )  P ( X  x )
Where p( x )  PDF of x
F( x )  CDF of x
p  Probability of x
PDF satisfies the following properties:

i) P( x )  0 for all x

ii)  p(x) dx  1

151
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
x
iii)  p(x) d  where  is a dummy variable

x2

iv) p( x 1  x  x 2 )   p(x ) dx
 x1

If a random variable x is transformed by a monotonic transfer function T ( x ) to

produce a new random variable y , then the PDF of y can be obtained from
knowledge of T ( r ) and PDF of x
dx
p y ( y)  p x ( x )
dy

Let p r ( r ) be PDF of random variable r of input image.

Ps (s) be PDF of random variables r of output image.
If p r (r ) and T ( r ) is know then Ps (s) can be obtained by
dr
p s (s)  p r (r )
ds

The PDF of transformed variables s is determined by pdf of input image and

by chosen transformation function.

Let us now discuss the discrete transformation function.

Transformation function is a cumulative distribution function (CDF) of r .

r
s  T(r )   p r ( w ) d ( w )
0
Where w is dummy variable.

CDF is an integral of a probability function (always positive). Thus CDF is

always single valued and monotonically increasing function. CDF satisfies the
conditions of transformation function, hence can be considered as T () .
Differentiating above equation with respect to r , we get

ds dt (r ) d  
r
   p r ( w ) dw 
dr dr dr  0 
 p r (r )
dr 1
p s (s)  p r (r )  p r (r ) 1
ds p r (r )

Thus, p s (s)  1 , where 0  s  1

Fig. 7
152
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
p s (s) is PDF of output image, whose value is 1 for all values of s. Thus p s (s)
is a uniform PDF independent to the form of p r (r) as shown in Fig. 7.

Discrete Transformation Function

Although the above discussion shows that any image can be transformed to a
flat histogram, it is only possible if grey values are varying continuously. In
practice any image will have discrete grey values. Thus the integration above
is replaced by a summation function. This means that the output image will not
have a flat histogram but only an approximation to it.
Probability of occurrence of grey levels in an image given by
nk
p r (rk )  , where k  0.............L  1 .
n
Transformation is given by
k
s k  T(rk )   p r (rj ) , where k  0, ....L  1 .
j0

Substituting value of p r (rk ), we get

k nj
rk  
j 0 n
The output image is obtained by mapping each pixel with grey level rk in the
input image to a corresponding pixel with grey level s k in output image.

(a) Low Contrast Image (b) Histogram Equalized Image

(c) Histogram of (a) (d) Histogram of (b)

Fig. 8
Example 3: For the given 4 4 image having grey scales between [0, 9] , carry
out histogram equalization. Also, draw the histogram of image before and after
equalization.

153
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
2 3 3 2
4 2 4 3
3 2 3 5
2 4 2 4

Solution: Given is

a) The grey levels are between [0 , 9] .

b) Size of image is 4  4 .

Following steps are to be followed for histogram equalization:

Step 1: Find the histogram of grey levels of input image by making a table of
rk (grey levels) and number of pixels (n k ) in the input image.

k
Step 2: For each input grey level, compute the cumulative n
j 0
j sum by

adding number of pixels for each grey level.

Step 3: Divide the cumulative sum by the total number of pixels in the image
(4 × 4 = 16) to generate the CDF of the given image.

Step 4: Scale the output by multiplying step 3 to maximum grey level value (9
in this case)

Step 5: Round off the output of step 4 to the nearest integer value

We have summarized step 1 to step 5 in the following table.

Grey pixel r j 0 1 2 3 4 5 6 7 8 9
Step No. of pixels 0 0 6 5 4 1 0 0 0 0
1 nj
Step Commulative
2 k 0 0 6 11 15 16 16 16 16 16
sum n
j 0
j

Step k 0 0 6 11 15 16 16 16 16 16
s   nj
n
3 16 16 16 16 16 16 16 16
j 0

Step s9 0 0 6 11 15 16 16 16 16 16
4 9 9 9 9 9 9 9 9
16 16 16 16 16 16 16 16

Step Round off s k 0 0 3 6 8 9 9 9 9 9

5

Step 6: Map input levels (rk ) to output levels s k in a tabular form from 1st and
last row of table above to generate

rk 0 1 2 3 4 5 6 7 8 9
sk 0 0 3 6 9 9 9 9 9 9

Step 7: Formulate new image g ( x , y) by replacing each level of f ( x , y) by a

new one using table above.
154
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
2 3 3 2 3 6 6 3
equalized
4 2 4 3 8 3 8 6
3 2 3 5 6 3 6 9
2 4 2 4 3 8 3 8
f ( x , y) g ( x , y)

Step 8: We draw output and input histogram as shown in Fig. 9.

