improved clustering criterion for image clustering with artificial bee colony algorithm
TRANSCRIPT
THEORETICAL ADVANCES
Improved clustering criterion for image clustering with artificialbee colony algorithm
Celal Ozturk • Emrah Hancer • Dervis Karaboga
Received: 23 November 2012 / Accepted: 27 January 2014
� Springer-Verlag London 2014
Abstract In this paper, a new objective function is pro-
posed for image clustering and is applied with the artificial
bee colony (ABC) algorithm, the particle swarm optimi-
zation algorithm and the genetic algorithm. The perfor-
mance of the proposed objective function is tested on seven
benchmark images by comparing it with the three well-
known objective functions in the literature and the
K-means algorithm in terms of separateness and com-
pactness which are the main criterions of the clustering
problem. Moreover, the Davies–Bouldin Index and the XB
Index are also employed to compare the quality of the
proposed objective function with the other objective
functions. The simulated results show that the ABC-based
image clustering method with the improved objective
function obtains well-distinguished clusters.
Keywords Image clustering � Artificial bee colony
algorithm � Genetic algorithms � K-means �Particle swarm optimization � Validity indexes
1 Introduction
It is known that a large-scale data set needs to be stored on
account of high development in data acquisition, in com-
puter performances, in technological devices and on
account of requirement for detail information. For instance,
highly improved scanners are used to minimize data loss in
order to accomplish successful raster to vector conversa-
tion of topographic maps [1, 2]. Therefore, the requirement
of storing data plays important role in many areas. How-
ever, some pre-processing implementations are needed to
distinguish significant information in this type of data sets
before processing and analysing. Data mining, one of the
young and interdisciplinary fields of computer science,
simply aims to extract knowledge from data to overcome
difficulties of analysing data. Data mining involves various
approaches like clustering, data summarization, classifica-
tion, finding dependency networks, regression, analysing
changes, detecting anomalies, etc. [3]. This paper concen-
trates on clustering, specifically on image clustering
approach, which is considered as an unsupervised learning
problem.
Clustering is the task of separating groups or clusters
based on some similar measures like distance, intervals
within multidimensional data [4]. Data members have
maximum similarities among all data members assigned to
one group or one cluster. A cluster is generally distin-
guished by its cluster centre, namely centroid [5]. Due to its
crucial role, clustering is used in many applications,
including machine learning, image analysis, pattern rec-
ognition and statistical data analysis.
Most clustering algorithms are based on two popular
techniques: hierarchical and partitional clustering. In hier-
archical clustering algorithms, the output is ‘‘a tree show-
ing a sequence of clustering, each of which is a partition of
the data set [6]’’. Hierarchical clustering algorithms can be
categorized into two groups: divisive or agglomerative.
Agglomerative hierarchical clustering algorithms (bottom
up approaches) start with each data member as a separate
cluster and the most similar clusters are merged into larger
C. Ozturk � E. Hancer (&) � D. Karaboga
Engineering Faculty, Computer Engineering Department,
Erciyes University, Kayseri, Turkey
e-mail: [email protected]
C. Ozturk
e-mail: [email protected]
D. Karaboga
e-mail: [email protected]
123
Pattern Anal Applic
DOI 10.1007/s10044-014-0365-y
clusters. The most known agglomerative algorithms are
single link and complete link. Single link algorithms are
based on merging two clusters which have the smallest
maximum pairwise distance. On the other hand, complete
link algorithms are based on merging two clusters which
have the smallest minimum pairwise distance [7]. Hierar-
chical algorithms have the following advantages: the
number of clusters does not need to be specified and they
are not very dependent to initial conditions [8]. However,
they are not dynamic, i.e. after a data point assigned to a
cluster, its cluster cannot be modified, and the lack of
information about global size and shape can cause over-
lapping clusters [9]. The other disadvantage of hierarchical
algorithms is the time complexity (O(Np2 log Np)), i.e. time
computation is expensive.
Partitional clustering attempts to decompose the data set
into a set of disjoint clusters by generating various partitions.
The algorithms evaluate the partitions by some criterions
which are tried to be minimized by assigning clusters to
peaks in the probability density function or the global
structure [10]. Typically the global criteria involve mini-
mizing a measure of dissimilarity in the samples within each
cluster while maximizing the dissimilarity of different
clusters. Since a criterion (e.g. square error function) is tried
to be minimized, the partitional clustering algorithms can be
regarded as optimization problems [11]. In the usage of the
partitional clustering algorithms, the disadvantages of the
hierarchical algorithms are the advantages of the partitional
clustering algorithms, and vice versa [11, 12].
The most widely used partitional clustering algorithm
developed by MacQueen is known as K-means [13]. In
K-means, K centroids are first initialized with the help of
random number generator or data information. After that,
each pattern is assigned to its corresponding closest centroid
based on similarity measure, Euclidean distance. Finally, the
positions of centroids are recalculated. These steps are
repeated until convergence. K-means algorithm is simple to
implement and is convenient for large data sets on account of
its time complexity (O(Np)). However, the algorithm is data
dependent and depends on the initial conditions, which cause
the algorithm to converge to the nearest local optimum from
the starting position of the search [14, 15]. The fuzzy version
of K-means algorithm, developed by Dunn [16] and
improved by Bezdek [17], is the Fuzzy C-means (FCM)
algorithm. FCM aims to minimize fuzzy version of the least-
square error criterion and is less influenced by existence of
uncertainty in the data set [18]. In this way, FCM performs
better than K-means. However, it might also converge to
local minima and it is highly sensitive to noise and imaging
artefacts as in K-means [19]. Researchers have developed
new versions of FCM by modifying equations based on
objective functions such as possibilistic C-means (PCM)
[20], robust fuzzy C-means (RFCM) [21], generalized FCM
(GFCM) [22], etc. K-means also has some other versions:
K-harmonic means (KHM) [23], spatial kernel-based KHM
(SKKHM) [24], etc.
