Accepted Manuscript

A Novel Color Image Multilevel thresholding based Segmentation

using Nature Inspired Optimization Algorithms

A.K. Bhandari , A. Kumar , S. Chaudhary , G.K. Singh

PII: S0957-4174(16)30331-1
DOI: 10.1016/j.eswa.2016.06.044
Reference: ESWA 10746

To appear in: Expert Systems With Applications

Received date: 27 December 2015

Revised date: 26 June 2016
Accepted date: 26 June 2016

Please cite this article as: A.K. Bhandari , A. Kumar , S. Chaudhary , G.K. Singh , A Novel Color
Image Multilevel thresholding based Segmentation using Nature Inspired Optimization Algorithms,
Expert Systems With Applications (2016), doi: 10.1016/j.eswa.2016.06.044

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights
 This paper introduces the comparative performance of different objective functions (Kapur’s & Otsu).
 Evolutionary algorithm based multilevel thresholding for a color satellite image has been presented.
 DE, WDO, PSO and CS algorithms are exploited with Kapur’s and Otsu method.
 CS based Kapur’s entropy was found to be more accurate for colored satellite image segmentation.

T
IP
CR
US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

A Novel Color Image Multilevel thresholding based Segmentation

using Nature Inspired Optimization Algorithms
A. K. Bhandari1, A. Kumar2, S. Chaudhary2 and G. K. Singh3
1
Department of Electronics & Communication Engineering, National Institute of Technology Patna, Bihar, (800005), India
2
PDPM Indian Institute of Information Technology Design and Manufacturing Jabalpur -482011, MP (India)
3
Department of Electrical Engineering, Indian Institute Technology Roorkee, Uttrakhand 247667, India
Email: [email protected], [email protected], [email protected], and [email protected]

T
IP
Abstract
Multilevel thresholding for segmentation is an essential task and indispensable process in various applications.

CR
Conventional color multilevel thresholding based image segmentations are computationally expensive, and lack
accuracy and stability. To address this issue, this paper introduces the comparative performance study of different
objective functions using cuckoo search and other optimization algorithms to solve the color image segmentation

US
problem via multilevel thresholding. During the optimization process, solutions are evaluated using Otsu or Kapur’s
method. Performance of the proposed approach has been assessed using a variety of benchmark images, and
compared against three other nature inspired algorithms namely differential evolution (DE), wind driven optimization
AN
(WDO) and particle swam optimization (PSO) algorithms. Results have been analyzed both qualitatively and
quantitatively, based on the fitness values of obtained best solutions and four popular performance measures namely
PSNR, MSE, SSIM and FSIM indices as well. According to statistical analysis of different nature inspired
M

optimization algorithms, Kapur’s entropy was found to be more accurate and robust for multilevel colored satellite
image segmentation problem. On the other hand, cuckoo search was found to be most promising for colored satellite
image segmentation.
ED

Keywords: Color image segmentation, Multilevel thresholding, Nature inspired optimization algorithms, Cuckoo
search algorithm, Kapur’s and Otsu method.
PT

1. Introduction
Multilevel thresholding for colored image segmentation is frequently used method among all the approaches.
CE

However, the conventional multilevel colored image thresholding methods are time consuming as they exhaustively
finds the optimal threshold to optimize certain fitness functions. Segmentation is one of the furthermost essential
tasks in image processing applications. It is the process of categorizing the image pixels into different regions on the
AC

basis of their intensity information and a threshold value. Now a day’s multilevel thresholding is a powerful and
famous tool, and it is extensively used in image processing areas (Agrawal, Panda, Bhuyan, & Panigrahi, 2014;
Bhandari, Kumar, & Singh, 2015). Basically, thresholding approaches can be classified into bi-level and multilevel
thresholding class, based upon the number of image fragments. In an image, if the object is distinguished from the
background by computing one threshold value, it is known as bi-level thresholding method (Kumar, Kumar, Sharma,
& Pant, 2013); while, classifying the image into several different areas according to color by locating numerous
threshold values, then it is called as multilevel colored image thresholding (Kurban, Civicioglu, Kurban, & Besdok,

2
 ACCEPTED MANUSCRIPT

2014). In case of bi-level thresholding approach, image is mainly separated into two distinct classes. In this concept,
pixel with gray level values higher than a certain value T are categorized as object of the image and rest gray level
values, which are lesser than the threshold criteria T are categorized as background image (Arora, Acharya, Verma, &
Panigrahi, 2008). However, in case of remote sensing images or real life images, bi-level thresholding does not give
appropriate performance. As a result, there is strong requirement of multilevel thresholding.
In past few decades, numerous kind of statistical properties have been introduced for choosing a threshold on
the basis of image histogram such as moment based approach (Tsai, 1985), entropy based approach (Kapur, Sahoo, &

T
Wong, 1985) and between-class variance based technique (Otsu, 1979). Amongst all the remarkable thresholding
methods, entropy-based image segmentation schemes have drawn the attention of numerous researchers. In 1988, an

IP
effective survey on thresholding based techniques has been presented by Sahoo et al. ( Sahoo, Soltani, Wong, &
Chen, 1988). In case of global thresholding, Sahoo et al. (1988) (Sahoo, Soltani, Wong, & Chen, 1988) found that

CR
Otsu’s technique performs very efficiently especially for general real world images. However, inefficient formulation
of Otsu’s technique leads to exhaustive and high processing time. In 2004, Sezgin and Sankur (Sezgin, & Sankur,

US
2004) presented a review of image thresholding methods and their performance analysis. In 2013, Akay (Akay, 2013)
has reported a study for multilevel thresholding using optimization algorithm. In 2014, Bhandari et al. (Bhandari,
Singh, Kumar, & Singh, 2014a) presented evolutionary algorithm based study of satellite images for multilevel
AN
thresholding. These all methods were initially introduced for the bi-level image thresholding process, and then
explored to compute the multilevel (Fan, Ouyang, Zhang, & Xiao, 2014; Sathya, & Kayalvizhi, 2011) and
furthermore extended for colored image multilevel thresholding problems (Kurban, Civicioglu, Kurban, & Besdok,
M

2014).
The Otsu method (Otsu, 1979) is based on the disjoint investigation to maximize division level of the groups in
gray scale image (Akay, 2013; Liu, Mu, Kou, & Liu, 2014; Huang, & Wang, 2009). It chooses finest threshold by
ED

between-class variance maximization process. On the other hand, Kapur’s entropy approach depends upon the
maximization of entropy through probability scattering of gray level histogram (Bhandari, Singh, Kumar, & Singh,
2014a; Manikandan, Ramar, Willjuice Iruthayarajan, & Srinivasagan, 2014). A variety of segmentation using
PT

Kapur’s entropy and Otsu method has appeared in the literature for determination of optimum threshold for
segmentation problems (Sri Madhava Raja, Rajinikanth, & Latha, 2014; Oliva, Cuevas, Pajares, Zaldivar, & Perez-
CE

Cisneros, 2013). Over the years, researchers have proposed many methods to improve standard Otsu’s method. In
2014, a new iterative scheme, based on Otsu’s method, has been presented which iteratively searches for subregions
of the image for segmentation, instead of treating the full image as a whole region for processing (Cai, Yang, Cao,
AC

Xia, & Xu, 2014). In favor of Kapur’s and Otsu methods, recently, a modified version of both the methods is
exploited for cloud model-based thresholding algorithm, which determines the binary threshold within the
constrained gray level range (Wu, Xiao, Qin, & Chen, 2014).
In general, Otsu approach has been found to be the finest thresholding based segmentation methods for
homogeneity and shape detections (Otsu, 1979; Sezgin, & Sankur, 2004). Though, due to exhaustive searching
policy, the effect of computation cost increases exponentially when the thresholding number is extended. This is the
main disadvantage of the multilevel thresholding applications. Kapur’s entropy scheme is used to give better

3
 ACCEPTED MANUSCRIPT

performance than the other approaches on the test images which constantly works on the basis of positive
probabilities and a global maximum rule for the entropy. In 2004, the researcher Sezgin and Sankur (Sezgin, &
Sankur, 2004) have reported the superior performance of Kapur’s entropy that produces superior average results in
comparison with the existing entropy-based calculations on nondestructive sample images. In 2013, Agrawal et al.
(Agrawal, Panda, Bhuyan, & Panigrahi, 2014) have presented a cuckoo search algorithm based optimal multilevel
thresholding by maximizing the Tsallis entropy (Tsai, 1985). A wide range of multilevel thresholding based
segmentation methods via numerous models have been reported in literature (Hammouche, Diaf, & Siarry, 2010; Yin,

T
2014). In 2006, Chang et al. (Chang, Du, Wang, Guo, & Thouin, 2006) have reported a review and comparative
performance survey of entropy and similar thresholding based entropy schemes.

IP
The exploitation of meta-heuristic computing algorithms has been very successful throughout the last few eras.

CR
To achieve optimum multilevel threshold, many heuristic optimization techniques have been applied for solving the
multilevel image segmentation tasks. Over the years, in literature, numerous works, based on swarm based systems
such as genetic algorithm (GA) (Zhang, 2014 Tang, Yuan, Sun, Yang, & Gao, 2011), differential evolution (DE)

US
(Storn, & Price, 1997; Cuevas, Zaldivar, & Pérez-Cisneros, 2010), ant colony optimization (ACO) (Tao, Jin, & Liu,
2007), bacterial foraging optimization (BFO) (Sathya, & Kayalvizhi, 2011a; Sathya, & Kayalvizhi, 2011b; Bakhshali,
& Shamsi, 2014), harmony search algorithm (HSA) (Oliva, Cuevas, Pajares, Zaldivar, & Perez-Cisneros, 2013),
AN
honey bee mating optimization (HBMO) (Horng, 2010a; Horng, 2010b) firefly algorithm (Sri Madhava Raja,
Rajinikanth, & Latha, 2014; Horng, & Liou, 2011), artificial bee colony (ABC) (Kumar, Kumar, Sharma, & Pant,
2013; Akay, 2013; Karaboga; 2005; Bhandari, Soni, Kumar, & Singh, 2014b; Soni, Bhandari, Kumar, & Singh,
M

2014a), PSO (Poli, Kennedy, Blackwell, 2007), WDO (Bhandari, Singh, Kumar, & Singh, 2014a; Bayraktar, Turpin,
& Werner, 2011) and CS (Bhandari, Singh, Kumar, & Singh, 2014a; Yang, & Deb, 2009; Bhandari, Soni, Kumar, &
Singh, 2014c), have been reported to tackle many multilevel image segmentation problems for determination of
ED

optimum threshold. Inspired by their effective applications, furthermore this paper examines the possibility of using
swarm or meta-heuristic methods for dealing color satellite image segmentation problem via different multilevel
thresholding approach.
PT

The evolutionary (Bhandari, Kumar, Singh & Soni, 2016b; Pare, Bhandari, Kumar, Singh, & Khare, 2015) and
swarm-based computing approaches (Bhandari, Kumar, Kumar, & Singh, 2016) stand out for their ability to search
CE

best solution from any objective function. In 2015, a new technique for color image segmentation using CS algorithm
supported by Tsallis entropy for multilevel thresholding has been proposed toward the effective colored segmentation
of satellite images (Bhandari, Kumar, & Singh, 2015). An Artificial bees for multilevel thresholding of iris images
AC

has been also presented (Bouaziz, Draa, & Chikhi, 2015). In 2015, to overcome the common drawbacks of thresholding
methods when images are corrupted with artifacts and noise, a new multiregion thresholding methodology is
presented (Aja-Fernández, Curiale, & Vegas-Sánchez-Ferrero, 2015). In 2015, the multilevel thresholding problem
has been solved using GA, PSO and hybrid optimization techniques for real-time imaging applications (Mala, &
Sridevi, 2015). Once again towards the advancement in the fast gray image segmentation via multilevel thresholding,
a membrane computing is introduced to propose an efficient and robust multilevel thresholding method (Peng, Wang,
& Pérez-Jiménez, 2015).