Fig. 9
It may be noted that while in the original image the histogram spans between
grey values 2 to 5, in the output image, the grey values span from 3 to 9.
Although some equalization has been achieved, some of the grey values are
still missing. It is because the image has only 9 levels. Histogram equalization
will be very effective in an image with 256 grey levels.

Now, try the following exercises.

E17) Write steps for a procedure to perform histogram equalization.

E18) What is the role of histogram equalization in image enhancement?
Why this technique yields a flat histogram?
E19) Equalize the histogram of 8 × 8 image g ( x , y) shown below. The input
image has grey levels 0, 1, ..........., 7.

4 4 4 4 4 4 4 0
4 5 5 5 5 5 4 0

4 5 6 6 6 5 4 0
 
4 5 6 7 6 5 4 0
4 5 6 6 6 5 4 0
 
4 5 5 5 5 5 4 0
4 4 4 4 4 4 4 0
 
 4 4 4 4 4 4 4 0

In the next section we shall see how to find transformation for an image, such
that the new histogram follows a specific shape. It is called Histogram
Specification.

155
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

4.7 HISTOGRAM SPECIFICATION

We shall begin with the definition of histogram specification.
Definition: Histogram specification is a point operation that maps input image
f ( x , y) into an output image g ( x , y) with a user specified histogram. An
example of specified histogram is shown in Fig. 9.

Fig. 9: Histogram specification

The main application of histogram specification is that it improves contrast and
brightness of images. It is a pre processing step in comparison of images.
Comparison to histogram equalization

Histogram equalization has a disadvantage that it can generates only one type
of output image where histogram is flat, which may not be the best approach
under all circumstances. Histogram specification is more general than
equalization. It is an interactive enhancement technique where user can draw
desired histogram. We can specify the shape of histogram which need not be
uniform. Thus, histogram specification gives us the flexibility to choose
histogram for reference and output image is mapped accordingly. The basic
idea is to carry out histogram equalization of input image histogram, then carry
out histogram equalization of specified histogram. The two transformations are
used to find the required transformation.

Algorithm for histogram equalization: In this procedure, there are three

variables:

Pr (r ) is pdf of grey level r of input image

Pz (z) is pdf of grey level z of specified image
Ps (s) is pdf of grey level s of output image


The transformation is s  T ( r )  p r ( r ) dr
0
s


Histogram equalization of input image G ( z)  p z ( z) dz
0
To find the histogram equalization of specified image, we equate G ( z) is
equated to s and an inverse transformation is computed as given below:
156
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
G (z)  s  T(r )
 z  G 1[ s ]  G 1[T (r )]

Assuming that G 1 exists, we can map input grey levels r to output grey levels
s.
Procedure for histogram specification

Step 1: Obtain the transformation T ( r ) by doing histogram equalization of

input image.
r
s  T (r )   p r (r ) dr
0
Step 2: Obtain the transformation G ( z) by doing histogram equalization of
specified image.
z


Step 3: G ( z)  p z ( z) dz
0
Equate G (z)  s  T ( r )
Step 4: Obtain inverse transformation function G 1 .

z  G 1[ s ]  G 1[T(r )]
Step 5: Obtain the output image by applying inverse transformation function
to all pixels of input image.
Let us now apply these steps in the following example.