Sorting out local minima problem of clustering process
is another demanded subject for researchers. To overcome
local minima problem, evolutionary algorithms, particu-
larly swarm-based algorithms, were employed in recent
years. Particle swarm optimization (PSO), proposed by
Kennedy and Eberhart [25, 26], is a population-based
approach inspired by social behaviours of bird flicks and
fish swarms [27]. It has been widely applied to image
clustering. The first PSO-based image clustering algorithm
is proposed by Omran et al. [28]. The other PSO-based
image clustering studies of Omran et al. can be seen in [29–
31]. Ant colony optimization (ACO) [32] is based on the
behaviours of ants searching a path between their colony
and a source of food, also applied to image clustering [3,
33]. The other applied algorithm to the clustering problems
is the differential evolution (DE) algorithm with the mod-
ifications [34–37]. Artificial bee colony (ABC) algorithm,
introduced by Karaboga [38], is known to be a new swarm
intelligence-based evolutionary algorithm. The model of
the ABC algorithm is inspired by the intelligent behaviours
of honeybee swarm and it has been applied to many
problems such as optimization of numerical problems [39],
data clustering [14], neural networks training for pattern
recognition [40], image analysis [41], and wireless sensor
network deployment [42] and routing [43]. The ABC
algorithm was also applied to the image clustering problem
in Refs. [44, 45]. However, in these works [44, 45] the
objective fitness function was only based on quantization
error which was not sufficient to satisfy the simple prop-
erties of the clustering process: compactness and sepa-
rateness. For instance, the maximum distance between the
centroids is generally very closest in Ref. [44] which
cannot be acceptable for the separateness property. On the
other hand, Hancer et al. [46] applied the ABC and PSO
algorithms to the clustering problem with the widely used
objective fitness function, which was also used in the
studies of Omran et al. [29–31]. Yet, the ABC and PSO
algorithms with this objective function could not obtain
well-minimized quantization error. Therefore, a new
objective function is proposed to obtain well-separated and
well-compacted clusters in this study. The effectiveness of
the proposed objective function is verified on seven
benchmark images (airplane, house, Lena, Morro Bay,
MRI, 42049 and 48025).
The rest of the paper organized as follows: Sect. 2
describes the ABC algorithm briefly; Sect. 3 describes the
clustering problem, objective functions and validity
indexes; Sect. 4 summarizes the results and comparisons of
the ABC, K-means, PSO and GA algorithms; finally, Sect.
5 concludes the paper.
Pattern Anal Applic
123
2 Artificial bee colony algorithm
As in the colony of real honey bees, the colony of real
artificial bees consists of three kinds of bees: employed
bees, onlooker bees and scout bees. A bee associated with a
food resource is called employed bee, a bee waiting on
dance area to get information concerning food sources is
called onlooker bee, and a bee performing random search is
called scout. The area of food sources denotes the possible
solutions of optimization problem and the nectar amount of
a food source represents the quality of relevant solution. It is
assumed that only one employed bee is associated with one
food source. Therefore, the number of employed bees is
equal to the number of food sources which are equal to half
of the colony size. After randomly generation of food
sources, each employed bee searches a new position in the
neighborhood of its associated food source. The employed
bee keeps its new food source position when the nec-
tar amount of new food source is more abundant than the
position of old one (greedy selection). After all employed
bees complete their search process, they share the infor-
mation belonging to nectar amount, direction and distance
of food sources with onlooker bees by waggle dance in
dancing area. Each onlooker bee selects a food source
probabilistically and searches for a new food source and
evaluates its quality. Then, greedy selection is applied same
as in the implementation of employed bees. After that, the
position of each food source is checked whether it is
improved or not. If a position of food source cannot be
improved for a number of predetermined trials (limit), that
food source is assigned as abandoned and a scout randomly
produces a new food source position for the abandoned one.
After the food production, scout bee becomes an employed
bee. The flow chart of ABC algorithm can be seen in Fig. 1
and the main steps of ABC algorithm [47] can be defined as
1. The production of initial food sources (x0is) and
evaluation of population by Eq. (1)
xij ¼ xminj þ rand 0; 1ð Þ xmax
j � xminj
� �ð1Þ
where i ¼ 1; . . .; SN and SN is the number of food sources;
j ¼ 1; . . .;D; and D is the number of parameters; xjmin is the
minimum and xjmax is the maximum values of parameter j.
2. Produce new food source in the neighbourhood of xi by
Eq. (2)
mij ¼ xij þ /ij xij � xkj
� �ð2Þ
where j and k are randomly chosen parameters and
neighbourhoods, respectively, and /ij is a random number
within [-1,1].
3. Compute the fitness value of mi, if mi is better than xi,
employed bee memorizes mi and removes xi from its
memory.Fig. 1 Flow chart of ABC algorithm
Pattern Anal Applic
123
4. The probability of pi is calculated by Eqs. (3) and (4).
pi ¼fitnessiPSNj¼1 fitnessi
ð3Þ
fitnessi ¼1
1þ fiti
; fiti� 0
1þ abs fitið Þ; fiti\0
8<: ð4Þ
where fiti is the fitness value of source xi.