4
 ACCEPTED MANUSCRIPT

In 2016, a multilevel color image segmentation technique based on cuckoo search algorithm and energy curve
has been proposed (Pare, Kumar, Bajaj, & Singh, 2016). Apart from satellite image segmentation, a cooperative semi-
supervised fuzzy clustering framework has been proposed for dental X-ray image segmentation (Son, Tuan, 2016) in
the field of medical. In 2016, towards the Hyper-spectral image segmentation, a new approach of unsupervised
classification using Rényi entropy with differential evolution has been presented for multilevel thresholding (Sarkar,
Das, & Chaudhuri, 2016). In favor of fast satellite image processing, an efficient cuckoo search algorithm based
multilevel thresholding for segmentation of satellite images using different objective functions has been given

T
(Suresh, & Lal, 2016). Towards the gray scale image segmentation using Otsu criterion, a beta differential evolution
(BDE) algorithm has been proposed for determining the n-1 optimal n-level threshold on a given image (Ayala, dos

IP
Santos, Mariani, & dos Santos Coelho, 2015). A performance study has been presented for two new optimization
algorithm named social spiders optimization and flower pollination algorithm for gray level multilevel image

CR
thresholding (Ouadfel, & Taleb-Ahmed, 2016). In 2016, towards efficient segmentation scheme, a fast scheme for
multilevel thresholding, based on a modified bees algorithm, has been proposed (Hussein, Sahran, & Abdullah,

US
2016). In 2016, to solve the computational cost problem, an advanced version of gravitational search algorithm
(GSA), namely hybrid algorithm of GSA with genetic algorithm (GA) (GSA-GA) has been presented for fast
multilevel thresholding (Sun, Zhang, Yao, & Wang, 2016). Recently, in favor of color image segmentation, a new
AN
quantum inspired meta-heuristic techniques for multilevel colour image thresholding has been proposed (Dey,
Bhattacharyya, & Maulik, 2016).
In the literature, there have been several methods reported for simple image segmentation usually for gray
M

scale images by the histogram based multilevel thresholding process. However, they are computationally time
consuming and their effectiveness is reduced when applied to a complex image especially satellite images and when
the number of different regions composing this image is high. Remote sensing or satellite images (Bhandari, Kumar,
ED

& Singh, 2015b; Ashish, Kumar, Padhy, 2012; Bhandari, Kumar, & Singh, 2012b) are an essential cause of
understanding of different regions, which are exploited in several environmental valuation such as climate survey,
forest resources monitoring, inspecting marine environment, agriculture, metrology, mapping, military, etc (Bhandari,
PT

Kumar, & Singh, 2012a; Bhandari, Gadde, Kumar, & Singh, 2012; Bhandari, Kumar, Chaudhary, & Singh, 2015a).
Therefore, this paper focusses on the color image multilevel thresholding where the segmented color images expose
CE

better features in comparison with the previous grayscale segmented images. It can be noticed that the segmented
color images are more informative and their fidelity parameters are better than the results claimed in previous papers
such as (Bhandari, Kumar, & Singh, 2015). Although the computational complexity is increased when 3 band or 3
AC

channels RGB image are processed for segmentation. Hence, determination of such a widely varying regions or
clusters presents a challenging task in formulation of segmentation algorithms (Bhandari, Kumar, & Singh, 2015b;
Kumar, Bhandari, & Padhy, 2012). Therefore, motivated by the wide range of application, this paper has proposed a
novel approach for colored satellite image segmentation using cuckoo search and other optimization algorithm based
on multilevel thresholding process.
This paper has significant application in the field of expert and intelligent systems such as machine vision,
medical imagine, object detection, traffic control systems, video surveillance, image analysis, understanding pattern

5
 ACCEPTED MANUSCRIPT

recognition, computer vision field. So far, image segmentation technique has been successfully applied in various
domains of expert and intelligent systems applications. Some of them may include face detection, image retrieval and
object recognition to name a few.
The endeavor of this paper is concentrated on color image multilevel thresholding using Kapur’s entropy and
between-class variance criterions. The reminder of the paper is structured as follows: Section 2 presents the color
image thresholding problem formulation. In Section 3, methodology of swarm intelligence such as DE, PSO, WDO
and CS is discussed. Section 4 compares the visual, qualitative and quantitative outputs of the techniques used that

T
are carried by PSNR, MSE, SSIM and FSIM, with brief explanation of the results. Lastly, conclusions are drawn in
Section 5.

IP
2. Problem Formulation for Color Image Segmentation

CR
Thresholding is a method of classifying the pixels belonging to gray scale image or colored image into classes
or sets on the basis of their intensity level (L). For appropriate selection of classes in an image, the optimal threshold
value (th) should be selected in a manner that follows Eq. (1).
C1  p
C2  p
if
if
0  p  th
th  p  L  1 US (1)
AN
where, p indicates one of the m x n pixels that belongs to gray scale image Ig, which can be denoted in L gray scale
levels L={0, 1, 2, …, L-1}. C1 and C2 represent the two different classes, which are separated using pixel p, whereas
th indicates the threshold.
The concept of bi-level thresholding technique can be simply extended for multilevel thresholding, which uses
M

multiple threshold values and generates an output image through multiple groups using Eq. (2):
C1  p if 0  p  th1
ED

C2  p if th1  p  th2
(2)
Ci  p if thi  p  thi 1
Cn  p if thn  p  L  1
PT

where, {th1 th2 … thi thi+1 thk} are the various thresholds. The aim is to choose the th values that appropriately classify
the different region in an image using bi-level or multilevel thresholding method. The two most widely used objective
functions Otsu's (Otsu, 1979) and Kapur's (Kapur, Sahoo, & Wong, 1985) have been considered to calculate the
CE

multiple threshold values. These two methods that present different criterion, which needs to be maximized in order
to obtain optimal th value, are discussed in detail below:
AC

2.1 Between-class variance (Otsu's method)

Otsu (Otsu, 1979) proposed a new approach towards thresholding based on between-class variance. It is a
nonparametric segmentation technique that separates the entire image into classes in such a way that maximum
variance can be achieved for different classes (Oliva, Cuevas, Pajares, Zaldivar, & Osuna, 2014). Let L be the
intensity levels from a gray scale image or from separate channels that are red, green or blue of RGB image; then by
considering simple rules of histogram of an image, probability distribution of the intensity values can be obtained.
The normalized histogram for a given image can be defined by Eq. (3):

6
 ACCEPTED MANUSCRIPT

1, 2,3 if Multispectral or RGB image

NP
hic
Phic 
NP
, i  0  i  L 1  Phi 1
c
i  1, c
1 if Gray sacle image
(3)

where, i corresponds to a particular intensity level between the range 0 to L-1, NP represents the total number of
pixels in an image. c indicates the image component that varies according to single channel (gray scale image) or

multiband channel (RGB image), while number of pixels are represented by hic (Histogram), which corresponds to i

intensity level in c. Probability distribution, Phic normalizes the image histogram. The simplest bi-level thresholding

T
that classifies the image into two classes are defined as:

Ph1c Ph0c Phthc 1 PhLc

IP
C1  ,..., and C2  ,..., (4)
0c (th) 0c (th) 1c (th) 1c (th)

CR
where, ω0(th) and ω1(th) represents the cumulative probability distributions for C1 and C2, which is specified in Eq.
(5).
th L
0c (th)   Ph , c
1c (th)   Ph c
(5)

US
1 i
i 1 th 1

It is necessary to calculate the mean intensity levels  c0 and 1c from the input image. The mean level of each class for
AN
the bi-level thresholding can be computed by using Eq. (6).
th L
iPhic iPhic
0c    (th)
i 1
c
0
, 1c    (th)
i th 1
c
1
(6)
M

c
When these values are computed, the between-class variance  B2 is obtained using Eq. (7).
c
 B2   1c   2c (7)
ED

It is noted that in above two Eqs. (6) and (7), c depends on the nature of image (Gray or RGB). Furthermore,  1c and

 2c in Eq. (7) indicates the variances of C1 and C2 that are expressed as

PT

1c  0c (0c  Tc )2 ,  2c  1c (1c  Tc )2 (8)

where, Tc is mean intensity of the input image defined as Tc  0c 0c  1c 1c and 0c +1c =1. On the basis of above
CE

Eq. (8), the objective function can be defined using Eq. (9).
c
fotsu ( th )  max{ B2 ( th )}, 0  th  L  1 (9)
AC

c
where,  B2 (th) represents the Otsu’s variance for a given value of th. Thus, with an aim to gain the intensity level (th),
which maximizes Eq. (9), the optimization problem is reduced.
The above between-class variance based bi-level thresholding can be easily extended for multilevel thresholding of
image. Original image can be easily divided into k number of classes by Eq. (2) with k threshold values obtained.
Afterward it is compulsory to calculate the k variances and irrespective elements. Therefore, the between-class
variance fOtsu(th) considered as an objective function can be extended for multilevel thresholds using Eq. (10):
c
fotsu (TH)  max{ B2 (TH)}, 0  thi  L  1, i  1,2,...,k (10)

7
 ACCEPTED MANUSCRIPT

here, TH=[th1, th2, …, thk-1] represents a vector having multiple values of variance, and thresholds are determined
using Eq. (11):
k k

   (
c
 B2  c
i
c
i
c
i  Tc )2 (11)
i 1 i 1

where, i gives indication of specific class. Correspondingly, ic and ic are the mean of a class and probability of
occurrence, respectively. In case of multilevel thresholding, these values are computed as following:
th1 th2 L

T
0c (th)  i 1
Phic , 1c (th)  
i th1 1
Phic , , kc1 (th)   Ph
i th1 1
c
i (12)

IP
and the mean values are expressed as:

CR
th1 th2 L
iPhic iPhic iPhic
0c    (th ) , 
i 1
c
0 1
c
1    (th ) ,
i th1 1
c
0 2
, kc1  
i thk 1 1 (thk )
c
(13)

Similarly, in case of MT using Otsu’s method, c=1, 2, 3, where, c represents the image channels (RGB), and for gray
scale image c=1.

2.2 Entropy criterion method (Kapur's method)

US
AN
In multilevel thresholding problem, the main aim is to obtain the optimum threshold value in order to segment
different classes so that it can meet the expected features (Oliva, Cuevas, Pajares, Zaldivar, & Osuna, 2014). Let an
image I is chosen for the thresholding purpose which has N pixels and L units of gray levels. In the segmentation
M

process, thresholds of (t1, t2, . . ., tk)|t1< t2< . . . < tk separates the L units of gray levels into (C1, C2, . . .,Ck) units of
classes.
In this paper, Kapur’s entropy is considered to determine the optimum thresholds. This scheme works over the
ED

concept of entropy criterion and probability distribution in the histogram of an image. Objective of this technique is
to obtain the optimal threshold value using which the overall entropy can be maximized. In an image, basically
computation of the entropy decides the separability and compactness between each class. Thus, if the entropy
PT

contains maximum value, it means that the selected threshold has optimal value, which appropriately divides the
distinct classes. In case of multispectral satellite images, there are three bands or three colored channel combination
CE

(red, green and NIR bands, respectively) unlike the gray scale images having only one intensity channel. Therefore,
in order to determine the bi-level thresholds, Kapur’s entropy, which acts as an objective function must be optimized
as given in Eq. (14):
AC

1, 2,3 if Multispectral or RGB image

f Kapur (th)  H1c  H 2c , c (14)
1 if Gray sacle image

where, entropies H1 and H2 are calculated using the following formula:

th
Phic  Phic  L
Phic  Phic 
H1c  
i 1
ln 
0c  0c
 ,

H 2c  
i th 1
ln 
1c  1c


(15)

8
 ACCEPTED MANUSCRIPT

As mentioned in Eq. (3), Phic represents probability distribution of the intensity levels. 0c (th) and 1c (th) are the

probability distributions for C1 and C2, whereas ln() indicates the natural logarithm. Furthermore, the objective
function fKapur(th) can be extended for multilevel thresholding using the Eq. (16).