Example 4: Assume an image having given grey level PDF Pr ( r ) . Apply

histogram specification with given desired PDF function Pz ( z) given below

 2r  1; 0  r  1 2 z ; 0  z 1
p r (r )   p s (r )  
 0 ; otherwise  0; otherwise

Apply histogram specification with the desired PDF function p z ( z) given in Fig.
10:

(a) (b)
Fig. 10
Solution: We shall solve it step wise.
Step 1: Obtain transformation function T ( r ) by doing histogram equalization
of input image.
r r
s  T (r )   p r (r ) dr   (2r  2) dr  [ r 2  2r ]0r  r 2  2r
0 0
Step 2: Obtain transformation function G ( z)
157
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
z z
G (z)   p z (z) dz   2z dz  [z 2 ]0z  z 2
0 0
Step 3: Equate

s  T(r )  G (z)
 r  2r  z 2
2

Step 4: Obtain inverse transformation G 1 .

z  G 1[T(r )]
z   r 2  2r
***
Now, we shall find the discrete transformation function for an actual image.

Discrete Transformation Function for an actual Image

Histogram equalization of input image.

k
s k  T(rk )   p r (rj ), k  0,..., L  1
j0
k nj
sk  
j0 n

where n  total number of pixels in input image

n j  number of pixels having level j

Histogram equalization of specified image

q
v q  G (z q )   p z (z i ), q  0,.., L  1
i 0

Equate G ( z q )  s k  T ( rk )

Inverse transformation z q  G 1 [s k ]  G 1[T ( rk )]

The block diagram of histogram specification is given in Fig 11.

v * is that value of v q so that min [ v q  s]  0

p  Corresponding value of z .

Fig. 11: Histogram Specification 158

UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Procedure for histogram specification of an image

Step 1: Equalize input image histogram (s k )

Step 2: Equalize specified image histogram ( v q )
Step 3: For min q[ v q  s]  0 find corresponding v * and p .
Step 4: Map input pixels to output pixels to get output image.

We shall apply these steps in the following example.

Example 6: Apply histogram specification on given image having

0 1 0 2
ri  z i  0,1, 2, 3
2 3 3 2
p r (ri )  .25 for i  0,1, 2, 3
3 1 0 1
p z (z 0 )  0, p z (z1 )  .5, p z (z 2 )  .5, p z (z 3 )  0
1 3 2 0

Solution: We shall apply each step of the procedure.

Step 1: Equalize input image histogram.

rk 0 1 2 3
p r (rk ) .25 .25 .25 .25
sk .25 .5 .75 1

Step 2: Equalize specified image histogram

zq 0 1 2 3
p z (rz ) 0 .5 .5 0
vq 0 .5 1 1

Step 3: Find minimum value of q such that ( v q  s)  0 . First three columns

are filled by step 1, next three columns are filled by step 2. In this
step, last two columns are filled by the fallowing procedure.

rk p r (rk ) sk zq p z (z q ) vq v4 p
0 .25 .25 0 0 0 .5 1
1 .25 .5 1 .5 .5 .5 1
2 .25 .75 2 .5 1 1 2
3 .25 1 3 0 1 2 2
a) q  0, k  0 ( v 0  s 0 )  (0  .25)  0  No  increase q
q  1, k  0 ( v1  S0 )  (.5  .25)  0  Yes  v * 0  v1  .5 and p 0  z1  1

b) q  1, k  1 ( v1  s1 )  (.5  .5)  0  Yes  v *1  v1  .5 and p1  z1  1

c) q  1, k  2 ( v1  s 2 )  (0  .75)  0  No  increase q
q  2, k  2 ( v 2  s 2 )  (1  .75)  0  Yes  v * 2  v 2  1 and p 2  z 2  2

d) q  2, k  3 ( v 2  s 2 )  (1  .75)  0  Yes  v * 3  v 2  1 and p 3  z 3  2

159
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

Step 4: Map input level values to output level values

rk 0 1 2 3
p 1 1 2 2

Step 5: Map input pixels to new values to get new image using step 4 output.

0 1 0 2 1 1 1 2
2 3 3 2 2 2 2 2
0 1 0 1 2 1 1 1
1 3 2 0 1 2 2 1
f ( x , y) g ( x , y)

This pixel mapping is used to map the pixels of input image to generate output
image with specified histogram. Histogram specification is a trial and error
process. There are no rules for specifying histogram and one must resort to
analysis on a case by case basis for any given enhancement task.
***
Try the following exercises.

E20) Define the concept of Histogram matching with appropriate example.