5. Select food sources probabilistically depending on pi
values. After that, produce new sources by Eq. (2) and
update the position of food sources using the greedy
selection approach as in step (3).
6. Check whether there exists an abandoned source or
not; if so, it is replaced with a new randomly generated
food source by the scout using Eq. (1).
7. Repeat stages (2)–(6) until the maximum number of
cycles.
3 The clustering problem using the ABC algorithm
Patterns are the physical definition of an object used by
clustering and they are distinguished from one another
through their attributes known as features. Given a data set
Z ¼ z1; z2; . . .; zp; . . .; zN
� �; zp is a pattern with D-dimen-
sional feature space, and N is the number of patterns in Z. To
partition Z into K number of clusters as C ¼C1;C2; . . .;Ckf g via a crisp clustering process, the following
properties are need to be considered:
• Each cluster should have at least one pattern,
Ci 6¼ ;8i ¼ 1; 2; . . .;K
• Each pattern should be assigned to a cluster,
[Kk¼1
Ck ¼ Z
• Each pattern should be assigned to only one cluster,
Ci \ Cj ¼ ; where i 6¼ j
Having these properties, clustering is done based on
some similarity measures. Distance measurement is
generally used for evaluating similarities between
patterns, which constructs fundamental part of an
objective function (fit) in a clustering problem. In this
paper, four objective functions are used in experiments. The
first two of them (fit1 [30] and fit2 [48]) are very popular in
the literature, the third one (Pfit) is the proposed objective
function in this study by improving fit1 and fit2 together, and
the last one (CSfit [49]) is employed as an objective
function, recently used in automatic clustering problem
[34]. The objective functions can be described as follows:
1. The first applied objective function: fit1 [30] applies
the three main criterions, comprising error measure,
intra-cluster distance and inter-cluster separation in a
weighted manner;
fit1 xi; Zð Þ ¼ w1dmax Z; xið Þ þ w2 zmax � dmin Z; xið Þð Þ þ w3Je
ð5Þ
where Z is the image data set; xi is the ith food source and
defined by xi ¼ ðmi;1;mi;2; . . .;mi;k. . .;mi;KÞ, mi,k is the
centre of kth cluster of the ith food source; zmax is the
maximum value of data set (for a s-bit image, zmax = 2s - 1);
and w1, w2 and w3 are user-defined parameters having
constant values. In addition, dmax(Z, xi), known as intra-
cluster distance, is the maximum average distance of a
cluster calculated by distances of all pixels to their
respective centre. To satisfy well-collected pixel groups,
intra-cluster distance [29] should be minimized, defined by
Eq. (6);
dmax Z; xið Þ ¼ maxX8zp2Ci;k
d zp;mi;k
� �ni;k
8<:
9=; ð6Þ
where ni;k is the number of pixels in cluster, Ci,k and
d(zp, mi,k) is the Euclidean distance between zp and mi,k.
The second component of fit1 is the inter-cluster separation
which is known as the minimum average Euclidean
distance (dmin) [50] between any pairs of clusters. The
dmin criterion, shown as follows, should be maximized to
obtain well-separated clusters;
dmin Z; xið Þ ¼ min d mi;j;mi;k
� �� �; j 6¼ k: ð7Þ
And the last component of fit1, known as the quantization
error (Je) and defined by Eq. (8), is used to express the
general quality of a clustering algorithm;
Je ¼PK
k¼1
P8zp2Ck
dðzp;mkÞ=nk
K: ð8Þ
2. The second applied objective function: The parameters
of fit1 are needed to be selected empirically for every
image, which is not flexible on account of high time
consumption. To overcome this problem, fit2 [48] was
improved;
fit2 xi; Zð Þ ¼ dmax Z; xið Þ þ Je;i
dmin Z; xið Þ : ð9Þ
3. The proposed objective function: Although fit2requires no user-defined parameters, the performance
of fit2 is almost similar to fit1. To tackle the drawbacks
of the above-mentioned objective functions, a new
fitness function (Pfit) is improved. Contrary to the
others, Pfit considers the mean square error (MSE) and
the quantization error in an efficient way. Moreover,
Pattern Anal Applic
123
Pfit aims to keep the error rate in minor rates while
maximizing the dmax and minimizing the dmin
criterions at the same time. The equation of Pfit is as
follows:
Pfit xi; Zð Þ ¼ Je �dmax Z; xið Þdmin Z; xið Þ
� dmax Z; xið Þ þ zmax � dmin Z; xið Þ þMSEð Þð10Þ
where MSE is defined by Eq. (11);
MSE ¼ 1
N
XK
k¼1
X8zp2Ck
dðzp;mkÞ2 ð11Þ
where N is the total number of patterns in a data set.