1, 2,3 if Multispectral or RGB image

k
f Kapur (TH)  H
i 1
c
i , c
1 if Gray sacle image
(16)

In Eq. (16), TH=[th1, th2, …, thk-1] indicates the k units of threshold value. The computation of k level of optimum

T
thresholds can be carried out as a k-dimensional problem for one channel images and 3k-dimensional optimization
problem for three channel images respectively. Each of the entropy is determined individually with respective th

IP
value; therefore, to determine k entropies Eq. (15) is expanded as shown below:
th1
Phic  Phic  th2
Phic  Phic  L
Phic  Phic 
  

CR
H1c  ln   , H 2c  ln   , , H kc  ln  c  (17)
i 1 0c  0c  i th1 1 1c  1c  i thk 1 k 1
c
 k 1 

where, probability distribution Phic can be obtained using Eq. (3), and the probability occurrence values

US
(0c ,1c ,...,kc1 ) for k units can be computed by Eq. (12). Finally, Eq. (2) is needed for segmentation of pixels in the
respective classes.
AN
3. Multilevel Thresholding using DE, PSO, WDO and CS Algorithms
In the presented technique, segmentation of an image is processed with an optimization technique that is given
as follows:
M

max imize fOtsu (TH) or f Kapur (TH), TH  [th1 , th2 ,..., thk ] where, TH  X  { k
0  thi  255 (18)

where, i=[1, …, k] and X represents the bounded feasible region, limited in interval 0 to 255. As a result, the CS
ED

methodology is utilized in order to search the intensity levels (TH), which can resolve the problem formulated using
Eq. (3).
In literature, several methods have been reported for simple image segmentation by the histogram based
PT

multilevel thresholding process. However, they are computationally time consuming and their effectiveness is
reduced when applied to a complex color images especially in case of color satellite image. In this paper, the first
CE

time, a fast and efficient method for segmenting color satellite or remote sensing images is proposed. Multilevel
thresholding has been used widely as a tool to segment gray level images, but a limited amount of work has been
published on multilevel thresholding of color images. In gray level images, multilevel thresholding is relatively
AC

straightforward and realized by some peak-picking or valley-searching algorithm on the histogram. For multiband
color satellite images, the task is considerably more complicated as the centers of concentrated gray level regions
must be extracted.
This paper presents the multilevel thresholding based color image segmentation, which is important in image
analysis and in understanding pattern recognition and computer vision field. The method is based on segmentation of
subsets of bands using multilevel thresholding, followed by the fusion of resulting segmentation channels. For color

9
 ACCEPTED MANUSCRIPT

images, the band subsets are chosen as RGB pairs, whose two-dimensional histograms are processed via a peak-
picking algorithm to affect multilevel thresholding.
3.1. Differential Evolution Algorithm (DE)
In past few years, much attention has been paid towards evolutionary algorithm’s applications in various image
processing fields. Storn and Price (Storn, & Price, 1997) have proposed the Differential Evolution algorithm, recently
emerged as a simple yet very competitive evolutionary optimizer. The DE algorithm can be effectively used to solve
many multilevel thresholding and image processing problems as this algorithm have revealed an extensive acceptance

T
and popularity (Cuevas, Zaldivar, & Pérez-Cisneros, 2010).
In this paper, version of DE algorithm being employed is DE/best/l/exp or ‘‘DE1”. This algorithm begins with

IP
the initialization of population of Np, and a D-dimension vectors that have uniformly and randomly distributed
parameter values which lies in between the upper initial parameter bound xj,high and lower initial parameter bound

CR
xj,low, whose values are pre-defined:
x j ,i ,G  x j ,low  rand(0,1)  ( x j ,high  x j ,low );
(19)
j  1, 2,..., D; i  1, 2,..., N p ; G  0.

US
Here, subscript G is known as generation index to which the population belongs, ith solution of population indexes is
represented by index i, and j indicates the parameter index. The three significant operators of DE algorithm are
AN
mutation, crossover and selection.
Mutation: To produce a trial solution, the mutation operator is used for generating mutant vector v𝑖 for every
target vector 𝑥𝑖 within the current population. Therefore, on adding to scaled difference of other two vectors among
M

the current population, DE algorithm initially mutates the best solution.

In accordance with Eq. 20, a mutant vector is generated as:

ED

Vi ,G  X r1,G  F  ( xr 2,G  xr 3,G ); where, r1 , r2 , r3 {1, 2,..., N p } (20)

where, r1, r2 are selected randomly in such a way that r1≠r2≠r3≠i, i = 1...NP, and F acts as the mutation scale factor;
PT

where, F  [0 1] .

Crossover: if the disturbed individual Vi ,G  (v1,i ,G ,..., vn,i ,G ) is generated once, it fits in with crossover operation

along with target individual X i ,G  ( x1,i ,G ,..., xn,i ,G ) , which creates trial solution at the end Ui ,G  (u1,i ,G ,..., u n,i ,G ) as
CE

given:


v j , i , G if rand Cr  j  jj 

AC

u j ,i ,G    (21)

 x j ,i ,G otherwie 

where, random parameter’s index j  1,..., n, jj {1,..., n} , which is once selected for every i. User adjusts the

crossover rate Cr  [0,1] , which controls the section of parameters that belongs to mutant vector contributing to trial
vector.
Selection: at last, solutions are improved with the help of a selection scheme. In DE algorithm, selection process
is different from other evolutionary algorithms. The trial vector is used for replacing the target vector over the

10
 ACCEPTED MANUSCRIPT

succeeding generation. In case cost function of trial vector is either equal or less than the target vector, then on
succeeding generation, target vector is replaced by the trial vector else for minimum one new generation, the target
vector would remain in the population.
U i ,G
 if f (U i ,G )  f ( X i ,G ) 

X i ,G 1    (22)
 X i ,G
 otherwie 

where, f denotes the cost function. The overall procedures repeat till either a predetermined generation number is
obtained or a termination criterion is fulfilled.

T
3.2 Particle swarm optimization (PSO)

IP
In PSO, possible solution and search space (collection of possible solutions) are called as swarm and particle
position correspondingly. Two basic updating equations for particle position i basically dominates the PSO : first is

CR
velocity updating equation, that can be represented as:
Vi (k  1) = w Vi (k )  c11 ( Pibest (k )  Pi (k ))  c22 (Gbest (k )  Pi (k )) (23)
And other equation is position updating equation represented as:
Pi (k  1)  Pi (k )  Vi (k  1)
US (24)

Here, c1 and c2 denotes the cognitive and social acceleration factors respectively, w denotes the inertia weight factor,
AN
that linearly varies from 0 to 1; while, ϕ1 and ϕ2 denotes the uniformly distributed random numbers having range 0 to
1. Eq. (23) is used here to calculate the next velocity Vi (k+1) of particle I whereas the next position Pi(k+1) is trailed
using Eq. (24).
M

Essentially, for optimization problem, one possible solution is specified by the particle position Pi, and position
vector Pi(k) measures the fitness function at every iteration. ‘pbest’ is known to be the position, which corresponds to
best fitness and ‘gbest’ is known to be the overall best solution of all the particles in population. The success to
ED

search the global optimum value relies exceptionally over the initial values of control parameters like w, c1, c2, ϕ1, ϕ2,
maximum iteration number and swarm size (s).
PT

3.3 Wind driven optimization (WDO) technique

WDO is a modern nature inspired global search optimization technique proposed by Bayraktar in 2010
CE

(Bayraktar, Turpin, & Werner, 2011). It is basically motivated by the motion of infinitesimally tiny air particle
present in earth's environment. In WDO, using Newton's second law of motion, the trajectory of these particles can be
expressed. In this optimization algorithm, “wind” corresponds to the horizontal air motion that lies in the troposphere
AC

which is the lowest layer of the earth's atmosphere. Wind blows in an attempt to create equal air pressure in the
atmosphere. Exclusively, the air is used to move at a velocity that is proportional to the pressure gradient, and with
this velocity it blows from high pressure to low pressure. The beginning step of WDO is supported by the Newton’s
second law of motion that determines the precise results specifically to study the atmospheric motion.
Now, in accordance with the Newton's Law, acceleration a results towards the direction of applied net force
when the net force is applied on air parcel, and it is mathematically represented as

a F i (25)

11
 ACCEPTED MANUSCRIPT

where, ρ denotes the air density of an infinitesimal element of volume, a represents the acceleration vector, and Fi
represents the individual forces that acts on the air parcel. Air pressure is related to its temperature and density with
the help of Eq. (26) given using ideal gas law:
P RT (26)
where, P denotes the pressure, R denotes the universal gas constant and T shows the temperature.
This motion of air particle in Eq. (25) is basically affected by four major forces that are FPG, FF, FG and FC
either draw away the wind from its path or make the wind to move in a particular direction. Pressure Gradient Force

T
(FPG), is the most noticeable force among the other forces, which causes the air to move. Although the Friction

IP
Force (FF) merely acts to resist such motion as described in Eq. (26). Whereas in three-dimensional atmosphere, the
Gravitational Force (FG) behaves like a vertical force when mapped to N-dimensional space. The rotation of earth

CR
causes the Coriolis Force (FC), which deflects the trail of the wind from one dimension to other dimension. However,
Coriolis force, in case of WDO, is employed like a motion in one dimension which influences the velocity
(Bayraktar, Turpin, & Werner, 2011).

FG    V g
US
Each of the above mentioned forces are governed using the equations given below:

FPG  P  V (27)

(28)
AN
Fc  2 u (29)

FF    u (30)
M

Here, δV is an infinitesimal air volume, ∇P denotes the pressure gradient, g represents the gravitational acceleration
constant, Ω denotes the earth’s rotation, u shows the velocity vector of the wind and α shows the friction coefficient.
ED

By putting above mentioned forces in Eq. (25), summation of all the four forces can be obtained by Eq. (31):
 u t = (  V g )  (P V )  ( u )  (2 u ) (31)
If an infinitesimal air parcel is considered which is moving along the wind, then the velocity update equation can be
PT

determined by (32). On the basis of Eq. (26) that is an ideal gas law equation, unity time step (Δt = 1) can be assumed
for simplifying the equation, and if ρ is written in terms of pressure, then updated velocity equation becomes
(Komurcu, Bossard, & Werner, 2013):
CE

P   cother dim 
unew   (1   )uold   g   xold    max  1 RT ( xmax  xold )    old  (32)
 Pold   Pold 
AC

which are velocity and position updating equation for WDO respectively. Here, P represents the pressure, T
represents temperature, R represents universal gas constant, c=-2ΩRT called Coriolis coefficient. In Eq. (32), unew, an
updated velocity for the subsequent iteration rely on uold that is the current iteration velocity, current location xold of
air parcel in search space, xmax which is the distance that is found from the highest pressure point, Pmax that is the
maximum pressure, Pold that is the pressure at current location, gravitational acceleration (g), temperature (T) and
constants R, α, and c.

12
 ACCEPTED MANUSCRIPT

The pressure term in Eq. (32) of the WDO algorithm is equivalent to fitness of a chromosome in GA. In WDO
expression (32), if PSO is compared with WDO algorithm; the analogous velocity update equations can be realized.
Eq. (33) can be used to update the position of air parcel after updating the velocity of parcel using Eq. (32)
(Komurcu, Bossard, & Werner, 2013),
xnew  xold  (unew  t ) (33)

where, xnew and xold are the current and new position of air parcel in search space respectively. In Eq. (33), xold is
used to represent the air parcel that continues to move over its previous path along with some resistance that is

T
created due to friction. An attractive force unew pulls against the center of coordinate system. For optimization

IP
problem, Δt indicates the force against position of maximum pressure that is supposed to be the global best position
xnew follows the Coriolis force, in fact, which is a deflecting force. In this way, WDO algorithm presents a highly

CR
effective yet a simple way to resolve the complex problems associated with the optimization. From the above
discussion, it can be concluded that WDO is controlled by RT, c, α, g and number of iteration. Detailed description of
WDO is given in [62].

3.4 Cuckoo Search (CS) Algorithm US

An evolutionary optimization algorithm is used in this paper which is motivated with the lifestyle of cuckoos.
AN
Cuckoo search algorithm is an elitist search algorithm based on population (Bhandari, Singh, Kumar, & Singh,
2014a; Yang, & Deb, 2009). Every new solution in this algorithm attempts to search about previously found best
solution. CS algorithm is consisting of two control parameters. The problem solving efficiency and general structure
M

of CS technique is elaborated in (Bhandari, Singh, Kumar, & Singh, 2014a). Cuckoo birds are most popular because
of their attractive voice and fascinating singing style; moreover, their reproduction policy is also one of the most
ED

aggressive among the birds. Cuckoos can engage indirect conflict with the host birds. Usually, host birds either throw
the alien eggs from the nest or they leave the nest to make a new nest.
The three essentially idealized rules of CS algorithm.
PT

1. Every cuckoo lays single egg at an instant and abandons it in a nest, which it chooses randomly.
2. The best nests that contain the eggs of high quality will only be bear to the subsequent generation.
3. There are fix number of available host nest, and the host bird having probability of pa  [0,1] discovers the egg that
CE

was laid by a cuckoo.

