I.J.

Image, Graphics and Signal Processing, 2017, 2, 9-18

Published Online February 2017 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2017.02.02

The Image Segmentation Techniques

Shiv Gehlot
Netaji Subash Institute of Technology, New-Delhi, India
Email: [email protected]

John Deva Kumar

PUSA Institute of Technology, New-Delhi, India
Email: [email protected]

Abstract—Image segmentation has a crucial role in is characterized by unique set of properties. Over the
image processing. Classical segmentation techniques years, several image segmentation techniques have
based on thresholding have been extensively used but evolved [6] but none of them can be termed as a
they fail drastically for noisy or non-uniformly generalized algorithm.
illuminated images. Several alternatives presented over In the broad sense, image segmentation algorithms can
the time have filled this void but with increased be categorized into two parts, edge based segmentation
complexity. In this paper we present an algorithm to and region based segmentation. The first method is based
address the above issues with minimum complexity. We on boundary detection between the regions which is
propose normalized self correlation function (NSCF) characterized by high frequency content of the image as
which forms a basis for the progress of the algorithm. We explained above. The classical methods in this category
also introduce relative error function (REF) which is used include, Sobel operator, Roberts operator, Laplace filter,
for qualitative assessment of the algorithm and its Marr-Hilderth method and Canny edge detector [1], [7].
comparison with other algorithms. We also propose a The region based segmentation partition the image into
second algorithm named piecewise image segmentation regions satisfying a predefined criterion. One of the
(PIS) which is a generalized edge-based method able to classical techniques in this domain is thresholding which
generate any desired edge map. The results show that the has multiple extensions like multivariable thresholding
proposed algorithms are able to perform well for different and Otsu‟s method [1].
scenarios and at the same time better than traditional Later, energy minimization technique, originated with
algorithms. snake model [8], became quite popular in this domain.
Among all other methods in this category, the level set
Index Terms—Image segmentation, normalized self method (LSM) has attracted the researchers most and it
correlation function, relative error function, piecewise segments the image by minimizing an energy function
image segmentation, Laplace filter, Otsu‟s method. [9]-[15]. Image segmentation techniques exploiting LSM
are also categorized into two types, edge based models
and region based models [16]-[18]. The first depends on
I. INTRODUCTION edge information while second is based on region
description for the progress of the algorithm [19], [20].
Image processing is simply the mapping of image
In this paper, we present two algorithms, one based on
pixels by a predefined function so that the result is
edge-based models and other based on region-based
suitable for a particular application [1]-[4]. The Image models. The first algorithm begins with calculation of
processing finds uses in multiple areas including normalized self correlation function (NSCF). NSCF map
computer vision, medical science, astronomy, remote
the image intensities to 0-1. NSCF calculates the
sensing etc. [1]-[4].
correlation between the pixels in a predefined
An image can be decomposed into high frequency and neighborhood [21]. The size of neighborhood also affects
low frequency components. High frequency information the performance of the algorithm. The results
is responsible for contrast, reflectance and local details of
demonstrate that algorithm is also able to combat the
the image. Similarly, low frequency part contributes to
noise and non-uniform illumination in the images.
the smoothing, illumination and global naturalness of the The second algorithm is an edge-based method called
image [5]. A delicate balance between the two is required piecewise image segmentation (PIS) and follows the
for the naturally enhanced image. An image rich in low
footsteps of classical algorithms like Sobel operator,
frequency component will be unnaturally blurred, and an
Robert operator and Laplace filter. The algorithm use
image biased towards high frequency component will Laplace filter as the intermediate step but initially pixel
result in a highly bright image. are checked to satisfy a predefined condition. Unlike
One important area of image processing is image
other edge-based methods which produce a complete
segmentation. Image segmentation, in simple words, is
edge-map, PIS can generate any desired edge map. In this
segregating of image into multiple regions such that sense, PIS is a generalized algorithm of which other
intersection of any two regions is null set i.e. each region techniques like Laplace filter forms a special case.