4. The last applied objective function: The CS measurement
[49] looks for clusters having minimum within cluster-
scatter and maximum between-cluster separation;
CS ¼
PKk¼1
1nk
Pzp2Ck
maxzq2Ck
dðzp; zqÞPK
k¼1 minj2K;j 6¼k
dðmk;mjÞð12Þ
CSfit xi; Zð Þ ¼ CS: ð13Þ
In addition to the Je, dmax and dmin criterions, two
clustering validity indexes such as the Davies–Bouldin
index (DBI) [51] and the XB index (XBI) [52] are chosen
to measure and compare the clustering qualities of the
above-mentioned objective functions:
1. The Davies–Bouldin Index: The Davies–Bouldin
index (DBI) [51], one of the most well-known
internal indexes, is a function of the ratio of the
sum of within cluster-scatter to between-cluster
separation. The scatter within the ith cluster (Si) and
between the ith and jth cluster (Rij) are defined by
Eqs. (14) and (15);
Si ¼1
ni
Xxj2Ci
d xj;mi
� �2 ð14Þ
Rij ¼Si þ Sj
dðmi;mjÞ2; i 6¼ j: ð15Þ
Last, DBI can be described as
DBI ¼ 1
K
XK
k¼1
Rk ð16Þ
where Rk ¼ maxj¼1;2;...;KRij and i ¼ 1; 2; . . .;K:2. The XB Index: The sum squares within (SSW) clusters
is a measurement of compactness and the sum squares
between (SSB) clusters is a measurement criterion of
separateness. The equations of SSW and SSB are given
as follows:
SSW ¼XK
k¼1
X8zp2Ck
dðzp;mkÞ2 ð17Þ
SSB ¼XK
k¼1
nkdðmk;MÞ2: ð18Þ
According to Zhao et al. [52], the SSW in the ratio of
SSW/SSB plays more important role in clustering and the
indexes such as Calinski–Harabasz [53], Ball–Hubert [54],
Hartigan [55], etc. based on the SSW and SSB criterions
have such drawback that they can be monotonously
increased or decreased. Therefore, they proposed XB
index (XBI) [52];
XBI ¼ K � SSW
SSB: ð19Þ
The fast implementation of image clustering algorithm
has two stages: histogram process of an image and creation
of data structure including grey levels, related clusters and
total pixel numbers. When an empty cluster is created for a
xi solution, this solution is generated again until receiving a
non-empty cluster set. Figure 2 briefly explains the basic
stages of the ABC-based clustering algorithm.
4 Experimental results
Experimental results can be considered in six stages: (1)
comparing the ABC, PSO and GA algorithms according to
the obtained fitness values, (2) analysing the objective
functions such as fit1, fit2, Pfit and CSfit within the ABC,
PSO and GA algorithms in terms of the separateness and
compactness properties (Je, dmax and dmin) as in [29–31], (3)
Fig. 2 The main steps of ABC-based image clustering algorithm
Pattern Anal Applic
123
measuring the qualities of the objective functions within the
ABC, PSO and GA algorithms according to their DBI and
XBI values, (4) determining whether there exist significant
differences between the results of the ABC, PSO and GA
algorithms over Pfit, (5) comparing the results obtained by
ABC, PSO and GA algorithms with the results of K-means
algorithm, and (6) judging whether there exist significance
differences between the employed fitness functions and the
proposed fitness function over the above-mentioned crite-
rions. The experimental results are summarized through the
Fig. 3 a 1. Airplane Image; 2. Clustered by ABC with Pfit, b 1.
House Image; 2. Clustered by ABC with Pfit, c 1. Lena Image; 2.
Clustered by ABC with Pfit, d 1. Morro Bay Image; 2. Clustered by
ABC with Pfit, e 1. MRI Image; 2. Clustered by ABC with pfit, f 1.
42049 Image; 2. Clustered by ABC with Pfit, g 1. 48025 Image; 2.
Clustered by ABC with Pfit
Pattern Anal Applic
123
following images: the five of them are airplane, house,
Lena, Morro Bay and MRI shown in Fig. 3a.1, b.1, c.1, d.1
and e.1, respectively; and the last two images are 42049
(3f.1) and 48025 (3g.1) taken from Berkeley segmentation
data set and benchmark images [56].
To evaluate the results in a statistically significant way,
Friedman test [57], which is a non-parametric statistical
test and based on ranks between the independent samples
(blocks or rows), is employed;
FR ¼12
rcðcþ 1ÞXc
j¼1
R2j � 3r cþ 1ð Þ ð20Þ
where Rj is the square of the sum of ranks for each group j,
r is the number of blocks (rows) and c is the number of
groups. Through the FR value, P value is obtained. When
the obtained P value is smaller than 0.05, the difference
between the independent samples is significant. Otherwise,
there exists no such a significant difference between the
compared results.
In the experiments, the following parameters are used
for the ABC, PSO, GA and K-means algorithms: the colony
size is taken 30 and the value of limit is 75 for ABC; and the
number of particles (chromosomes) is taken 30 for PSO and
GA. The maximum number of iterations (cycles) for ABC,
PSO and GA is set to 250, i.e. the number of evaluations is
7,500. Therefore, the number of iterations of the K-means
algorithm is selected as 7,500. The number of clusters is
taken 5 in all cases. The parameters of the PSO algorithm are
chosen as used in [30] (c1 = 1.49, c2 = 1.49, wstart = 0.72,
wend = 0.4 and Vmax = 255) and the parameters of GA such
as crossover rate and mutation rate are chosen as 0.8 and 0.2,
respectively. The weights of fit1 are chosen as w1 = 0.4,
w2 = 0.2, w3 = 0.4 for each image.