Now, for such cases, either the host bird throws away the egg or would abandon that nest in order to build a
AC

totally new nest at new location. In order to simplify the last assumption, many authors have assumed that it can be
supposed with probability pa for n nest that are replaced with the latest nests (random new solutions). Every egg
inside the nest stands for one solution while a new solution is indicated by cuckoo egg, which aims to exploits
potentially better new solution. A new nest is generated by Levy flight in initial iteration while keeping the current
best. A Lévy flight is performed by Eq. (34): for generating new solution, xi (t+1) for cuckoo i,
xi (t  1)  xi (t )    Lévy( ) (34)

13
 ACCEPTED MANUSCRIPT

where, α is the step size (α>1) and is related to size of the problem. α=1 is used for most of the cases. The product
 deals with multiplications process considered entry wise. However, a random walk is provided by the Lévy flights,
whereas their random step lengths are taken from Lévy distribution for large steps defined by (35):

Lévy( )  t  where 1    3 (35)

In accordance with the application, Lévy function can be modified. One of the L’evy function is known as
Mantegna's algorithm with an infinite variance and mean. Lévy flights are more proficient in case of extensive search
space in comparison with Brownian random walks, because variance (σ2) of Levy flights increases at greater rate than

T
Brownian random walks. Moreover, variance of Levy flights distribution is expressed as:

IP
 2 (t ) ~ t 2 where 1    2 (36)
Which expand at a greater rate than the linear relationship of Brownian random walks followed by σ2(t) →t. On the

CR
basis of above expression, it can be noticed that the CS algorithm is mainly controlled by four parameters such as
number of iterations, step size (α), number of host nests (NS), and probability (pa).

4. Study of Experiments and Results

US
In this part, the experimental outcomes are presented. The multilevel colored satellite image thresholding
problem is the method of finding multiple numbers of thresholds (TH) for each band (i.e., red, green and NIR) within
AN
the range [1, L]. In this study, 6, 8, 12 and 16 level of thresholds are used to maximize the entropy for different
satellite images by utilizing DE, PSO, WDO and CS algorithm with different fitness criteria like Kapur’s entropy and
between-class variance. Tables 1 depicts the constraints used for DE, PSO, WDO and CS algorithms. The results
M

achieved are then examined for DE, PSO, WDO and CS algorithms as shown visually in Figs. 3-8 and quantitatively
in Tables 2-5.
Furthermore, with rapid development of remote sensing technology, the high-resolution satellite images have
ED

been employed to many areas which need accurate segmented part in various remote sensing applications. Moreover,
in satellite images, it is normally the occasion that certain area covers only a few pixels, whereas the nearby areas are
PT

considerably huge. Hence, automatically identifying the segmented parts or clusters of such usually changing pixel
variations creates a challenge in multilevel thresholding problem. Therefore, the proposed CS and Kapur’s based
approach has been tested on different satellite images which reveal robustness of the segmented results for multilevel
CE

thresholding problem.
Multilevel thresholding for color image segmentation is one of the most widely used techniques. However, the
conventional multilevel color image thresholding methods are time consuming as they exhaustively search for
AC

optimal thresholds to optimize the objective functions. In favor of multilevel thresholding problem, various
techniques have been developed for gray scale segmentation of the images. In this paper, the very first time, a distinct
study on the application of colored image segmentation using cuckoo search and other optimization algorithms to
solve the color image segmentation problem via Otsu or Kapur’s method has been carried out. For the entire test of
satellite images that have been considered, the CS-Kapur’s entropy based segmented results are better than those
obtained with DE, WDO and PSO. The experimental results reflect the evidence of outstanding performance and
accuracy of the proposed technique.

14
 ACCEPTED MANUSCRIPT

Satellite images often need segmentation in presence of uncertainty, caused due to the factors like highly
dependent on environmental conditions, poor resolution and poor illumination and have very low spatial resolution.
Multilevel thresholding can be used precisely to locate the area of interest, detect, identify and assist in satellite image
segmentation. Furthermore, color satellite image segmentation using Otsu and Kapur’s method has been never
examined with CS, DE, WDO and PSO algorithms for color image segmentation case. In satellite images, rate of
information are very high. Because of this, existing features in the image is very dense, and due to that, the rate of
change from one region to other region is very rapid. Therefore, in case of segmentation of remote sensing images or

T
satellite images, accurate segmentation is a very challenging task. Therefore, in this paper, a new approach for color
image segmentation using CS algorithm with Otsu or Kapur’s method for multilevel thresholding has been introduced

IP
towards the effective colored segmentation of satellite images.

CR
4.1. Image dataset
In order to get better comparison of the segmented results, experiments have been examined on completely 10
different satellite images having miscellaneous kind of features as shown in Fig. 2 (a1-a10)

US
(http://earthobservatory.nasa.gov/?eocn=topnav&eoci=logo). Each image shows an unique histogram Fig. 2 (b1-b10).
On the basis of grey level distribution of the satellite images, it can be seen that as compared to simple gray level
images, the remote sensing pictures are more challenging to fragment accurately because of histogram multimodality
AN
properties. The images used in this paper are encoded at 24-bit true color (red, green, NIR band or RGB). The
segmented results are obtained at 6, 8, 12 and 16 level of thresholds for each satellite images. Since optimization
algorithms deals with randomness, the initial solutions have been generated through random search. Therefore, to get
M

best segmented output, each experiment has been reiterated for 10 times for each level of threshold and image
respectively.
ED

4.2. Segmented Image quality measures

In this experimental section, different well known image processing matrices such as peak signal-to-noise ratio
PT

(PSNR) and mean-squared error (MSE), structural-similarity index (SSIM) and feature similarity index (FSIM)
similarity metrics have been used. PSNR and MSE calculated directly through intensity values of the image which
usually indicate the strength and accuracy of final reconstructed signal or image. On the other hand, SSIM and FSIM
CE

are exploited to calculate the resemblance among the images computed after thresholding process that is followed by
numerous objective functions (Kapur’s and Otsu) and original images. On the basis of following noteworthy
parameters, performance of the proposed method is checked:
AC

M N

[I (i, j)  I (i, j)]

1
MSE  2
(37)
MN i 1 j 1

 2552 
PSNR  10 log10   (38)
 MSE 
where, M, N represents the size of image, I denotes the original image and Ĩ stands for thresholded image at a specific
level.

15
 ACCEPTED MANUSCRIPT

SSIM is a picture feature valuation, which works on the basis of structural information degradation. The SSIM
is used to analyse structures region of the original and thresholded image (Bhandari, Singh, Kumar, & Singh, 2014a;
Oliva, Cuevas, Pajares, Zaldivar, & Osuna, 2014). The SSIM index is determined as followed:
(2 x  y  C1 )(2 xy  C2 )
SSIM( x, y)  (39)
( 2
x   2 y  C1 )( 2 x   2 y  C2

where µx and µy stands for mean intensity of image of x and y correspondingly, σx and σy indicates the standard
deviations of x and y correspondingly, σxy is the local sample correlation coefficient among x and y. C1 and C2 are the

T
constants, and are selected to escape unpredictability when  2 x   2 y are very near to zero. Here, C1 = C2 = 0.065.

IP
The SSIM can take the values in [−1, 1] range and a greater rate of SSIM represents superior result. For real color
RGB images, SSIM metric can be extended as following:

CR
SSIM   SSIM(xc , y c ) (40)
c

Here, xc, yc correspond to cth frame of the input image and segmented image obtained through multilevel thresholding

US
task, where, c stands for frame quantity (i.e., c = 1, 2, 3 in real color RGB pictures).
FSIM is used to calculate the feature similarity between the input and processed images (Bhandari, Singh,
Kumar, & Singh, 2014a; Oliva, Cuevas, Pajares, Zaldivar, & Osuna, 2014), which is formulated as:
AN
FSIM 
 X 
S L ( X ) PCm ( X )
(41)
 X 
PCm ( X )

where, Ω represents the entire image, and SL(x) indicates the similarity between the segmented image obtained
M

through multilevel thresholding task and input image. The FSIM metric can be extended for true color RGB images
as following:
ED

FSIM   FSIM(xc , y c ) (42)

c

4.3 Experimental Setup

PT

In the proposed methodology, DE, PSO, WDO and CS algorithms have been used individually, and comparison
of their search productivities have been made. Two most well-known objective functions named as Kapur’s entropy
and Otsu method have been tested. In the presented method, every candidate solution relates to an m-level of
CE

threshold values (6, 8, 12, 16) to be optimized. In initialization steps of the algorithms, the populations of solutions
are created in arbitrary way inside the location [0, L−1] on every aspect. When a population is produced, solutions in
population are allotted a fitness point value, thereafter, algorithms begins to find the best threshold values through
AC

their collection and perturbation procedures.

4.4 Evaluation of the experimental results using Kapur’s entropy

The results acquired using Kapur’s objective function for various satellite images are discussed in this section.
Table 2 gives the maximum entropy values (objective values), based on Kapur’s entropy using nature inspired
algorithms DE, WDO, PSO and CS obtained at various levels of threshold for each satellite image. In Table 2, the
SSIM and FSIM results are also included which are calculated between the original (input) image and its thresholded

16
 ACCEPTED MANUSCRIPT

image for number of thresholds and each nature inspired algorithms. Once the thresholded images have been gained,
every pixel of the input image has been fixed to threshold value that is nearest in terms of Euclid distance. A review
of Table 2 reveals that CS has obtained maximum entropy and best SSIM, FSIM value for threshold numbers of all
the images whereas, DE, WDO and PSO algorithms have obtained acceptable results.
Furthermore, CPU time, PSNR and MSE of Kapur’s entropy based nature inspired algorithms DE, WDO, PSO
and CS are included in Table 3. For nearly all the images and at every threshold numbers, CS has obtained highest
PSNR and minimum MSE values. DE, WDO as well as PSO follow the CS’s success. SSIM and FSIM represents the

T
quality metric values in Table 2, which is very well associated with the outcome of maximized entropy values shown
in Tables 2 (objective values). In other words, it can be seen that the results of entropy (objective) values are also

IP
confirmed by the values of SSIM and FSIM. As low running times are needed by the real-time applications, the
analysis of CPU time required by the algorithms is significant. Table 2 provides the information of CPU time to

CR
obtain the thresholded images using Kapur’s entropy.
Figs. 3 and 5 show the Kapur’s entropy based 6 and 8-level thresholding results using DE, WDO, PSO and CS

US
algorithms for each test image, respectively. Fig. 7 shows the comparison of various threshold numbers (12 and 16).
Visual results of the images in Figs. 3, 5 and 7 indicates that the images obtained with Kapur’s entropy using DE,
WDO, PSO and CS are better than the Otsu based method. Furthermore, the Kapur’s entropy using CS algorithm
AN
provides best segmented results for almost all the test images.
In few cases, the optimization algorithms that are computing the best results for several threshold levels are
similar for each other. The fitness value that is maximized through the optimization technique is neither SSIM nor
FSIM is the fundamental cause of such kind of condition. Between-class variance or Kapur’s entropy are expected to
M

be maximized by such algorithms. Conversely, for all the algorithms, as the threshold number is higher, quality of the
obtained result also goes higher as expected.
ED

4.5 Evaluation of the experimental results using between-class variance

In this section, the results acquired for various satellite images using between-class variance objective function
PT

are discussed. The same experiments were repeated using between-class variance for each satellite image as given in
Figs. 4 and 6. Table 4 gives the Otsu based maximized between-class variance values (objective values), SSIM and
FSIM values using nature inspired algorithms DE, WDO, PSO and CS acquired at different levels of threshold for
CE

each satellite image. A review of Table 4 reveals that the CS has obtained the best between-class variance and best
SSIM, FSIM value for all the images and threshold numbers. However, DE, WDO and PSO algorithms have obtained
AC

acceptable results. CPU time, PSNR and MSE of Otsu method based nature inspired algorithms DE, WDO, PSO and
CS are included in Table 5. CS has obtained highest PSNR and minimum MSE values for nearly all the images and at
each threshold numbers. DE, WDO and PSO follow the CS’s success.

4.6 Comparison of the experimental results using Kapur’s Entropy and Otsu methods
The Kapur’s criterion and Otsu’s criterion are taken into account with an aim to verify the effectiveness as well
as efficiency of the nature inspired optimization algorithms. Beside the unique histogram plots given for each image
[Figs. (b1-b10)], their respective histograms equalized (enhanced) images are also presented for clear interpretation

17
 ACCEPTED MANUSCRIPT

of the segmented results (Figs. (c1-c10)). Performance of both the objective functions has been tested with 10
different satellite images and evaluated with DE, WDO, PSO and CS algorithms. A comparison has been made in this
study between the two versions of nature inspired algorithms, one with Kapur criterion and other using Otsu objective
function. The results demonstrate that Kapur criterion gives better results than the Otsu function.
Table 2, Table 3, Table 4 and Table 5 represents the computed values for 10 different satellite images. The best
values found at the end of entire test value are depicted in bold. Tables show how CS algorithm, in case of colored
satellite image segmentation, has given better performance than others. Such values are computed using Kapur's

T
method as objective function, which is mentioned in Table 2 and Table 3. On the other hand, the same experiment has
been repeated using Otsu's method. The results of this experiment are presented in Table 4 and Table 5. From Table 2

IP
and Table 3, it can be seen that CS algorithm using Kapur’s entropy can obtain higher objective values, SSIM, FSIM,
PSNR and minimum MSE than the other mentioned methods, whereas CS algorithm with Otsu method gives

CR
acceptable result in comparison with Kapur’s based approach. However, for Otsu based colored multilevel
thresholding processing, CS algorithm has again performed better than DE, WDO and PSO based Otsu approach.