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
10 The Image Segmentation Techniques

The rest of the paper is organized as follows: Section II 1

deals with the introduction of normalized self correlation | A( x, y) | (2)
 N [ I ( x, y)]
function (NSCF), its development and its various aspects.
This section also builds the remaining algorithm taking
NSCF as the basis. In section III relative error function is Both, fig.2(a) and fig.2(b) have A( x, y)  1 as all
proposed for performance analysis of the method. Section
pixels have same intensities in neighborhood around
center pixels. Absence of constant  in (1) for fig.2(b)
IV discusses the results and performance of the algorithm.
In section V the piecewise segmentation algorithm is
leads to indeterminate form, making introduction of
introduced and its methodology is developed. This
constant  compulsory.
section also analyzes the experimental results of the PIS.
Section VI finally concludes the paper.

1 1 1 0 0 0
II. NSCF
In this section we introduce normalized self correlation 1 1 1 0 0 0
function (NSCF) which can be exploited for image
segmentation. Consider an image I of size M  N . 1 1 1 0 0 0
Fig.1 shows a pixel I ( x, y ) of image I in a mask of
predefined size 3  3. Fig.2 (a) Fig.2 (b)

A. Edge detection using NSCF

I ( x  1, y  1) I ( x  1, y) I ( x  1, y  1) Edges in an image are replica of sharp discontinuities
or local intensity variations. Processing an image having
edges using NSCF, maps edge pixels to low intensities
I ( x, y  1) I ( x, y) I ( x, y  1) due to their very low correlation with neighboring pixels.
For the same reasons, background pixels are mapped to
high intensities as they are highly correlated to each other.
Let image I ' is the result after replacing every pixel
I ( x  1, y  1) I ( x  1, y) I ( x  1, y  1)
I ( x, y) of image I by A( x, y) .
Fig.1. Pixel I ( x, y ) of an image I
I I'
A

We define normalized self correlation function (NSCF)

as: As explained above I ' has low intensity pixels at
corresponding edge pixels in I and high intensity pixels
k 1 l 1

  ( f (x, y) f ( x  k , y  l ))  
k 1 l 1
at remaining coordinates. The edge map can be generated
by taking negative of I ' .
k 0 l 0
A( x, y )  (1)
8. f ( x, y )2  
I e   I n' (3)

such that 0  A( x, y)  1 and  is a very small

where I e is edge map of image I , I n' is negative of I ' and
constant, i.e.  0.
After applying A( x, y) to whole image I , every pixel  is a diagonal matrix of same size as I n' .  can be used
I ( x, y) is replaced with A( x, y) . For applying A( x, y) to to adjust the intensity of pixels for display purpose.
boundary pixels, the image has to be suitably padded. B. Image segmentation using NSCF
NSCF will capture the variation in intensities around Above we explain how NSCF can be used for edge
pixel I ( x, y) in a predefined neighborhood. If detection or edge based segmentation of an image.
 N [ I ( x, y)] denotes the intensity variation of pixels Similarly, NCSF can be successfully exploited for region
around pixel I ( x, y) in neighborhood N , then based segmentation. Fig.3 shows an image divided into
two regions such that „region 1‟ and „region 2‟ have
A( x, y)  1 if  N [ I ( x, y)]  0 i.e. A( x, y) . attains its different sets of intensity variations i.e.
maximum value for minimum variation in pixel
intensities. Conversely, minimum value is assigned to  N [ R1( x, y)]   N [ R2( x, y)].
A( x, y) for maximum variation in intensities or
Clearly, application of NSCF produces different
outputs for each region and

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
 The Image Segmentation Techniques 11

A( x, y) R1  A( x, y) R 2 Averaging, indirectly increases the contribution of each

pixel to the center pixel and hence increases the
correlation between the pixels or reduces the variation of
This divides image into two set of intensities, intensities in a predefined neighborhood. As expected,
producing a segmented image. For the region based application of (5) after averaging will produce a brighter
segmentation, we can either use I ' or I n' depending upon image which will produce a poor edge map.
suitability of results. Similarly, we can use I ' or E. Procedure for calculating NCSF of an image
I '
n for further enhancement of results. The steps for calculating NSCF of an image I can be
summarized as follows:

1. Decide the neighborhood of odd size m  m.

2. Calculate a and b using (2).
3. Pad the image I to accommodate the boundary
pixels in the neighborhood defined in step 1.
4. At every pixel I ( x, y ) calculate A( x, y) using (5)
Fig.3. An image having two regions and replace I ( x, y) with A( x, y ) . This will give
image I '.
C. Effect of neighborhood size
5. Using (3) find I e or directly use I ' .
For simplification purpose we consider only odd size
square neighborhood. We define variables a and b such
that III. RELATIVE ERROR FUNCTION
For quantitative description of algorithm, we propose
(m  1)
ab (4) relative error function (REF) which will also contrast the
2 performance of two comparative algorithms. The
motivation behind the approach is the assumption that a
where m  m is neighborhood size. given algorithm applied to an image for an intended
For this general case (1) can be modified to: application will do so in the best possible manner leaving
no scope for further modification by the second pass of
k  a l b the same algorithm. An image I can be mapped by a
  ( f (x, y) f ( x  k , y  l ))  
k  a l  b
function F ( I ) to I m .
k 0 l  0
A( x, y )  (5)
(m2  1). f ( x, y ) 2  
I 
F (I )
Im (6)
Considering larger neighborhood for calculating NSCF
will take into effect the intensities of more pixels. This We define image I m ' as
will help to take a better judgment about intensity
variations around a particular pixel. This indirectly leads I m'  I  | I m | (7)
to better results as compared to smaller neighborhood
size.
As the next step, we apply F ( I ) to I m ' to get I n .
D. Effect of averaging
Fig.4(a) shows a 3  3 averaging filter which is used to I m '  In
F (I )
(8)
replace a pixel I ( x, y ) with average value as calculated
by the filter. F ( I ) has already extracted all the relevant
m m
information in I and subtraction of I from I will
1/ 9 1/ 9 1/ 9 1 1 1

1/ 9 1/ 9 1/ 9 1 8 1

1/ 9 1/ 9 1/ 9 1 1 1

Fig.4(a). A 3  3 averaging filter Fig.4(b). A 3  3 Laplace Filter

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
12 The Image Segmentation Techniques

make I void of any information on which In this sense, I Bm and I Bn should have minimum
F ( I ) flourishes and therefore (8) should show maximum possible deviation because a better algorithm in (8) will
possible deviation from (6) i.e. not spread the pixels to a greater range or otherwise
M 1 N 1

  I
x 0 y 0
m
B ( x, y)  I Bn ( x, y )  should be maximum because deviation between I Bm and I Bn can be used as quantitative
measure of the algorithm. We tap this deviation in form
smaller deviation implies the inability of the algorithm to of REF and define it as:
target the relevant information in the very first pass. Next,
m n
we convert I and I to respective binary images by 1  M 1 N 1 m 
REF    ( I B ( x, y)  I B ( x, y))  (11)
n
calculating a suitable threshold such that both images ( M  N )  x 0 y 0 
have same pixels values for a particular range of
intensities.
where,  denotes ex-or operation and M  N is size of
I n 
T
 I Bn (9) the image I . Clearly, REF measures the deviation
between two images and should have minimum value for
the betterment of the algorithm.
I m 
T
 I Bm (10)

(a) (b) (c) (d) (e)

Fig.5. Results for image Bubbles. (a) Original image. (b) Result of Laplace filter. (c) Zoomed version of (b).(d) Result of NSCF. (e) Zoomed version
of (d).

(a) (b) (c) (d) (e)

(f) (g) (h)

Fig.6. Results for image Shaded Text. (a) Original Image. (b) Result of Otsu‟s method. (c) Result of 3  3 Laplace filter. (d) Result of 5  5 Laplace
filter. (e) Result of 3  3 NSCF. (f) Result of 5  5 NSCF. (g) Result of 5  5 Laplace filter to image smoothed by 3  3 averaging filter (h) Result of
5  5 NSCF to image smoothed by 3  3 averaging filter.

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
 The Image Segmentation Techniques 13

IV. SIMULATIONS AND RESULTS edges also. Overall, it has produced a natural boundary.
The image Shaded text is used to check viability of NSCF
NSCF is applied to various images for checking its
for noisy case. Fig.6(b) shows the result of Otsu‟s method
correctness. We use objective assessment as well as
which has clearly defeated its purpose. For the case of
quantitative assessment for the analysis of algorithm.
Laplace filter, fig.6(c) and fig.6(d) show that although
A. Objective assessment method is able to threshold the image, the shaded strips
are still appearing giving an unpleasant view. Further, the
For image Bubbles, the purpose is to extract the intensities of strips are enhancing with filter size, making
boundaries of objects. Comparison of fig.5(c) and fig. 5(e)
the method unsuitable for noisy case. Fig.6(e) and fig.6(f)
shows that both Laplace filter and NSCF are able to are for NSCF. It is evident that algorithm has naturally
extract the boundary of image but Laplace filter has segmented the image by successfully combating the
distorted the shape of the boundary and has produced
shaded strips and there is no evidence of shaded strips in
thick edges. It is also unable to extract the fine edges and
the results. Moreover, the results of NSCF are improving
has eliminated small objects. NSCF, on other hand, has with increased neighborhood size as already predicted.
produced undistorted image and is able to extract fine Fig.6(g) shows the result of Laplace filter after averaging