The simulation runs are studied on the computer with
29CPU 2.4 Ghz Dual (6 cores), 4 9 8 GB Ram, and it is
observed that the simulation times of the algorithms are
very close to each other. The results of the mean values and
standard deviations of the considered criterions are pre-
sented over the 30 simulation runs in the tables. Further-
more, the best mean values for the compared algorithms
and the statistically significant results are written in bold,
and standard deviation values of produced results are
reported in parenthesis in the tables. The obtained fitness
values of the ABC, PSO and GA algorithms are summa-
rized for each objective function in the Table 1. In addi-
tion, the separateness and compactness properties (Je, dmax,
dmin, DBI and XBI) of each objective function are dem-
onstrated in the Table 2 for the images: airplane (Fig. 3a),
house (Fig. 3b), Lena (Fig. 3c), Morro Bay (Fig. 3d), MRI
(Fig. 3e), 42049 (Fig. 3f) and 48025 (Fig. 3g) in terms of
ABC,PSO and GA algorithms, respectively. The obtained
results of the K-means algorithm for each criterion with the
results of the algorithms such as ABC, PSO and GA based Ta
ble
1T
he
resu
lts
of
fitn
ess
val
ues
for
the
ob
ject
ive
fun
ctio
ns
(Fit
CS
,F
it1,
Fit
2an
dP
Fit
)
Imag
esF
itC
SF
it1
Fit
2P
Fit
AB
CP
SO
GA
AB
CP
SO
GA
AB
CP
SO
GA
AB
CP
SO
GA
Air
pla
ne
0.4
99
(0.0
01)
0.5
01
(0.0
03
)
0.5
20
(0.0
21
)
46
.15
9(0
.07
9)
46
.17
5
(0.2
73
)
47
.32
8
(1.7
41
)
0.4
98
(0.0
05)
0.5
00
(0.0
13
)
0.5
70
(0.0
59
)
79
0.6
18
(8.7
89
)
77
7.7
67
(8.6
76)
92
4.5
96
(13
3.1
09
)
Ho
use
0.6
05
(0.0
01)
0.6
06
(0.0
03
)
0.6
19
(0.0
20
)
50
.79
5(0
.15
5)
51
.11
6
(0.8
41
)
51
.73
2
(0.9
93
)
0.4
90
(0.0
08)
0.5
05
(0.0
47
)
0.5
45
(0.0
53
)
1,0
70
.08
(11.8
40
)
1,0
86
.34
(13
7.1
40
)
1,1
92
.32
4
(21
2.8
51
)
Len
a0
.57
8
(0.0
01
)
0.5
77
(0.0
01)
0.5
96
(0.0
20
)
50
.06
5(0
.08
8)
50
.08
7
(0.2
36
)
50
.66
6
(0.6
16
)
0.5
03
(0.0
06)
0.5
03
(0.0
11
)
0.5
44
(0.0
44
)
91
9.3
07
(13
.98
7)
89
9.3
65
(11
.46
4)
1,0
25
.06
5
(10
0.0
02
)
Mo
rro
Bay
0.6
04
(0.0
01)
0.6
05
(0.0
03
)
0.6
19
(0.0
20
)
50
.09
7(0
.10
4)
50
.16
8
(0.1
86
)
50
.74
1
(0.5
22
)
0.4
54
(0.0
06)
0.4
58
(0.0
08
)
0.4
83
(0.0
23
)
88
6.0
52
(9. 8
25)
89
8.8
66
(64
.64
8)
99
8.2
67
(13
0.4
35
)
MR
I0
.49
0(0
.00
1)
0.4
96
(0.0
09
)
0.5
25
(0.0
41
)
40
.49
0
(0.0
73)
40.4
48
(0.0
96
)
40
.92
9
(0.4
66
)
0.4
46
(0.0
03)
0.4
48
(0.0
05
)
0.4
90
(0.0
55
)
52
3.1
45
(7.2
82)
52
3.7
51
(16
.78
6)
59
5.7
81
(61
.69
5)
42
04
90
.52
5
(0.0
01
)
0.5
23
(0.0
01)
0.5
96
(0.0
11
)
50
.78
9
(0.0
58
)
50.7
68
(0.1
70
)
52
.08
9
(1.0
65
)
0.4
87
(0.0
05)
0.4
88
(0.0
11
)
0.6
03
(0.0
92
)
67
3.3
70
(7.7
85
)
66
7.9
83
(9.1
10
)
71
4.8
39
(45
.81
8)
48
02
50
.58
6
(0.0
01
)
0.5
85
(0.0
01
)
0. 5
61
(0.0
04
)
51
.06
4(0
.10
4)
51
.07
4
(0.2
77
)
51
.67
5
(0.5
62
)
0.5
05
(0.0
07)
0.5
06
(0.0
16
)
0.5
31
(0.0
29
)
1,1
92
.57
(16.6
02
)
1,2
02
.15
(24
.53
6)
1,3
32
.13
(20