US
Although, in case of Otsu based segmentation, MSE is found to be minimum with PSO technique only, but other
fidelity criteria reveals the best performance of CS algorithm.
A review of data in Table 3 and Table 5 reveals that, for both objective functions, as the threshold number
AN
increases, calculation time also increases in its direct proportion. DE using Kapur’s entropy and Otsu method is the
most effective algorithm in terms of CPU time. CS, PSO and WDO are slower than DE. While in terms of running
time, the slowest algorithm with Kapur’s entropy was found to be PSO and slowest algorithm with Otsu method was
WDO. Conversely, as compared to Kapur’s entropy, Otsu method is slower in terms of CPU time. The numerical
M

illustrations demonstrate that the proposed CS algorithm and Kapur’s entropy based colored multilevel thresholding
for satellite image segmentation outperforms the other methods.
ED

The segmentation results obtained for various satellite images illustrate that CS algorithm has comprehensively
outperformed the other algorithms in terms of solution quality, feature preservation and stability. The validity and
accuracy of the proposed technique is reported both qualitatively and quantitatively. Another advantage of CS
PT

algorithm is that it has very less control parameters than DE, WDO and PSO algorithms. The segmentation results of
nature based optimization algorithm for colored satellite image multilevel thresholding are promising, and it
CE

encourages further research for applying these algorithms to complex and real-time image processing application
such as satellite image enhancement, satellite image denoising, optimization based image classification and also in
various computer vision problems. In addition, other new objective functions can also be implemented using these
AC

optimization algorithms for colored multilevel thresholding.

The achieved results can be summarized as follows:
 Four methods, based on nature inspired algorithms called Differential Evolution, Wind Driven Optimization,
Particle Swarm Optimization and Cuckoo Search have been evaluated with two well-known objective functions
in terms of solution of multilevel colored image thresholding problem.
 These four methods with both the objective functions have been subjected to detailed comparative analysis using
10 different satellite images.

18
 ACCEPTED MANUSCRIPT

 Although each algorithm has statistically similar results for practical applications, the algorithms are sorted as
CS>PSO>DE>WDO with Kapur’s entropy, and as CS>DE>PSO>WDO with Otsu method in terms of
robustness in descending order.
 When the algorithms are sorted from good to bad in terms of running time, it is seen that they are sorted in order
of DE<CS<WDO<PSO with Kapur’s entropy, whereas order of complexity from low to high in case of Otsu
based method is DE<CS<PSO<WDO. However, Kapur’s entropy was found to be faster than Otsu method.

5. Conclusion

T
In previous works, authors have developed various techniques using multilevel thresholding for gray scale image

IP
segmentation. However, this paper focuses on color image multilevel thresholding for satellite and natural images
using different optimization algorithms. In this study, the problem of color image segmentation through multilevel

CR
thresholding has been cast as a continuous optimization problem, where two objective functions namely between-
class-variance and Kapur’s entropy have been considered using nature inspired algorithms DE, WDO, PSO and CS.
Mainly, the aim realized by each algorithm is to obtain the best values of a given number of thresholds that maximize

US
the used objective function. A thorough performance based experimental study has been conducted with the aim to
assess the ability of proposed algorithms to solve efficiently the tackled problem and to compare their performance to
AN
other state-of-the-art population based algorithms such as DE, WDO and PSO algorithms. The results estimated by
statistical tools recommend that CS algorithm with both Kapur’s entropy and Otsu criterion can be powerfully
utilized in segmentation by multilevel thresholding operation. Firstly, in order to test the performance of CS with both
the objective functions, the number of thresholds is set as 6, and 8 respectively. However, in comparison with Otsu
M

methods, CS algorithm with Kapur’s entropy outperforms the other methods in terms of objectives function, SSIM,
FSIM, MSE and PSNR. Secondly, in order to compare the finding capability of every population-based algorithm on
ED

a huge aspect, the number of thresholds is set as 12 and 16 respectively. The experimental results show that the
performance of CS is obviously better than other algorithms with the increase in number of thresholds.
The segmentation results obtained for various satellite images illustrate that CS algorithm has comprehensively
PT

outperformed the other algorithms in terms of solution quality, feature preservation and stability. The validity and
accuracy of the proposed technique is reported both qualitatively and quantitatively. Another advantage of CS
algorithm is that it has very less control parameters than DE, WDO and PSO algorithms. The segmentation results of
CE

nature based optimization algorithm for colored satellite image multilevel thresholding are promising, and it
encourages further research for applying these algorithms to complex and real-time image processing application
AC

such as satellite image enhancement, satellite image denoising. Numerous combinations of entropies may be used
along with multi-objective optimization based approaches in this context to improve the quality of color image
segmentation. Also, fuzzy entropy or Renyi’s entropy based multilevel thresholding frameworks can be adopted in
order to obtain the better outcomes.

References

Agrawal, S., Panda, R., Bhuyan, S., & Panigrahi, B. K. (2013). Tsallis entropy based optimal multilevel thresholding using cuckoo
search algorithm. Swarm and Evolutionary Computation, 11, 16–30.

19
 ACCEPTED MANUSCRIPT

Aja-Fernández, S., Curiale, A. H., & Vegas-Sánchez-Ferrero, G. (2015). A local fuzzy thresholding methodology for multiregion
image segmentation. Knowledge-Based Systems, 83, 1-12.
Akay, B. (2013). A study on particle swarm optimization and artificial bee colony algorithms for multilevel thresholding. Applied
Soft Computing, 13(6), 3066-3091.
Arora, S., Acharya, J., Verma, A., & Panigrahi, P. K. (2008). Multilevel thresholding for image segmentation through a fast
statistical recursive algorithm. Pattern Recognition Letters, 29(2), 119-125.
Ashish B. K, Kumar A, Padhy P. K. (2011). Satellite image processing using discrete cosine transform and singular value
decomposition”. In Processing Advances in Digital Image Processing and Information Technology, Springer Berlin

T
Heidelberg, 205, 277-290.
Ayala, H. V. H., dos Santos, F. M., Mariani, V. C., & dos Santos Coelho, L. (2015). Image thresholding segmentation based on a

IP
novel beta differential evolution approach. Expert Systems with Applications, 42(4), 2136-2142.
Bakhshali, M. A., & Shamsi, M. (2014). Segmentation of color lip images by optimal thresholding using bacterial foraging

CR
optimization (BFO). Journal of Computational Science, 5(2), 251-257.
Bayraktar, Z., Turpin, J. P., & Werner, D. H. (2011). Nature-inspired optimization of high-impedance metasurfaces with ultrasmall
interwoven unit cells. IEEE Letters on Antennas and Wireless Propagation, 10, 1563-1566.

US
Bhandari, A. K., Kumar, A., & Padhy, P. K. (2011). Enhancement of low contrast satellite images using discrete cosine transform
and singular value decomposition. World Academy of Science, Engineering and Technology, 55, 35–41.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2012a). Feature extraction using normalized difference vegetation index (NDVI): A
AN
case study of Jabalpur city. Procedia Technology, 6, 612–621.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2012b). SVD Based Poor Contrast Improvement of Blurred Multispectral Remote
Sensing Satellite Images. In Processing IEEE 3rd International Conference on Computer and Communication Technology
(ICCCT), 156-159.
M

Bhandari, A. K., Singh, V. K., Kumar, A., & Singh, G. K. (2014a). Cuckoo search algorithm and wind driven optimization based
study of satellite image segmentation for multilevel thresholding using Kapur’s entropy. Expert Systems with Applications,
ED

41(7), 3538-3560.
Bhandari, A. K., Soni, V., Kumar, A., & Singh, G. K. (2014b). Artificial Bee Colony-based satellite image contrast and brightness
enhancement technique using DWT-SVD. International Journal of Remote Sensing, 35(5), 1601-1624.
Bhandari, A. K., Soni, V., Kumar, A., & Singh, G. K. (2014c). Cuckoo search algorithm based satellite image contrast and
PT

brightness enhancement using DWT–SVD. ISA transaction, 53(4), 1286-1296.

Bhandari, A. K., Gadde, M., Kumar, A., Singh, G. K. (2012). Comparative analysis of different wavelet filters for low contrast and
CE

brightness enhancement of multispectral remote sensing images. In Proceding IEEE Inernational Conference on Machine
Vision and Image Processing (MVIP), 81-86.
Bhandari, A. K., Kumar, A., Chaudhary, S., & Singh, G. K. (2015). A new beta differential evolution algorithm for edge preserved
AC

colored satellite image enhancement. Multidimensional Systems and Signal Processing, 1-33.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2015a). Modified Artificial Bee Colony based Computationally Efficient Multilevel
Thresholding for Satellite Image Segmentation using Kapur's, Otsu and Tsallis Functions. Expert Systems with Applications,
42(3), 1573–1601.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2015b). Improved feature extraction scheme for satellite images using NDVI and
NDWI technique based on DWT and SVD. Arabian Journal of Geosciences, 8(9), 6949-6966.
Bhandari, A. K., Kumar, A., & Singh, G. K. (2015c). Improved knee transfer function and gamma correction based method for
contrast and brightness enhancement of satellite image. AEU-International Journal of Electronics and Communications,
69(2), 579-589.

20
 ACCEPTED MANUSCRIPT

Bhandari, A. K., Kumar, A., & Singh, G. K. (2015d). Tsallis entropy based multilevel thresholding for colored satellite image
segmentation using evolutionary algorithms. Expert Systems with Applications, 42(22), 8707-8730.
Bhandari, A. K., Kumar, D., Kumar, A., & Singh, G. K. (2016). Optimal sub-band adaptive thresholding based edge preserved
satellite image denoising using adaptive differential evolution algorithm. Neurocomputing, 174, 698-721.
Bhandari, A. K., Kumar, A., Singh, G. K., & Soni, V. (2016a). Dark satellite image enhancement using knee transfer function and
gamma correction based on DWT–SVD. Multidimensional Systems and Signal Processing, 27(2), 453-476.
Bhandari, A. K., Kumar, A., Singh, G. K., & Soni, V. (2016b). Performance study of evolutionary algorithm for different wavelet
filters for satellite image denoising using sub-band adaptive threshold. Journal of Experimental & Theoretical Artificial

T
Intelligence, (ahead-of-print), 1-25.
Bouaziz, A., Draa, A., & Chikhi, S. (2015). Artificial bees for multilevel thresholding of iris images. Swarm and Evolutionary

IP
Computation, 21, 32-40.
Cai, H., Yang, Z., Cao, X., Xia, W., & Xu, X. (2014). A New Iterative Triclass Thresholding Technique in Image

CR
Segmentation,’ IEEE Transactions on Image Processing, 23(3), 1038-1046.
Chang, C. I., Du, Y., Wang, J., Guo, S. M., & Thouin, P. D. (2006). Survey and comparative analysis of entropy and relative
entropy thresholding techniques. IEE Proceedings Vision Image Signal Processing, 153(6), 837–850.

US
Cuevas, E., Zaldivar, D., & Pérez-Cisneros, M. (2010). A novel multi-threshold segmentation approach based on differential
evolution optimization. Expert Systems with Applications, 37(7), 5265-5271.
Dey, S., Bhattacharyya, S., & Maulik, U. (2016). New quantum inspired meta-heuristic techniques for multilevel colour image
AN
thresholding. Applied Soft Computing, 46, 677–702.
Fan, C., Ouyang, H., Zhang, Y., & Xiao, L. (2014). Optimal multilevel thresholding using molecular kinetic theory optimization
algorithm. Applied Mathematics and Computation, 239, 391-408.
Hammouche, K., Diaf, M., & Siarry, P. (2010). A comparative study of various meta-heuristic techniques applied to the multilevel
M

thresholding problem. Engineering Applications of Artificial Intelligence, 23(5), 676-688.

Horng, M. H. (2010a). A multilevel image thresholding using the honey bee mating optimization. Applied Mathematics and
ED

Computation, 215(9), 3302-3310.

Horng, M. H. (2010b). Multilevel minimum cross entropy threshold selection based on the honey bee mating optimization. Expert
Systems with Applications, 37(6), 4580-4592.
Horng, M. H., & Liou, R. J. (2011). Multilevel minimum cross entropy threshold selection based on the firefly algorithm. Expert
PT

Systems with Applications, 38(12), 14805-14811.