E21) Perform histogram specification from the given data.

r  0,1, 2, 3, 4
p(ri )  0.2 for i  0.......4
z i  0, 2, 4, 5, 7
p(z 0 )  0, p(z 2 )  0.2, p(z 4 )  0.4, p(z 5 )  0.4,
p( z 7 )  0

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

4.8 SUMMARY
In this unit, we have discussed the following points.

1) Filtering is a process that removes some unwanted components or small

details in an image. In digital image processing, filter is basically a
subimage or a mask or kernel or template or window. Filters can be of two
types: Spatial filters and frequency domain filters.

2) A filter that passes low frequencies is called a lowpass filter. This type of
filter is used to blur (smooth) the image, therefore called smoothing filter
also. These filters are also used for noise reduction. Noise reduction can
be done by blurring with a linear filter, (in which operation performed on
the image pixels is linear) or a non-linear filter.

3) This filter is used to highlight transitions in intensity. These are based on

first and second derivatives.

160
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
4) Linear filtering is a spatial domain process where a filter (mask/ kernel/
template) with some integer coefficient values is applied to input image to
generate the filtered/ output image.

5) Each pixel in the smoothened image g ( x , y) is obtained from the average

pixel value in the neighborhood of ( x , y) in input image. Such a mask is
also known as a Mean filter.

6) Like all other spatial filters, non linear filters compute the result at some
position ( x , y) from the pixels inside the moving region S of the original
image. These filters are called non-linear because source pixels are
processed by some non-linear function.

7) Median filters are edge preserving smoothing filters, where the level is set
to the median of pixel values in the neighborhood of that pixel.

8) Min filter removes salt noise (white dots with large grey level values)
because any large grey level with in a 3  3 filter region is replaced by one
of its surrounding pixels with smallest value. As a side effect, min filter
introduces dark structures in the image. The reverse effect is expected
from a max filter. It removes pepper noise (black dots with small grey
level values) because any black dot within 3  3 filter region is replaced by
one of its surrounding pixels with the largest value. White dots/bright
structures are widened as a side effect and black dots (pepper noise) will
disappear.

9) Image sharpening is opposite of image smoothing. This is done to

highlight fine details and edges in an image.

10) Defined histogram

11) Understood how to draw histogram from the image
12) Perform histogram equalization
13) Perform histogram specification

4.9 SOLUTIONS/ANSWERS
E1) Linear filtering is a spatial domain process where a filter
(mask/kernel/template/window) with some coefficient values is applied to
input image to generate the filtered/ output image. Generally, filter size is
either 3  3 or 5  5, 7  7 or 21 21 (odd sizes) and filter is centered at a
coordinate ( x , y) called ‘ Hot Spot’.

E2) Smoothing linear filters are used to reduce ‘sharp’ transitions in grey
levels, reduce noise, bluredges, help in smoothing false colours, reduce
‘irrelevant’ details in an image.

Sharpening linear filters are used to detect edges, lines, points in the
image. It is also used to highlight all high frequency components in the
image.

E3) Each pixel in the smoothened image g ( x , y) is obtained from the

average pixel value in the neighborhood of ( x , y) in input image.

161
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
1
g ( x , y)   f ( m, n ) w ( m, n )
M ( m ,n )Sxy

1 1 1 1 2 1 
1   1  
Various masks are  1 1 1
9   16  2 4 2
1 1 1 1 2 1 

Following steps are required for spatial averaging

1) Position the mask over the current pixel such that hotspot w(0,0)
coincides with current pixel.
2) Form all products of filter elements with the corresponding elements
in the neighbourhood.
3) Add up all the products and store it at current position in the output
image. This must be repeated for every pixel in the image.

E4) The aim of image smoothing is to diminish the effect of camera noise,
spurious pixel values, missing pixel values etc.

E5) Median filters are edge preserving smoothing filters, where the level is
set to the median of pixel values in the neighborhood of that pixel. It is
impossible to design a filter that removes only noise and retains all the
important image structures intact, because no filter can discriminate
which image content is important to the viewer and which is not. Median
filter replaces every image pixel by median of the pixels in the
corresponding filter region Sxy.

g ( x , y )  median {f ( x  i, y  j) (i, j)  Sxy }

E6) Max and min filters are defined as:

g ( x , y)  min{f ( x  i, y  j) (i, j)  S xy }
g ( x , y)  max {f ( x  i, y  j) (i, j)  S xy }

where S xy denotes the filter region, usually a size of 3  3 pixels.