2.3
9)
Pattern Anal Applic
123
Ta
ble
2T
he
resu
lts
of
com
par
iso
ncr
iter
ion
s(J
e,
dm
ax,
dm
in,
DB
Ian
dX
BI)
Fit
s.A
BC
PS
OG
A
J ed
max
dm
inD
BI
XB
IJ e
dm
ax
dm
inD
BI
XB
IJ e
dm
ax
dm
inD
BI
XB
I
Air
pla
ne
Fit
CS
19.7
01
(0.1
54)
36.9
82
(0.9
12)
40.6
02
(1.4
87)
0.2
401
(0.0
117)
0.4
951
(0.0
223)
19.7
60
(0.2
96)
34.7
68
(1.9
03)
51.8
83
(3.1
99)
0.2
276
(0.0
069)
0.4
699
(0.0
503)
21.1
09
(1.6
87)
36.6
44
(6.0
05)
49.1
80
(9.0
33)
0.3
254
(0.1
452)
0.8
844
(0.4
591)
Fit
19.8
89
(0.2
04)
11.0
16
(0.4
38)
42.0
13
(1.2
79)
0.1
590
(0.0
021)
0.2
826
(0.0
340)
9.9
13
(0.3
56)
11.2
12
(0.4
93)
42.3
73
(1.2
86)
0.1
592
(0.0
045)
0.2
792
(0.0
377)
10.4
84
(1.0
17)
12.0
51
(1.3
63)
40.6
89
(2.7
39)
0.1
739
(0.0
222)
0.3
101
(0.0
624)
Fit
29.9
34
(0.3
62)
11.0
77
(0.5
72)
42.1
05
(1.6
51)
0.1
586
(0.0
023)
0.2
828
(0.0
349)
10.0
81
(0.3
38)
11.5
13
(0.4
71)
43.1
91
(1.3
49)
0.1
597
(0.0
036)
0.2
770
(0.0
512)
11.3
36
(1.4
46)
12.9
48
(2.0
81)
42.5
46
(3.4
00)
0.1
799
(0.0
227)
0.3
287
(0.0
627)
PF
it9
.710
(0. 0
46
)
10.7
34
(0.3
22
)
40.5
31
(1.2
34)
0.1
569
(0.0
008)
0.2
533
(0.0
111
)
9.7
23
(0.1
69)
10.8
35
(0.3
61)
41.3
29
(1.2
02)
0.1
571
(0.0
012)
0.2
566
(0.0
153)
10.0
83
(0.5
02)
11.4
35
(0.6
59)
39.6
97
(2.3
02)
0.1
66
(0.0
126)
0.2
535
(0.0
236)
House
Fit
CS
18.2
88
(0.1
46)
26.4
47
(1.1
73)
56.0
87
(2.2
14)
0.1
973
(0.0
047)
0.4
808
(0.0
118)
18.0
94
(0.3
77)
27.4
71
(2.0
92)
54.3
77
(4.9
98)
0.1
943
(0.0
048)
0.5
464
(0.0
966)
18.0
81
(0.9
79)
28.0
84
(4.5
47)
49.5
75
(7.9
69)
0.2
174
(0.0
389)
0.5
912
(0.2
031)
Fit
111.3
47
(0.2
91)
11.9
42
(0.4
16)
47.6
01
(1.3
53)
0.1
470
(0.0
033)
0.2
998
(0.0
086)
11.6
04
(0.7
76)
12.4
36
(0.9
76)
47.4
98
(2.7
09)
0.1
636
(0.0
298)
0.3
179
(0.0
429)
11.6
36
(0.7
83)
13.0
24
(0.8
87)
45.6
61
(4.3
51)
0.1
78
(0.0
039)
0.3
322
(0.0
553)
Fit
211.3
19
(0.3
06)
11.9
63
(0.3
97)
47.4
94
(1.3
77)
0.1
468
(0.0
029)
0.2
992
(0.0
089)
11.8
42
(0.9
05)
12.7
14
(1.3
91)
48.6
59
(1.8
29)
0.1
610
(0.0
260)
0.3
226
(0.0
514)
12.4
10
(1.1
40)
13.9
19
(1.7
22)
48.3
49
(3.2
31)
0.1
71
(0.0
028)
0.3
520
(0.0
478)
PF
it10.8
18
(0.1
09)
11.4
27
(0.0
99
)
44.2
63
(1.6
11)
0.1
452
(0.0
027
)
0.2
759
(0.0
083)
10.8
21
(0.1
48)
11.4
83
(0.4
68)
44.6
80
(1.9
86)
0.1
506
(0.0
229)
0.2
831
(0.0
279)
11.0
21
(0.3
65)
11.9
50
(0.7
38)
43.6
47
(3.1
36)
0.1
57
(0.0
272)
0.3
161
(0.0
625)
Len
a
Fit
CS
18.2
51
(0.1
18)
26.5
51
(0.7
97)
59.8
92
(1.2
23)
0.2
059
(0.0
034)
0.5
023
(0.0
046)
18.2
21
(0.1
06)
26.4
71
(0.7
71)
60.4
29
(1.1
83)
0.2
046
(0.0
034)
0.5
003
(0.0
046)
18.8
79
(1.5
79)
30.3
69
(10.3
1)
52.3
87
(9.3
32)
0.2
831
(0.2
452)
0.6
689
(0.2
935)
Fit
110.2
09
(0.3
84)
11.1
20
(0.4
72)
42.3
33
(1.5
98)
0.1
584
(0.0
042)
0.2
474
(0.0
112)
10.6
08
(0.4
62)
11.4
96
(0.6
09)
43.7
73
(1.1
35)
0.1
628
(0.0
049)
0.2
589
(0.0
192)
10.8
09
(0.9
73)
11.9
15
(1.3
41)
42.1
17
(2.3
72)
0.1
66
(0.0
121)
0.2
852
(0.0
386)
Fit
210.1
39
(0.3
03)
11.0
96
(0.3
62)
41.9