Huang, D. Y., & Wang, C. H. (2009). Optimal multilevel thresholding using a two-stage Otsu optimization approach. Pattern
CE

Recognition Letters, 30(3), 275-284.

Hussein, W. A., Sahran, S., & Abdullah, S. N. H. S. (2016). A fast scheme for multilevel thresholding based on a modified bees
algorithm. Knowledge-Based Systems, 101, 114-134.
AC

<http://earthobservatory.nasa.gov/?eocn=topnav&eoci=logo>
Karaboga D. (2005). An idea based on honey bee swarm for numerical optimization. Technical Report-TR06, Kayseri, Turkey:
Erciyes University, 200.
Kapur, J. N., Sahoo, P. K., & Wong, A. K. C. (1985). A new method for gray-level picture thresholding using the entropy of the
histogram. Computer Vision Graphics Image Processing, 29(3), 273–285.
Kumar, A., Bhandari, A. K., & Padhy, P. (2012). Improved normalised difference vegetation index method based on discrete
cosine transform and singular value decomposition for satellite image processing. IET Signal Processing, 6(7), 617-625.
Kumar, S., Kumar, P., Sharma, T. K., & Pant, M. (2013). Bi-level thresholding using PSO, Artificial Bee Colony and MRLDE
embedded with Otsu method. Memetic Computing, 5(4), 323-334.

21
 ACCEPTED MANUSCRIPT

Kurban, T., Civicioglu, P., Kurban, R., & Besdok, E. (2014). Comparison of Evolutionary and Swarm Based Computational
Techniques for Multilevel Color Image Thresholding. Applied Soft Computing, 23, 128-143.
Liu, Y., Mu, C., Kou, W., & Liu, J. (2014). Modified particle swarm optimization-based multilevel thresholding for image
segmentation. Soft Computing, 1-17.
Mala, C., & Sridevi, M. (2015). Multilevel threshold selection for image segmentation using soft computing techniques. Soft
Computing, 1-18.
Manikandan, S., Ramar, K., Willjuice I. M., & Srinivasagan, K. G. (2014). Multilevel thresholding for segmentation of medical
brain images using real coded genetic algorithm. Measurement, 47, 558-568.

T
Oliva, D., Cuevas, E., Pajares, G., Zaldivar, D., & Perez-Cisneros, M. (2013). Multilevel Thresholding Segmentation Based on
Harmony Search Optimization. Journal of Applied Mathematics, 1-24.

IP
Oliva, D., Cuevas, E., Pajares, G., Zaldivar, D., & Osuna, V. (2014). A Multilevel Thresholding algorithm using electromagnetism
optimization. Neurocomputing, 139, 357-381.

CR
Otsu, N. (1979). A threshold selection method from gray-level histograms. IEEE Transactions on Systems Man Cybernetics, SMC-
9(1), 62–66.
Ouadfel, S., & Taleb-Ahmed, A. (2016). Social spiders optimization and flower pollination algorithm for multilevel image

US
thresholding: A performance study. Expert Systems with Applications, 55, 566-584.
Pare, S., Bhandari, A. K., Kumar, A., Singh, G. K., & Khare, S. (2015). Satellite image segmentation based on different objective
functions using genetic algorithm: A comparative study. In 2015 IEEE International Conference on Digital Signal Processing
AN
(DSP) (pp. 730-734). IEEE.
Pare, S., Kumar, A., Bajaj, V., & Singh, G. K. (2016). A multilevel color image segmentation technique based on cuckoo search
algorithm and energy curve. Applied Soft Computing, 47, 76-102.
Peng, H., Wang, J., & Pérez-Jiménez, M. J. (2015). Optimal multilevel thresholding with membrane computing. Digital Signal
M

Processing, 37, 53-64.

Poli R., Kennedy J., Blackwell T. (2007). Particle swarm optimization: an overview. Swarm Intelligence, 1(1), 33–57.
ED

Raja, N. S. M., Rajinikanth, V., & Latha, K. (2014). Otsu Based Optimal Multilevel Image Thresholding Using Firefly Algorithm.
Modelling and Simulation in Engineering, 1-17.
Sahoo, P. K., Soltani, S., Wong, A. K. C., & Chen, Y. C. (1988). A survey of thresholding techniques. Computer Vision Graphics
and Image Processing, 41, 233–260.
PT

Sathya, P.D., & Kayalvizhi, R. (2011a). Modified bacterial foraging algorithm based multilevel thresholding for image
segmentation. Engineering Applications of Artificial Intelligence, 24, 595–615.
CE

Sathya, P. D., & Kayalvizhi, R. (2011b). Optimal multilevel thresholding using bacterial foraging algorithm. Expert Systems with
Applications, 38(12), 15549-15564.
Sezgin, M., & Sankur, B. (2004). Survey over image thresholding techniques and quantitative performance evaluation. Journal of
AC

Electronic Imaging, 13(1), 146–165.

Sarkar, S., Das, S., & Chaudhuri, S. S. (2016). Hyper-spectral image segmentation using Rényi entropy based multilevel
thresholding aided with differential evolution. Expert Systems with Applications, 50, 120-129.
Son, L. H., Tuan, T. M. (2016). A cooperative semi-supervised fuzzy clustering framework for dental X-ray image segmentation.
Expert Systems with Applications, 46, 380-393.
Soni, V., Bhandari, A. K., Kumar, A., & Singh, G. K. (2013). Improved sub-band adaptive thresholding function for denoising of
satellite image based on evolutionary algorithms. IET Signal Processing, 7(8), 720-730.
Storn, R., & Price, K. (1997). Differential evolution–a simple and efficient heuristic for global optimization over continuous
spaces. Journal of global optimization, 11(4), 341-35.

22
 ACCEPTED MANUSCRIPT

Sun, G., Zhang, A., Yao, Y., & Wang, Z. (2016). A novel hybrid algorithm of gravitational search algorithm with genetic
algorithm for multilevel thresholding. Applied Soft Computing, 46, 703–730.
Suresh, S., & Lal, S. (2016). An efficient cuckoo search algorithm based multilevel thresholding for segmentation of satellite
images using different objective functions. Expert Systems with Applications, 58, 184-209.
Tang, K., Yuan, X., Sun, T., Yang, J., & Gao, S. (2011). An improved scheme for minimum cross entropy threshold selection
based on genetic algorithm. Knowledge-Based Systems, 24(8), 1131-1138.
Tao, W., Jin, H., & Liu, L. (2007). Object segmentation using ant colony optimization algorithm and fuzzy entropy. Pattern
Recognition Letters, 28(7), 788-796.

T
Tsai, W. (1985). Moment-preserving thresholding: A new approach. Computer Vision Graphics and Image Processing, 29, 377–
393.

IP
Wu, T., Xiao, J., Qin, K., & Chen, Y. (2014). Cloud model-based method for range-constrained thresholding. Computers &
Electrical Engineering. 2014.

CR
Yang, X. S., & Deb, S. (2009). Cuckoo search via levey flights. In Proceedings of the World congress on nature and biologically
inspired computing, NABIC: Coimbatore, vol. 4, pp. 210– 214, 2009.
Yin, S., Zhao, X., Wang, W., & Gong, M. (2014). Efficient multilevel image segmentation through fuzzy entropy maximization

US
and graph cut optimization. Pattern Recognition, 47(9), 2894-2907.
Zhang, J., Li, H., Tang, Z., Lu, Q., Zheng, X., & Zhou, J. (2014). An improved quantum-inspired genetic algorithm for image
multilevel thresholding segmentation. Mathematical Problems in Engineering, 1-12.
AN
M
ED
PT
CE
AC

23
 ACCEPTED MANUSCRIPT

Table 1: Specific values for given parameters used in each algorithm.

Algorithms Parameters Values

Number of objectives 1
Number of constraints 0
Number of decision variables 6, 8, 12, 16
DE Scaling factor 0.5
Crossover Probability 0.2
Generation bound 10000

T
Function evaluations bound 150

IP
Swam size 200
No. of iterations 1000

CR
Cognitive, social and neighborhood acceleration (C1, C2, C3) 2, 2, 1
Lower bound lb (Wmin) and Upper bound ub (Wmax) 1 & 256
Error goal and Max Trial limit 1e-7 & 500
PSO
Value of velocity weight at the beginning 0.95

US
Value of velocity Weight at the end of the PSO iterations 0.4
The fraction of maximum iterations, for which W is linearly varied 0.7
Maximum velocity step, Constriction factor and Neighborhood size 1
AN
Value of Global Minima 0

Population size 20
Maximum no. of iteration 500
M

RT coefficient 2
WDO Gravitational constant (g) 0.2
Constant in the update equation 0.5
ED

Coriolis effect coefficient (c) 0.4

Maximum allowed speed or velocity limit 0.3
PT

Number of nests 25
No. of iterations 3000
CS Step size (α) 1
CE

Mutation probability value (pa) 0.25

Scale factor (β) 1.5
AC

24
 ACCEPTED MANUSCRIPT

Table 2: Comparison of best objective values (entropy), SSIM and FSIM computed by DE, WDO, PSO and CS algorithm using Kapur’s entropy.

Kapur’s Entropy (Objective Function Value, SSIM & FSIM)

m Best objective values SSIM FSIM

Images
Test

DE WDO PSO CS DE WDO PSO CS DE WDO PSO CS

6 23.81754 24.17661 24.00123 24.23789 0.97062 0.97333 0.97179 0.97662 0.95067 0.93057 0.95037 0.95512
1 8 28.65868 28.22878 29.03273 29.23067 0.98048 0.98196 0.98313 0.98508 0.96032 0.95679 0.96049 0.96471
12 36.51251 36.14086 37.77768 37.87890 0.99420 0.99108 0.99145 0.99506 0.97122 0.96516 0.97168 0.97273
16 41.67917 42.34675 45.13322 45.27020 0.99512 0.98917 0.99727 0.99774 0.98600 0.96161 0.99265 0.99372
6 23.19584 24.17142 24.35919 24.44928 0.97286 0.97415 0.98109 0.98834 0.97075 0.96232 0.98530 0.97906

T
2 8 28.58406 28.68227 28.66316 29.42294 0.99055 0.98757 0.98222 0.99294 0.98624 0.97939 0.98094 0.98707
12 36.22494 36.25214 37.30920 38.02897 0.99413 0.98010 0.99379 0.99660 0.98750 0.96826 0.99074 0.99398

IP
16 41.23897 42.27089 45.63157 45.63968 0.99599 0.99095 0.99516 0.99800 0.99375 0.98497 0.99445 0.99656
6 23.72005 23.66865 23.96841 24.03465 0.95548 0.96392 0.97153 0.97690 0.92170 0.94031 0.94484 0.94456
3 8 27.88440 27.98234 28.69529 29.43079 0.96896 0.97975 0.98765 0.98815 0.95791 0.96163 0.95818 0.96086
12 33.47781 36.12154 37.31493 37.34753 0.99036 0.98919 0.99211 0.99281 0.97207 0.98077 0.97451 0.98383

CR
16 41.08386 41.28704 44.40156 44.44916 0.99204 0.99341 0.99183 0.99745 0.98039 0.98786 0.99071 0.99406
6 22.28589 22.84154 22.24289 23.49527 0.98338 0.96674 0.97278 0.98265 0.95915 0.92283 0.93260 0.96162
4 8 27.49287 28.05434 26.69044 28.16738 0.98552 0.98393 0.98205 0.99120 0.95451 0.93583 0.95516 0.96135
12 33.55422 34.70889 34.66046 36.61275 0.99103 0.98628 0.99105 0.99242 0.97031 0.95155 0.97447 0.97549
16 41.24827 41.99895 41.25177 43.71182 0.99143 0.99256 0.99392 0.99661 0.97089 0.97041 0.98095 0.98643