E7) Linear filters produce an output that is a linear combination of the input.
For example, smoothing linear filters produce output as the average of
input while sharpening linear filters produce output as the first or second
derivative of the input.

Non linear filters produce a statistical output where median filter produce
output as median of the input values, min filter produces output as
minimum of all input values and max filter produces output that is
maximum of all input values.

 2 f ( x , y)  2 f f 2
E8) 2 f   2 2
x 2 y 2 x y

162
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
E9)
1 2 1 0 1 0 1 1 1
2 4 2 1 4 1 1 8 1
1 2 1 0 1 0 1 1 1

Spatial filtering is one of the main tools used in variety of applications

such as noise removal, bridging the gaps in object boundaries,
sharpening of edges etc.The idea of spatial filtering is to move a ‘mask’,
a rectangle (usually with size of odd length) or other shape over the
image. By this process, we create a new image where grey level values
of the pixels are calculated from the values under the mask. The values
under the mask are modified by a function called ‘filter’. If this filter
function is a linear function of all grey level values in the mask, then filter
is called a ‘linear filter’, else it is called ‘non linear filter’.

E10) Laplacian (second derivative) is a rotation invariant and linear

operator and it is defined as
 2 f ( x , y)  2 f  2f
2 f   2 2
x 2 y 2 x y
 2f
 f ( x  1, y)  2f ( x , y)  f ( x  1, y)  1  2 1
x 2
1 
 2f
 f ( x , y  1)  2f ( x , y)  f ( x , y  1)   2
y 2
 1 
Thus

 2f  [f ( x  1, y)  f ( x  1, y)  2f ( x , y)]  [f ( x , y  1), f ( x , y  1)  2f ( x , y)]

 2f  f ( x  1, y)  f ( x  1, y)  f ( x , y  1)  f ( x , y  1)  4f ( x , y)

1 2 1 0 1 0 1 1 1
2 4 2 1 4 1 1 8 1
1 2 1 0 1 0 1 1 1

E11) Image sharpening is done to highlight fine details and edges in an

image. Applications of image sharpening are industrial applications,
medical imaging etc. Sharpening is reverse of smoothing which is
achieved by pixel averaging which is analogous to integration. Thus,
sharpening is achieved by differentiation. The derivative operation
enhances the degree of discontinuity in an image.
Sharpening
• Enhances edges and other discontinuities (noise) in an image.
• De-emphasizes area with slowly varying grey levels (background) in an
image.
Different image sharpening filters are Laplacian, gradient, Robert, sobel,
etc.
E12) For a 2D function f ( x , y) , the gradient (first derivative) is defined as
Laplacian (second derivative) is a rotation invariant and linear
operator and it is defined as 163
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

f ( x , y) f f
f   
xy x y

 2 f ( x, y)  2 f f 2
2 f   2 2
x 2 y 2 x y
E13)

1 0 1 1 1 0 0 1
1 0 0 0 0 1 1 0

Robert operator

E14) Sharpening filters are used for edge enhancement applications.

Sharpening linear filters are used to detect edges, lines, points in the
image. It is also used to highlight all high frequency components in the
image.

E15) Histogram is a graph between various grey levels on x -axis and the
number of times a grey level has occurred in an image, on y -axis.
Histogram of an image is defined as
nk
h (rk )  ,  0.............L  1
k
rk  k th grey level
n k  number of birds grey level values as rk .
E16) The histogram of an image represents the relative frequency of
occurrence of the various gray levels in the image. It also provides a
useful image statistics which are helpful in various image processing
applications, for example, thresholding, intensity level slicing,
segmentation etc. Besides it is very simple to calculate histogram.
Intuitively, it tells how vivid or washed out an image appears.
Histograms are the basis for many spatial domain processing
techniques.
E17) Histogram equalization is done by the transformation

k
s k  T (rk )   p r (rj ) , k  0,.......L  1
j 0

E18) Histogram equalization is a point operation that maps an input image

onto output image such that there are equal number of pixels at each
grey level in output. It is used for contrast enhancement.