87
(1.1
99)
0.1
586
(0.0
058)
0.2
416
(0.0
083)
10.5
37
(0.5
33)
11.6
05
(0.7
07)
44.0
03
(1.5
99)
0.1
625
(0.0
057)
0.2
621
(0.0
174)
11.2
41
(0.9
16)
12.3
92
(1.2
85)
43.8
93
(3.1
24)
0.1
69
(0.0
094)
0.3
062
(0.0
490)
PF
it9.7
36
(0.1
08)
10.6
76
(0.1
33
)
39.9
55
(0.9
67)
0.1
548
(0.0
044
)
0.2
297
(0.0
047)
9.7
93
(0.1
98)
10.6
91
(0.2
24)
40.5
64
(0.5
03)
0.1
554
(0.0
045)
0.2
361
(0.0
083)
10.2
11
(0.4
89)
11.0
46
(0.5
83)
40.2
23
(1.9
45)
0.1
629
(0.0
073)
0.2
585
(0.0
287)
Morr
o
Bay
Fit
CS
17.6
11
(0.0
63)
31.9
67
(0.0
75)
57.4
13
(0.7
24)
0.1
797
(0.0
016)
0.3
268
(0.0
013)
17.6
59
(0.0
31)
31.9
81
(0.0
01)
58.3
32
(0.3
47
)
0.1
804
(0.0
009)
0.3
264
(0.0
007)
17.2
87
(0.9
42)
30.2
56
(3.2
96)
50.0
74
(7.9
47)
0.2
071
(0.0
508)
0.3
478
(0.0
535)
Fit
110,1
92
(0.4
79)
12,2
85
(0.7
66)
49,4
63
(1,8
28)
0.1
37
(0.0
042)
0.1
626
(0.0
112)
10,4
79
(0.4
55)
13,1
96
(0.9
15)
51,5
09
(2,2
82)
0.1
422
(0.0
055)
0.1
839
(0.0
128)
10,9
49
(1.0
61)
13,5
51
(1.0
07)
50,2
94
(2,9
74)
0.1
468
(0.0
098)
0.1
853
(0.0
171)
Fit
210,0
10
(0.3
19)
12,0
84
(0.7
54)
48,6
88
(1.7
18)
0.1
361
(0.0
039)
0.1
566
(0.0
057)
10,6
23
(0.6
53)
13,2
98
(0.9
29)
52,1
96
(2,5
12)
0.1
423
(0.0
057)
0.1
759
(0.0
150)
11,2
14
(1.0
10)
13,8
24
(0.9
83)
51,8
08
(2,5
66)
0.1
480
(0.0
088)
0.1
822
(0.0
193)
PF
it9.6
96
(0.1
37)
11.7
10
(0.1
67
)
46,9
55
(0.4
66)
0.1
352
(0.0
013
)
0.1
536
(0.0
019)
9,7
63
(0.2
48)
12.0
9
(0.8
73)
47,9
72
(0.9
99)
0.1
360
(0.0
047)
0.1
559
(0.0
093)
10,0
04
(0.2
85)
12,4
51
(1.1
73)
47,1
32
(1.7
33)
0.1
412
(0.0
100)
0.1
673
(0.0
207)
MR
I
Fit
CS
18.0
94
(0.0
29)
47.9
93
(0.0
38)
52.1
74
(0.3
78)
0.1
98
(0.0
01)
0.3
12
(0.0
02)
18.1
63
(0.4
60)
47.7
74
(2.0
35)
50.3
44
(2.9
80)
0.2
009
(0.0
087)
0.3
178
(0.0
126)
18.4
64
(1.5
72)
48.7
53
(9.2
38)
46.2
21
(10.0
7)
0.2
513
(0.1
187)
0.3
912
(0.1
324)
Fit
18.4
69
(0.2
55)
11.4
76
(0.3
95)
44.4
38
(1.3
36)
0.1
322
(0.0
019)
0.2
653
(0.0
069)
8.6
45
(0.3
61)
11.6
51
(0.4
69)
45.3
52
(1.5
96)
0.1
316
(0.0
029)
0.2
668
(0.0
065)
8.7
99
(0.4
73)
11.6
21
(0.7
04)
43.1
96
(2.7
53)
0.1
410
(0.0
101)
0.2
712
(0.0
139)
Fit
28.3
31
(0.3
03)
11.2
43
(0.5
11)
43.5
09
(1.8
46)
0.1
321
(0.0
018)
0.2
654
(0.0
060)
8.6
07
(0.3
17)
11.6
44
(0.4
92)
45.1
65
(1.4
30)
0.1
329
(0.0
037)
0.2
667
(0.0
058)
9.0
26
(0.9
03)
12.1
74
(1.7
81)
43.3
16
(2.5
49)
0.1
465
(0.0
175)
0.2
715
(0.0
382)
Pattern Anal Applic
123
Ta
ble
2co
nti
nu
ed
Fit
s.A
BC
PS
OG
A
J ed
max
dm
inD
BI
XB
IJ e
dm
ax
dm
inD
BI
XB
IJ e
dm
ax
dm
inD
BI
XB
I
PF
it8
.049
(0.1
02
)
10.6
01
(0.3
73)
41.4
83
(1.1
94)
0.1
338
(0.0
024)
0.2
648
(0.0
051
)
8.1
06
(0.1
75)
10.6
93
(0.2
82)
41.5
78
(0.9
22)
0.1
327
(0.0
014)
0.2
661
(0.0
096)
8.4
84
(0.3
27)
11.3
23
(0.6
19)
40.3
69
(2.7
63)
0.1
442
(0.0
096)
0.2
685
(0.0
386)
42049
Fit
CS
18.7
52
(0.0
99)
27.1
19
(0.4
09)
62.9
44
(0.6
26)
0.2
213
(0.0
010)
0.3
351
(0.0
103)
18.6
85
(0.0
38)
27.1
05
(0.2
28)
63.6
12
(0.4
43
)
0.2
199
(0.0
004)
0.3