(a) (b) (c) (d)

Fig.7. Results for image Spot shaded text (a) Original image. (b) Results of 5  5 Laplace filter. (c) Result of 5  5 NSCF. (d) Result of Otsu‟s
method.

(a) (b) (c) (d)

Fig.8. Results for image Noisy septagon (a) Original image. (b) Results of 5  5 Laplace filter. (c) Result of 5  5 NSCF. (d) Result of Otsu‟s method.

(a) (b) (c) (d) (e)

Fig.9. Results for image Thumb (a) Original image. (b) Results of 5  5 Laplace filter. (c) Zoomed version of (b). (d) Result of 5  5 NSCF. (e)
Zoomed version of (d).

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
14 The Image Segmentation Techniques

(a) (b) (c) (d) (e)

Fig.10. Results for image Pattern (a) Original image. (b) Results of 5  5 Laplace filter. (c) Zoomed version of (b). (d) Result of 5  5 NSCF. (e)
Zoomed version of (d).

the original image. Same results are shown for NSCF by

fig.6(h). Laplace filter still has no effect on shaded strips Table 1. Calculation of REF for different images
while NSCF is producing even better result, a fact already
Algorithm
proved in section II. Image
For image Spot shaded text in fig.7(a), the Laplace NSCF Laplace filter
filter has produced ringing effect in background which is Bubbles .067 .192
again an unnatural effect. Also, this effect will dominate Shaded text .256 .470
with increased mask size. Otsu‟s method has again failed
for this case also and has produced a completely Spot shaded text .219 .459
unnatural image. Coming to NSCF, it has again Noisy septagon .479 .502
successfully segmented the image and has produced a Thumb .249 .516
uniform intensity image, leaving no clue of shaded spot
in resultant image. The results can be further improved by Pattern .054 .178
either increasing the neighborhood size or averaging the
original image before applying NSCF. From table.1. it can be concluded that the value of REF
Image Noisy septagon in fig.8(a) has non uniform for every image is less in case of NSCF. As verified
illumination with illumination increasing to right. The earlier in section III, a better algorithm should have
aim is to extract the septagon from background. Laplace minimum value of REF and hence quantitative
filter failed completely for this case and is able to extract assessment also proves that NSCF is way better than
only the boundary of septagon which is obviously not our Laplace filter. Table 2 summarizes the execution time for
sole motive. Otsu‟s method is able to threshold the image both the algorithms. Execution time of NSCF differs from
but with unnatural effect and clearly it is victim of non that of Laplace filter by a fraction of seconds which is
uniform intensity which is still appearing in the mapped negligible considering superior performance of NSCF
image. NSCF has produced a fine image and has over the later.
extracted septagon from background by assigning
different intensities to foreground and background. Also, Table 2. Execution time (in sec.) for different images
no sign of non uniform intensity exist in the results. In Algorithm
Image
effect, non uniform intensity has no effect on the NSCF Laplace Filter
performance of NSCF.
For image Thumb in fig.9(a), Laplace filter has Bubbles 3.36 3.16
produced a noisy image with no clear extraction of details Shaded text 4.56 4.38
while NSCF has generated an image which is noiseless Spot shaded text 4.68 4.35
and from which fine details can be easily extracted. For 4.80 4.42
Noisy septagon
image Pattern in fig 10(a), Laplace transform has
produced thick edges with distorted boundary. NSCF for Thumb 7.30 6.84
the same case has generated an image with perfect Pattern 2.69 2.32
boundaries and thin edges.
B. Quantitative assessment
V. PIECEWISE IMAGE SEGMENTATION
For quantitative assessment we use the relative error
function (REF) defined in section III. We also calculate In this section, we propose a new algorithm that is
the execution time of the algorithm. Both quantities are useful for generating a controlled edge map of an image
calculated for each image with a neighborhood of 3  3 using a recursive approach. Traditional sharpening filter
and no prior averaging. Results are depicted in table 1 produces a relevant edge map in most of the cases but it
and table.2. has no provision for controlling the number of edges
appearing in the edge map. Sometimes the requirement
may be a partial edge map containing only the important