Given information: Input image f ( x , y) from which we can calculate

its histogram h ( rk )

164
UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN
Goal: To obtain a uniform histogram for the output image. To device a
point operation s  T ( r ) that maps input image f ( x , y) to an output
image g ( x, y) that has a flat histogram

E19) Given size: 8 8 , levels [0 7]

rj 0 1 2 3 4 5 6 7
Step 1 nk 8 0 0 0 31 16 8 1
Step 2 k 8 8 8 8 39 55 63 64
n
j0
j

Step 3 k nj 8 8 8 8 39 55 63 64
Sk  
j 0 n 64 64 64 64 64 64 64 64
Step 4 sk  7 8 8 8 8 39 55 63 64
7 7 7 7 7 7 7 7
64 64 64 64 64 64 64 64

Step 5 Round off 1 1 1 1 4 6 7 7

Step 1: Find the histogram of input image by making a table of rk and n k .

Step 2: For each input level, find cumulative sum

Step 3: Divide cumulative sum by total number of pixels n  64 .

Step 4: Scale the output by multiplying step 3 by 7.

Step 5: Round off to nearest integer value to find s k .

Step 6: Map input levels rk to output levels s k using 1st and last row of
table above to make table below.
rk 0 1 2 3 4 5 6 7
sk 1 1 1 1 4 6 7 7
Step 7: Formulate a new image g ( x , y) by replacing each level of
f ( x , y) by a new one using table.

Step 8: Draw input and output histograms.

4 4 4 4 4 4 4 0 4 4 4 4 4 4 4 1
4 5 5 5 5 5 4 0 4 6 6 6 6 6 4 1
4 5 6 6 6 5 4 0 4 6 7 7 7 6 4 1
4 5 6 7 6 5 4 0 4 6 7 7 7 6 4 1

4 5 6 6 6 5 4 0 4 6 7 7 7 6 4 1
4 5 5 5 5 5 4 0 4 6 6 6 6 6 4 1
4 4 4 4 4 4 4 0 4 4 4 4 4 4 4 1
4 4 4 4 4 4 4 0 4 4 4 4 4 4 4 1
f (x, y) g(x, y)

f(x) g(x) 165

UNIT 4 IMAGE FILTERING OPERATIONS IN SPATIAL DOMAIN

E20) Histogram specification is more general than equalization. It is an

interactive enhancement technique where user can draw desired
histogram. We can specify the shape of histogram which needs not to
be uniform. Thus, histogram specification gives us the flexibility to
choose histogram for reference and output image is mapped
accordingly

E21) Step 1: Equalize input histogram.

rk 0 1 2 3 4
p(rk ) .2 .2 .2 .2 .2
sk .2 .4 .6 .8 1

Step 2: Equalize specified histogram

rk 0 1 2 3 4
p(rk ) .2 .2 .2 .2 .2
sk .2 .4 .6 .8 1

Step 3: Find minimum value of q such that [ vq  s]  0

1) q  0, k  0 ( v 0  s 0 )  (0  .2)  0  No  increase q
q  1, k  0 ( v1  s 0 )  (.2  .2)  0  Yes  v *0  v1  .2 and p 0  z1  2

2) q  1, k  1 ( v1  s1 )  (.2  .4)  0  No  increase q

q  2, k  1 ( v 2  s 2 )  (.2  .4)  0  Yes  v *1  v 2  .6 and p1  z 2  4

3) q  2, k  2 ( v 2  s 2 )  (.6  .6)  0  Yes  v * 2  v 2  .6 and p 2  z 2  4

4) q  3, k  3 ( v 2  s 2 )  (.6  .8)  0  No  increase q

q  3, k  3 ( v q  s q )  (1  .8)  0  Yes  v * q  v q  1 and p q  z q  5

5) q  3, k  4 ( v 3  s 4 )  (1  1)  0  Yes  v * 4  v 3  1 and p 4  z 3  5

S.No. rk p(rk ) sk zq p(z q ) vq v* p

1 0 .2 .2 0 0 0 .2 2
2 1 .2 .4 2 .2 .2 .6 4
3 2 .2 .5 4 .4 .6 .6 4
4 3 .2 .8 5 .4 1 1 5
5 4 .2 1 7 0 1 1 5

Step 4: Input and output grey level mapping from 2nd and last row of
previous output

r 0 1 2 3 4
p 2 4 4 5 5

166