US
6 23.35011 23.61419 23.58502 23.69380 0.98676 0.97344 0.97464 0.97807 0.93625 0.91055 0.91710 0.91299
5 8 28.05476 28.11991 28.57064 28.61105 0.99132 0.98052 0.98082 0.99088 0.94263 0.93637 0.95264 0.95375
12 35.82048 35.92937 37.25578 37.59843 0.99471 0.98790 0.99271 0.99543 0.96605 0.94256 0.97186 0.97837
16 42.40405 42.09759 44.73003 44.94067 0.99631 0.99188 0.99241 0.99761 0.98265 0.96572 0.98840 0.98953
6 23.90171 23.89492 23.97392 24.17052 0.98601 0.97399 0.98075 0.98135 0.94589 0.92726 0.92342 0.92612
6 8 28.43246 28.42510 28.90414 29.09930 0.99040 0.98257 0.99023 0.99073 0.96107 0.95266 0.96014 0.96350
AN
12 35.71635 35.94265 37.42972 37.69790 0.99418 0.98749 0.99589 0.99608 0.98021 0.96648 0.98462 0.98676
16 42.20033 41.96075 44.80418 45.00839 0.99587 0.99380 0.99650 0.99757 0.98575 0.97198 0.98996 0.99225
6 24.40915 24.74322 24.84637 24.87463 0.98561 0.97278 0.98753 0.98753 0.97712 0.93020 0.97857 0.97857
7 8 29.23968 29.33921 29.81024 29.82061 0.99123 0.98367 0.99240 0.99240 0.98124 0.96906 0.98670 0.98670
12 36.47080 36.80882 38.40419 38.41298 0.99433 0.98616 0.99633 0.99635 0.98554 0.97396 0.99365 0.99366
M

16 42.36018 43.48243 45.77859 45.78455 0.99598 0.99133 0.99781 0.99798 0.99153 0.97756 0.99614 0.99629
6 23.48762 23.59425 23.61006 23.91120 0.96305 0.96052 0.97454 0.98417 0.98617 0.90200 0.94162 0.94386
8 8 27.94529 28.32794 28.45202 29.08578 0.99055 0.98464 0.98099 0.98963 0.97489 0.96609 0.97050 0.97588
12 36.19108 35.48884 37.02756 37.42725 0.99460 0.98626 0.99591 0.99592 0.98903 0.96819 0.98879 0.98866
16 42.22627 41.08056 44.27310 44.69970 0.99659 0.99281 0.99750 0.99759 0.99149 0.97602 0.99363 0.99369
ED

6 23.54834 23.77179 24.03868 23.89147 0.97869 0.97390 0.97868 0.98968 0.94233 0.93422 0.96391 0.96391
9 8 26.57448 28.27914 28.81822 28.81924 0.98285 0.98675 0.98459 0.99459 0.95694 0.95504 0.97886 0.97886
12 35.51241 35.19261 37.16698 37.82642 0.98995 0.98715 0.99248 0.99745 0.97289 0.96791 0.98822 0.98802
16 40.66756 40.63702 44.24569 44.82529 0.99142 0.99164 0.99784 0.99847 0.97850 0.97641 0.99450 0.99234
6 23.28969 23.56912 22.92039 23.75823 0.97674 0.97197 0.94636 0.97582 0.93879 0.93174 0.89620 0.91278
10 8 27.53307 27.98776 27.38411 28.42332 0.98758 0.98150 0.98123 0.99133 0.96295 0.95503 0.97546 0.97583
PT

12 3.150067 34.31754 35.46899 36.67403 0.99246 0.99031 0.99602 0.99606 0.97446 0.96910 0.98925 0.98968
16 42.99926 41.95654 42.32682 43.59820 0.99110 0.99223 0.99769 0.99784 0.98021 0.97207 0.99474 0.99508
CE
AC

25
 ACCEPTED MANUSCRIPT

Table 3: Comparison of CPU Timing, PSNR and MSE computed by DE, WDO, PSO and CS algorithm using Kapur’s entropy.

Kapur’s Entropy (CPU Timing, PSNR & MSE)

m CPU Timing PSNR MSE

Images
Test

DE WDO PSO CS DE WDO PSO CS DE WDO PSO CS

6 6.97856 5.74182 55.8310 12.61359 25.40692 25.41014 25.53303 25.50413 1.87263 1.87195 1.81881 1.81116
1 8 7.17302 6.17163 60.0937 13.26601 26.03864 25.82255 26.01631 26.74175 1.61932 1.70167 1.82797 1.60540
12 7.80171 7.25083 69.9937 14.83862 26.98317 26.55704 27.23283 27.29052 1.30269 1.43741 1.32979 1.24183
16 8.79764 7.62185 76.5308 15.04575 27.76023 26.74186 28.46390 28.92692 1.09222 1.37737 93.2181 1.04274
6 6.68659 15.0345 55.3250 12.69407 25.23914 25.45664 25.53090 25.53539 1.81629 1.85110 1.82000 1.80037

T
2 8 6.71746 15.8701 58.5189 12.32620 25.90243 25.86779 26.02786 26.96624 1.68721 1.68399 1.62303 1.61612
12 6.75998 18.2126 65.8625 13.52117 26.94672 26.29577 27.33694 27.40058 1.31570 1.52679 1.25068 1.23901

IP
16 6.81671 19.3905 76.8907 14.83220 27.93116 26.95225 29.16506 29.93135 1.05229 1.31358 78.8211 1.00955
6 6.55278 16.5858 57.5810 16.33355 25.29609 25.30256 25.46895 25.47551 1.92158 1.91827 1.84601 1.82676
3 8 8.16258 17.4583 60.8638 18.34725 25.81468 25.68568 26.09108 26.84948 1.76619 1.75619 1.59953 1.59106
12 8.17028 20.2984 71.0286 17.79697 26.66158 26.51244 27.12255 27.20061 1.40415 1.45293 1.26213 1.24216

CR
16 8.60884 21.2001 79.9529 18.97865 27.43407 26.95669 28.74155 29.28023 1.18216 1.31056 86.9151 1.17924
6 6.69784 17.1429 59.8387 13.00859 25.42834 25.25911 25.50979 25.56873 1.86377 1.94050 1.82888 1.81617
4 8 6.99232 17.2552 67.2097 13.09878 25.80484 25.71114 25.98040 26.00518 1.70887 1.74700 1.64089 1.63496
12 7.51910 18.4222 78.3678 14.62582 26.63444 26.45754 27.37336 27.88949 1.41414 1.47200 1.19131 1.16306
16 7.52190 20.0393 80.2362 15.59520 26.84912 27.02243 28.61421 28.86526 1.34524 1.29419 89.7784 1.15517

US
6 6.63518 15.6369 54.5825 12.63928 25.47771 25.30525 25.49079 25.52336 1.84214 1.91690 1.83659 1.82590
5 8 6.80804 15.0508 60.0364 13.04159 26.03599 25.85335 25.87351 25.88332 1.62016 1.69036 1.67832 1.67797
12 6.90241 17.9120 68.1750 14.02908 27.00655 26.46849 27.26042 27.31760 1.29569 1.46964 1.22220 1.20324
16 6.92718 21.2350 74.9166 15.46030 27.99169 27.05381 28.88986 29.20103 1.23485 1.28417 84.2133 1.13997
6 6.70890 16.1020 64.1220 13.43390 25.55429 25.27590 25.29407 25.32554 1.81008 1.93044 1.92244 1.90818
6 8 7.33715 16.1708 70.1906 13.67942 26.03068 25.75083 25.92226 26.87204 1.62254 1.73014 1.66301 1.60290
AN
12 7.57675 17.5525 78.6613 14.27918 26.92128 26.45954 27.27129 27.30151 1.32199 1.47024 1.21995 1.21073
16 7.90855 22.2338 86.5904 15.86256 27.88111 27.53493 28.09958 28.98735 1.06037 1.14972 80.0580 1.01938
6 6.67061 14.8163 65.0406 13.69187 25.58726 25.42333 25.48931 25.48931 1.79687 1.86577 1.83719 1.83719
7 8 7.06479 15.7052 73.9154 14.15294 26.06157 25.88218 25.99252 26.99252 1.61059 1.67828 1.63620 1.60620
12 7.17424 17.4176 76.6022 14.84791 26.87958 26.35431 27.22812 27.24647 1.33577 1.50639 1.23114 1.22590
M

16 7.60782 18.7353 86.0012 15.22741 27.69962 27.10316 28.93144 29.28754 1.10604 1.27025 83.2262 1.09962
6 6.40782 15.7175 66.7514 13.71741 25.43606 25.24661 25.31243 25.37184 1.85989 1.94442 1.88409 1.86939
8 8 6.44316 16.2005 70.5169 14.21184 25.89035 25.67093 25.97514 26.76322 1.67562 1.76196 1.64375 1.62493
12 6.86572 18.0864 78.2263 15.05961 27.10275 26.19240 27.27013 27.28606 1.26717 1.56412 1.21930 1.21480
16 6.95938 19.4305 84.2423 15.12782 27.84460 26.95278 28.91239 29.03822 1.06831 1.31254 83.7284 1.05903
ED

6 6.65501 16.3091 64.3487 12.92111 25.17994 25.37041 25.27626 25.27626 1.88462 1.88838 1.93014 1.83014
9 8 6.89657 16.6485 71.6599 13.62258 25.90310 25.86947 25.80367 25.80367 1.67156 1.68356 1.70935 1.70935
12 7.57478 17.8093 80.3667 14.10515 26.59477 26.41022 27.06496 27.07813 1.42721 1.49109 1.27832 1.27637
16 7.65133 18.5611 90.9008 15.55444 27.31152 26.88982 28.65746 29.08395 1.21690 1.33138 88.9434 1.28611
6 6.48748 15.1280 65.5174 13.65042 25.47442 25.25963 25.29120 25.44887 1.84388 1.93746 1.92297 1.83512
10 8 6.95040 15.7345 67.3622 14.05377 26.05081 25.48321 26.03131 26.13783 1.61449 1.84012 1.62399 1.60145
PT

12 6.98030 18.2370 75.5676 14.51475 26.91390 26.38028 27.06806 27.36800 1.32522 1.49670 1.27765 1.19314
16 7.01416 19.8932 83.9983 15.31718 27.19550 26.99890 28.79169 29.48820 1.25084 1.29875 85.9051 1.20989
CE
AC

26
 ACCEPTED MANUSCRIPT

Table 4: Comparison of best objective values (entropy), SSIM and FSIM computed by DE, WDO, PSO and CS algorithm using between-class
Variance.

Between-class Variance (Objective Function Value, SSIM & FSIM)

m Best objective values SSIM FSIM

Images
Test

DE WDO PSO CS DE WDO PSO CS DE WDO PSO CS

6 0.9549 0.9501 0.9416 0.9603 0.9789 0.9713 0.9269 0.9746 0.9056 0.9111 0.9269 0.9333
1 8 0.9772 0.9847 0.9869 0.9998 0.9818 0.9888 0.9772 0.9898 0.9247 0.9270 0.9346 0.9405
12 1.0074 1.0024 1.0073 1.0095 0.9873 0.9858 0.9804 0.9930 0.9351 0.9439 0.9444 0.9488
16 1.0334 1.0345 1.0057 1.0477 0.9936 0.9939 0.9910 0.9954 0.9757 0.9757 0.9692 0.9798

T
6 0.9674 0.9625 0.9592 0.9634 0.9721 0.9680 0.9757 0.9792 0.9590 0.9109 0.9646 0.9701
2 8 0.9938 0.9934 0.9915 0.9952 0.9843 0.9847 0.9826 0.9897 0.9715 0.9742 0.9732 0.9761

IP
12 1.0069 1.0029 1.0075 1.0225 0.9905 0.9872 0.9870 0.9937 0.9838 0.9790 0.9781 0.9813
16 1.0265 1.0272 1.0292 1.0503 0.9920 0.9920 0.9930 0.9951 0.9843 0.9862 0.9930 0.9867
6 0.9589 0.9588 0.9504 0.9597 0.9736 0.9802 0.9643 0.9834 0.9110 0.9318 0.9339 0.9718
3 8 0.9649 0.9852 0.9851 0.9948 0.9826 0.9884 0.9741 0.9887 0.9547 0.9517 0.9554 0.9727

CR
12 0.9842 1.0280 0.9997 1.0293 0.9903 0.9874 0.9864 0.9910 0.9714 0.9777 0.9718 0.9820
16 1.0166 1.0438 1.0327 1.0551 0.9941 0.9903 0.9932 0.9946 0.9874 0.9793 0.9865 0.9859
6 0.9138 0.9572 0.9506 0.9582 0.9733 0.9806 0.9502 0.9814 0.9035 0.9165 0.8899 0.9486
4 8 0.9707 0.9819 0.9744 0.9938 0.9855 0.9854 0.9749 0.9867 0.9430 0.9485 0.9128 0.9505
12 0.9781 1.0124 1.0129 1.0276 0.9930 0.9915 0.9869 0.9881 0.9779 0.9708 0.9607 0.9726