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
 The Image Segmentation Techniques 15

edges like boundary pixels and void of all the minute controlling nature in the algorithm. For different values of
edges. In such cases, traditional filters may not be a T a unique edge map is generated. Now, problem is to
useful tools and we need to discover new methods. The select the initial value and final value of the T .
proposed algorithm comes handy in such applications and
presents a better alternative to the traditional filters. The As an initial step, we find the mean of  L obtained
algorithm operates recursively and can also produce using (13).
complete edge map apart from partial edge map. In the
remaining section develop the proposed algorithm and its 1
methodology along with simulation results. m     L ( x, y)
( M  N ) x  M  y  N 
(17)

A. Methodology
Consider an image I of size M  N and an extension Next, we find variance of L :
of the averaging filter f mentioned in fig.4 to size
n  n, where n is an odd number. Using this filter we can 1
   
( M  N ) x  M  y  N 
( L ( x, y)  m )2 (18)
calculate local mean at a pixel I ( x, y) by the equation:

a a We use  as the initial value of T to start the

mL ( x, y)    I ( x  k , y  l ) f (k , l )
k  a l  a
(12) iteration. Now, we find the maximum deviation between
local variance and global variance i.e.  L and  G :
and for odd n, a  b  (n  1) / 2. If we apply this
operation at each pixel, then we get an image mL having
  max{ L   G } (19)

local mean as the pixels. Eq. (12) can be exploited to

We use this deviation (  ) as the final value of T in
calculate the local variance at each pixel I ( x, y ) using:
the iterations. The remaining of the algorithm will
proceed as follows:
a a
1
 L ( x, y ) 
nn
  [ I ( x  k , y  l )  m ( x  k , y  l )]
L
2

1) With   as initial value of T and  as final value

k  a l  a

(13) of T , start the iterations. Decide a predefined step

Similarly we can calculate the global mean as: size (  ) and choose T   as the next threshold
value. Repeat the following steps for each value of
1 T till  is reached.
mG    I ( x, y)
( M  N ) x  M  y  N 
(14) 2) If an pixel I ( x, y ) satisfies (16) then replace
I ( x, y) with output of Laplace filter ( w ) of
and global variance as: predefined size (fig.4(b)) and centered at I ( x, y ) .

1
  ( I ( x, y)  mG )2
a b
G  (15) I ( x, y )    I ( x  k , y  l )w(k , l ) (20)
( M  N ) x  M  y  N  k  a l  b

Eq.(13) and (15) are basis for further development of 3) If an pixel I ( x, y) fails (16) then
the algorithm. Conceptually a pixel I ( x, y ) will belong
to an edge if I ( x, y)  0 (21)

T  ( L ( x, y)   G )  T (16) 4) Find global mean of the magnitude of image

'
obtained after applying (2) and (3), say mk and
i.e local variance at the pixel lies within a particular range
of global variance. This is an obvious condition because also calculate
an edge pixel is identified by sudden change of intensity
in its neighborhood and it will be reflected in its local   mk'  mk' 1 (22)
variance which will be higher than global variance of the
image.
5) Find the value of T , say Tmax , for which  is
Threshold T is used to check the output of the
algorithm and it is also responsible for generating partial maximum. This is the threshold value for which
major changes in edge map will occur. All the
or complete edge map. In other words T induces a

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
16 The Image Segmentation Techniques

other values of T including Tmax and previous to our preferred tool. The results of fig.12(c) were obtained
by applying presented algorithm with threshold
Tmax will produce a partial edge map. As we will Tmax which has successfully extracted the targeted region.
move away from Tmax in opposite direction, a For image Coins in fig.13, the aim is again to extract
lighter edge map will be obtained. the boundaries of the objects. Laplace filter as usual
(fig.13(b)) has produced a complete edge map. For this
The preferred value of T is particular image, our algorithm with threshold
Tmax   has extracted the boundaries of the coins but it
T  Tmax   (23) is broken at some places. The threshold Tmax would have
produced an unbroken boundary but it had also generated
This value will produce an edge map containing more inner edges. Nevertheless, for such cases we can
boundary pixels only and it can be exploited for boundary apply dilation to complete the boundary rather than
extraction of the object. increasing the threshold.
B. Objective assessment For image Moon in fig.14, the Laplace filter has again
produced a complete edge map. The application of our
algorithm with threshold Tmax   is able to extract the
The algorithm is tested on several images for
verification of its performance. The results are obtained
using filter of size 3  3. The results are summarized boundary but result (fig.14(c)) contains inner edges as
below: well. Here, either we can further reduce the threshold
For image Tower in fig.11, the Laplace filter has value to Tmax  2 or we can apply erosion operation to
produced a complete edge map (fig.11(b) and results are get rid of minute inner edges.
fair in this sense but it‟s not possible to extract boundary For image Cameraman in fig.15, the expectations are
of the object using the same. The presented algorithm same i.e. boundary extraction. The application of our
was applied with threshold Tmax   and it is successfully algorithm has completely eliminated the background
able to extract the boundary without concentrating on fine leaving only the boundary of main object in the image
(fig.15(c)). If we try to recover the boundaries of
edges. If we further increase the threshold beyond Tmax , background objects also by increasing the threshold value,
we can replicate the results of Laplace filter. then we may introduce unnecessary fine edges as in case
For image Kidney in fig.12, the complete edge map has of Laplace filter.
lots of information but sometime the requirement may be
only a particulate region of complete object like main
vessel in this image. For such a case, Laplace filter is not

(a) (b) (c)

Fig.11. Results for image Tower. (a) Original image. (b) Result of 3  3 Laplace filter. (c) Result of proposed algorithm with Tmax  1 threshold.

(a) (b) (c)

Fig.12. Results for image Kidney. (a) Original image. (b) Result of 3  3 Laplace filter. (c) Result of proposed algorithm with Tmax threshold.

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
 The Image Segmentation Techniques 17

(a) (b) (c)

Fig.13. Results for image Coins. (a) Original image. (b) Result of 3  3 Laplace filter. (c) Result of proposed algorithm wi th Tmax  1 threshold.

(a) (b) (c)

Fig.14. Results for image Moon. (a) Original image. (b) Result of 3  3 Laplace filter. (c) Result of proposed algorithm with Tmax  1 threshold

(a) (b) (c)

Fig.15. Results for image Cameraman. (a) Original image. (b) Result of 3  3 Laplace filter. (c) Result of proposed algorithm with Tmax  1 threshold.

VI. CONCLUSION REFERENCES

This paper proposes two algorithms for image [1] R.C. Gonzalez, R.E. Woods, and S.L. Eddins, Digital
segmentation. Simulation results demonstrated that NSCF Image Processing. Upper Saddle River, NJ, USA:
based algorithm performs well even in case of noisy or Prentice-Hall, 2004.
non-uniformly illuminated images. The REF also [2] A. Bovik, Handbook of Image and Video Processing.
New York, NY, USA: Academic, 2000.
concludes that NSCF based algorithm is better than [3] H.K. Sawant and M. Deore, “A comprehensive review of
classic algorithm like Laplace filter even in qualitative image enhancement techniques,” Int. J. Comput. Technol.
sense. Moreover, this algorithm maintains its simplicity Electron. Eng. Vol. 1, pp. 39-44, Mar. 2010.
and has very less computational complexity. We also [4] J.C. Caicedo, A. Kapoor, and S.B. Kang, “Collaborative
propose PIS as edge-based method which is successfully personalization of image enhancement,” in Proc. IEEE
used in boundary extraction. In can also produce Conf. Comput. Vis. Pattern Recognit.,Jun. 2011, pp. 249-
complete edge map like any other algorithm in this 256.
domain. However, we have not performed qualitative [5] Shuhang Wang, Jin Zheng, Hai-Miao Hu and Bo Li,
assessment of this algorithm. “Naturalness Preserved Enhancement Algorithm for Non-
Uniform Illumination Images,” IEEE trans. on Image
Processing, Vol. 22, No. 9, 3538-2548, Sep 2013.