5
16
6
8
12
16
1.0149
0.9903
1.0399
1.0089
1.0190
1.0534
0.9534
0.9867
1.0031
1.0241
1.0800
0.9534
0.9930
1.0040
1.0486
1.0537
0.9713
0.9982
1.0468
1.0821
0.9937
0.9785
0.9860
0.9874
0.9938
US
0.9919
0.9819
0.9837
0.9894
0.9944
0.9912
0.9611
0.9728
0.9894
0.9918
0.9948
0.9823
0.9854
0.9897
0.9949
0.9784
0.9179
0.9349
0.9416
0.9645
0.9784
0.9273
0.9268
0.9452
0.9596
0.9678
0.8960
0.8974
0.9496
0.9621
0.9818
0.9318
0.9554
0.9503
0.9821
AN
6 0.9688 0.9564 0.9648 0.9654 0.9842 0.9766 0.9590 0.9801 0.9328 0.9328 0.9031 0.9352
6 8 0.9838 0.9814 0.9930 0.9949 0.9868 0.9790 0.9830 0.9830 0.9429 0.9422 0.9530 0.9546
12 1.0073 1.0372 1.0161 1.0486 0.9907 0.9895 0.9897 0.9915 0.9481 0.9598 0.9602 0.9749
16 1.0225 1.0778 1.0463 1.0896 0.9926 0.9905 0.9910 0.9916 0.9742 0.9679 0.9641 0.9846
6 0.9456 0.9612 0.9613 0.9670 0.9754 0.9818 0.9663 0.9840 0.9424 0.9669 0.9485 0.9613
7 8 0.9872 0.9978 0.9985 0.9994 0.9882 0.9848 0.9791 0.9884 0.9689 0.9719 0.9657 0.9738
M

12 1.0162 1.0290 1.0289 1.0294 0.9911 0.9906 0.9822 0.9924 0.9833 0.9795 0.9708 0.9847
16 1.0535 1.0574 1.0914 1.0980 0.9951 0.9922 0.9903 0.9949 0.9826 0.9849 0.9828 0.9880
6 0.9610 0.9546 0.9551 0.9626 0.9784 0.9720 0.9730 0.9773 0.9261 0.9535 0.9440 0.9540
8 8 0.9273 0.9917 0.9928 0.9980 0.9826 0.9823 0.9753 0.9921 0.9454 0.9519 0.9440 0.9497
12 1.0057 1.0191 1.0191 1.0199 0.9892 0.9848 0.9808 0.9876 0.9553 0.9533 0.9581 0.9663
ED

16 1.0182 1.0574 1.0587 1.0592 0.9921 0.9924 0.9915 0.9937 0.9797 0.9797 0.9755 0.9832
6 0.9825 0.9573 0.9401 0.9888 0.9702 0.9718 0.9645 0.9741 0.9281 0.9461 0.8980 0.9488
9 8 1.0154 1.0108 1.0104 1.0151 0.9801 0.9809 0.9790 0.9834 0.9493 0.9414 0.9320 0.9647
12 1.0286 1.0415 1.0183 1.0454 0.9882 0.9886 0.9880 0.9902 0.9677 0.9696 0.9584 0.9731
16 1.0720 1.0632 1.0416 1.0847 0.9907 0.9896 0.9908 0.9911 0.9784 0.9717 0.9792 0.9791
PT

6 0.9743 0.9645 0.9411 0.9718 0.9729 0.9710 0.9763 0.9732 0.9157 0.9409 0.9287 0.9513
10 8 0.9971 0.9961 0.9909 0.9991 0.9803 0.9820 0.9789 0.9864 0.9391 0.9543 0.9385 0.9637
12 1.0296 1.0610 1.0358 1.0626 0.9871 0.9897 0.9867 0.9899 0.9684 0.9682 0.9498 0.9686
16 1.0534 1.0899 1.0706 1.0891 0.9931 0.9916 0.9923 0.9951 0.9728 0.9762 0.9801 0.9819
CE
AC

27
 ACCEPTED MANUSCRIPT

Table 5: Comparison of CPU Timing, PSNR and MSE computed by DE, WDO, PSO and CS algorithm using between-class Variance.

Between-class Variance (CPU Timing, PSNR & MSE)

m CPU Timing PSNR MSE

Images
Test

DE WDO PSO CS DE WDO PSO CS DE WDO PSO CS

6 5.53745 16.6068 11.4093 9.06883 25.4860 25.4308 25.0591 25.6997 1.83853 186.206 2.02872 1.75031
1 8 5.55494 15.9035 12.1618 9.78569 25.8322 25.7520 25.7533 26.7461 1.69830 173.106 1.72889 1.38719
12 5.62148 18.2205 12.4947 11.2108 26.3143 26.2243 25.9474 27.6674 1.52055 155.429 1.65598 1.11751
16 5.72966 20.6132 14.1489 12.6887 26.9903 27.1191 27.1072 28.0871 1.30148 126.886 1.26680 1.01993
6 5.74030 16.0232 11.2920 9.15101 25.5108 25.3998 25.3993 25.5124 1.82822 187.549 1.87563 182.831

T
2 8 5.76271 17.2846 12.5135 9.74270 25.8342 25.6951 25.8793 25.8322 1.69690 175.234 1.67940 169.867
12 5.78688 18.9034 14.0243 11.9264 26.3564 26.1614 26.5700 26.9580 1.50468 157.514 1.43253 131.053

IP
16 5.80453 21.9816 14.5886 12.3549 26.9707 26.9588 27.1730 27.5103 1.31016 131.290 1.25442 116.011
6 6.64540 17.1841 14.3751 12.0213 25.5250 25.4482 14.3751 25.0354 182.219 185.519 1.99309 169.509
3 8 6.78160 19.1799 15.0326 11.6691 25.6813 25.6491 25.6131 25.9061 175.950 177.296 1.78650 166.970
12 6.97372 26.7830 14.6446 13.4068 26.6281 26.4559 26.6031 26.7372 141.366 147.176 1.42346 102.212

CR
16 6.98405 26.1110 16.3718 14.9736 27.0898 26.7351 27.2522 27.1154 127.285 138.587 1.22535 127.174
6 6.01629 15.6261 14.7407 9.05395 25.4149 25.5085 25.3786 25.3972 187.099 182.911 1.88496 150.085
4 8 6.19861 17.2024 15.1578 10.1915 25.9423 25.8416 25.6511 26.7781 165.566 169.451 1.77099 137.202
12 6.58525 17.8172 15.6366 11.0489 26.6595 26.5896 26.4022 27.0501 140.463 142.729 1.49165 130.900
16 6.75414 23.4362 15.6764 13.1043 26.6607 27.1556 27.0716 27.9360 141.053 125.312 1.27886 109.394

US
6 4.98244 18.0654 12.3007 9.43104 25.2447 25.4205 25.2855 25.4081 194.409 186.684 1.92699 187.190
5 8 4.99663 18.8046 16.1929 10.6313 25.6751 25.7711 25.7180 25.6232 176.094 172.177 1.74319 178.416
12 5.04881 19.9066 16.3495 11.7710 26.0129 26.4524 26.3652 26.4614 162.925 147.233 1.50194 147.483
16 5.11188 24.7529 16.5321 13.3151 26.9718 27.3550 26.9690 27.0743 130.956 119.594 1.30981 127.897
6 4.18725 17.8132 11.9414 9.75393 25.4996 25.7938 25.3705 25.6776 183.299 171.387 1.88913 182.543
6 8 5.20719 18.7200 12.7905 9.79720 25.9377 26.2456 25.6495 25.8933 165.745 154.851 1.77094 167.872
AN
12 5.64749 20.0225 13.7776 11.4553 26.5116 26.4627 26.6174 27.1667 145.292 146.844 1.41947 127.191
16 5.69969 22.8381 13.8616 14.9578 26.8355 26.8312 26.8999 27.4430 135.173 135.169 1.33245 120.483
6 5.61342 16.2379 11.7382 9.38076 25.2743 25.5128 25.4051 25.6738 193.207 182.778 1.87331 176.305
7 8 5.68231 16.8358 12.3423 10.2070 25.7513 25.9486 25.6360 25.9720 172.967 165.368 1.77729 164.860
12 5.72612 23.7229 13.4117 12.9486 26.2354 26.3420 26.2152 26.4742 154.812 151.286 1.56211 146.825
M

16 5.74512 22.5321 14.2157 12.7772 27.3911 26.8000 26.9379 27.3064 119.000 136.060 1.31869 121.490
6 5.84869 18.1705 11.3835 9.82659 25.5290 25.3216 25.3919 25.7188 182.150 190.982 1.88011 174.425
8 8 5.91341 18.6456 12.1015 10.2861 25.7500 25.7397 25.6293 26.4255 173.059 173.722 1.78061 148.497
12 5.94924 21.9950 13.1734 11.4411 26.3221 26.4018 26.2138 28.0761 151.672 148.947 1.56250 102.749
16 5.96756 26.0290 13.5164 13.0421 27.1929 27.2215 26.8853 28.4424 124.392 123.583 1.33691 93.3592
ED

6 5.12800 18.8921 11.1912 9.45923 25.3886 25.4759 25.2724 25.9373 188.101 184.556 1.93147 165.847
9 8 5.32123 20.9860 11.9810 10.9076 25.5614 25.6086 25.7278 26.0376 181.294 179.033 1.73970 163.910
12 5.43793 21.2725 12.5277 11.5749 26.2485 26.6083 26.4025 27.1468 154.325 142.368 1.49082 127.494
16 5.68653 26.1447 13.1514 13.0935 27.0720 26.8177 26.7898 27.1468 127.946 135.500 1.36356 126.773
6 5.07118 19.3031 11.2243 9.41981 25.4940 25.4505 25.3596 25.9978 183.593 185.392 1.89312 164.922
10 8 5.16691 21.5338 11.4745 10.3549 25.7928 25.5757 25.6514 26.0104 171.372 180.189 1.77032 162.976
PT

12 5.72842 22.3219 12.3398 12.2584 26.2846 26.5421 26.3299 27.1249 153.306 144.200 1.51471 126.222
16 5.79747 23.0515 13.6116 13.8646 26.9369 27.1295 27.2539 27.8580 131.879 127.005 1.22504 106.499
CE
AC

28
 ACCEPTED MANUSCRIPT

Thresholded Image
Satellite Image into k levels
Segmented Image
Easy analysis
Simple Interpretation
Exposed features
Compressed data

T
Consume less memory

IP
Maximize entropy

CR
US
AN
Enhanced Image Multilevel Entropy Nature Inspired
Calculation Optimization Techniques
M

Fig. 1. A complete flowchart routine of optimal color-image thresholding.

ED
PT
CE
AC

29
 ACCEPTED MANUSCRIPT

(a1) (a2) (a3) (a4) (a5)

T
IP
CR
(b1) (b2) (b3) (b4) (b5)

US
AN
(c1) (c2) (c3) (c4) (c5)
M
ED

(a6) (a7) (a8) (a9) (a10)

PT
CE

(b6) (b7) (b8) (b9) (b10)

AC

(c6) (c7) (c8) (c9) (c10)

Fig. 2. (a1-a10) represent original satellite images, (b1-b10) illustrate corresponding histogram image and (c1-c10) represent the histogram
equalized images for segmentation purpose.

30
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
2

US
AN

3
M
ED
PT

4
CE
AC

(Continued)

31
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
7

US
AN

8
M
ED
PT

9
CE
AC

10

Fig. 3. The 6-levels thresholded test images obtained by Kapur’s entropy with different optimization techniques.

32
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
2

US
AN

3
M
ED
PT

4
CE
AC

(Continued)

33
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
7

US
AN
M

8
ED
PT

9
CE
AC

10

Fig. 4. The 6-levels thresholded test images obtained by between-class-variance with different optimization techniques.

34
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
2

US
AN

3
M
ED
PT

4
CE
AC

(Continued)

35
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
7

US
AN

8
M
ED
PT

9
CE
AC

10

Fig. 5. The 8-levels thresholded test images obtained by Kapur’s entropy with different optimization techniques.

36
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
2

US
AN

3
M
ED
PT

4
CE
AC

(Continued)

37
 ACCEPTED MANUSCRIPT

DE WDO PSO CS

T
IP
CR
7

US
AN

8
M
ED
PT

9
CE
AC

10

Fig. 6. The 8-levels thresholded test images obtained by between-class-variance with different optimization techniques.

38
 ACCEPTED MANUSCRIPT

DE WDO PSO CS
12-Levels

T
IP
CR
16-Levels

US
Fig. 7. The thresholded images for the test image-6 with 12 and 16 levels using Kapur’s entropy.
AN
DE WDO PSO CS
M
12-Levels

ED
PT
16-Levels

CE
AC

Fig. 8. The thresholded images for the test image-6 with 12 and 16 levels using between-class-variance.

39