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18
18 The Image Segmentation Techniques

[6] N. R. Pal and S. K. Pal, “A review on image segmentation polynomials,” IEEE Signal Process. Lett., vol. 22, no. 3,
techniques,” Patt. Recognit., vol. 26, no. 9, pp. 1277–1294, pp. 298–302, Mar. 2015.
1993. [19] D. Mumford and J. Shah, “Optimal approximations by
[7] Anil K. Jain, Digital Image Processing. piecewise smooth functions and associated variational
[8] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active problems,” Commun. Pure Appl. Math., vol. 42, no. 5, pp.
contour models,” Int. J. Comput. Vis., vol. 1, no. 4, pp. 577–685, 1989.
321–331, 1988. [20] T. F. Chan and L. Vese, “Active contours without edges,”
[9] S. Osher and J. A. Sethian, “Fronts propagating with IEEE Trans.Image Process., vol. 10, no. 2, pp. 266–277,
curvature dependent speed: Algorithms based on 2001.
hamilton-jacobi formulations,” J. Comput. Phys., vol. 79, [21] Ashish Thakur and Radhey Shyam Anand, “A local
no. 1, pp. 12–49, 1988. statistics based region growing segmentation method for
[10] V. Caselles, F. Catt, T. Coll, and F. Dibos, “A geometric ultrasound medical images,”Int. J. of Signal Process., Vol.
model foractive contours in image processing,” Numer. 1, No. 2, pp 141-146, 2004.
Math., vol. 66, no. 1, pp.1–31, 1993.
[11] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active
contours,” Int. J. Comput. Vis., vol. 22, no. 1, pp. 61–79,
1997. Authors’ Profiles
[12] R. Malladi, J. Sethian, and B. Vemuri, “Shape modeling
with front propagation: A level set approach,” IEEE Trans.
Shiv Gehlot received the B.Tech. degree in
Patt. Anal. Mach. Intell., vol. 17, no. 2, pp. 158–175, Feb.
Electronics & Communication engineering
1995.
from Uttar Pradesh Technical University in
[13] X. Yang, X. Gao, J. Li, and B. Han, “A shape-initialized
2011. He completed his M.Tech. degree in
and intensity adaptive level set method for auroral oval
Signal Processing at Netaji Subash Institute
segmentation,” Inf. Sci., vol.277, pp. 794–807, 2014.
of Technology in 2013. His present
[14] X. Yang, X. Gao, D. Tao, and X. Li, “Improving level set
research areas include Statistical Signal
method for fast auroral oval segmentation,” IEEE Trans.
Processing and Digital Image Processing. Presently, he is
Image Process., vol. 23, no. 7, pp. 2854–2865, 2014.
working as an Assistant Professor in the department of
[15] X.Yang, X.Gao ,D.Tao, X.Li,and J.Li,“An efficient MRF
Electronics &Communication Engineering at Noida Institute of
embedded level set method for image segmentation,”
Engineering & Technology, Greater Noida.
IEEE Trans.Image Process.vol. 24, no. 1, pp. 9–21, 2015.
[16] C. Li, R. Huang, Z. Ding, J. Gatenby, D. N. Metaxas, and
J. C. Gore,“A level set method for image segmentation in
John Deva Kumar completed his M.Tech.
the presence of intensity inhomogeneities with application
degree in Signal Processing at Netaji
to mri,” IEEE Trans. Image Process., vol. 20, no. 7, pp.
Subash Institute of Technology in 2014.
2007–2016, 2011.
Presently he is working as a Lecturer at
[17] Y. Wang, L. Liu, H. Zhang, Z. Cao, and S. Lu, “Image
PUSA institute of Technology. His present
segmentation using active contours with normally biased
research areas include microcontroller
GVF external force,” IEEE Signal Process. Lett., vol. 17,
based system design, OFDM systems
no. 10, pp. 875–878, 2010.
Signal processing and image processing.
[18] S. Mukherjee and S. Acton, “Region based segmentation
in presence of intensity inhomogeneity using legendre

How to cite this paper: Shiv Gehlot, John Deva Kumar,"The Image Segmentation Techniques", International Journal
of Image, Graphics and Signal Processing(IJIGSP), Vol.9, No.2, pp.9-18, 2017.DOI: 10.5815/ijigsp.2017.02.02

Copyright © 2017 MECS I.J. Image, Graphics and Signal Processing, 2017, 2, 9-18