363
(0.0
036)
17.7
63
(0.6
51)
27.9
17
(5.1
61)
51.6
88
(6.5
63)
0.2
133
(0.0
379)
0.5
284
(0.1
378)
Fit
18.9
44
(0.2
82)
10.3
25
(0.2
63)
39.5
96
(0.9
60)
0.1
554
(0.0
026)
0.1
349
(0.0
136)
9.3
75
(0.5
20)
10.5
55
(0.4
36)
41.0
18
(1.3
11)
0.1
574
(0.0
057)
0.1
514
(0.0
277)
10.9
20
(0.8
78)
12.3
71
(1.6
07)
41.1
34
(2.2
40)
0.1
799
(0.0
191)
0.2
074
(0.0
352)
Fit
28.8
92
(0.2
48)
10.2
01
(0.3
27)
39.1
42
(1.1
58)
0.1
569
(0.0
036)
0.1
327
(0.0
059)
9.4
88
(0.6
31)
10.6
62
(0.5
63)
41.2
51
(1.7
98)
0.1
582
(0.0
055)
0.1
517
(0.0
240)
11.6
67
(1.7
76)
13.8
11
(3.0
85)
42.2
31
(3.8
90)
0.1
912
(0.0
310)
0.2
425
(0.0
765)
PF
it8
.699
(0.0
53
)
9.9
00
(0.1
58
)
38.0
24
(0.7
66)
0.1
558
(0.0
025)
0.1
281
(0.0
025
)
8.7
44
(0.1
39)
10.0
26
(0.2
54)
38.9
30
(0.8
63)
0.1
558
(0.0
022)
0.1
286
(0.0
036)
9.1
58
(0.7
18)
10.5
21
(0.9
90)
37.5
43
(2.8
78)
0.1
655
(0.0
173)
0.1
486
(0.0
314)
48025
Fit
CS
17.7
81
(0.2
27)
31.3
01
(2.2
18)
51.3
20
(3.7
47)
0.1
987
(0.0
101)
0.5
423
(0.0
400)
17.9
82
(0.3
71)
29.6
26
(3.6
68)
54.6
18
(6.0
64)
0.2
011
(0.0
102)
0.5
318
(0.0
384)
17.7
31
(0.7
95)
28.0
26
(5.2
14)
50.5
84
(6.9
59)
0.2
180
(0.0
511)
0.5
299
(0.1
454)
Fit
111.1
49
(0.2
45)
13.3
62
(0.6
60)
48.6
99
(1.9
88)
0.1
537
(0.0
041)
0.2
935
(0.0
101)
11.5
15
(0.3
99)
13.9
31
(0.5
55)
50.6
72
(1.9
99)
0.1
575
(0.0
038)
0.3
024
(0.0
129)
11.6
78
(0.9
50)
13.2
51
(1.0
07)
46.4
87
(4.3
18)
0.1
612
(0.0
142)
0.2
991
(0.0
311)
Fit
211.2
01
(0.3
89)
13.4
49
(0.9
61)
48.8
36
(2.6
77)
0.1
542
(0.0
055)
0.2
971
(0.0
138)
11.7
78
(0.8
64)
14.1
17
(0.9
24)
51.2
47
(2.5
15)
0.1
578
(0.0
059)
0.3
089
(0.0
183)
12.1
75
(1.0
93)
14.1
04
(1.4
87)
49.5
00
(4.4
06)
0.1
639
(0.0
118)
0.3
180
(0.0
332)
PF
it10.8
95
(0.1
03)
12.2
76
(0.3
02)
44.7
16
(1.9
14)
0.1
504
(0.0
023)
0.2
781
(0.0
100)
10.8
72
(0.0
98
)
12.0
79
(0.2
72
)
41.9
17
(2.7
50)
0.1
552
(0.0
052)
0.2
699
(0.0
141)
11.2
77
(0.7
01)
13.0
82
(1.2
74)
45.7
98
(4.5
53)
0.1
594
(0.0
089)
0.2
871
(0.0
294)
Pattern Anal Applic
123
Table 3 The comparison
criterions values of ABC, PSO,
GA and K-means algorithms
ABC (Pfit) PSO (Pfit) GA (Pfit) K-means
Airplane Je 9.710 (0.046) 9.723 (0.169) 10.083 (0.502) 9.854 (0.121)
dmax 10.734 (0.322) 10.835 (0.361) 11.435 (0.659) 15.434 (0.577)
dmin 40.531 (1.234) 41.329 (1.202) 39.697 (2.302) 21.167 (3.573)
DBI 0.1569 (0.0008) 0.1571 (0.0012) 0.166 (0.0126) 0.233 (0.001)
XBI 0.2533 (0.0111) 0.2566 (0.0153) 0.2535 (0.0236) 0.2048 (0.0003)
House Je 10.818 (0.109) 10.821 (0.148) 11.021 (0.365) 10.330 (0.061)
dmax 11.427 (0.099) 11.483 (0.468) 11.950 (0.738) 12.545 (0.783)
dmin 44.263 (1.611) 44.680 (1.986) 43.647 (3.136) 31.833 (1.053)
DBI 0.1452 (0.0027) 0.1506 (0.0229) 0.157 (0.0272) 0.168 (0.001)
XBI 0.2759 (0.0083) 0.2831 (0.0279) 0.3161 (0.0625) 0.2454 (0.0001)
Lena Je 9.736 (0.108) 9.793 (0.198) 10.211 (0.489) 9.430 (0.036)
dmax 10.676 (0.133) 10.691 (0.224) 11.046 (0.583) 11.712 (0.261)
dmin 39.955 (0.967) 40.564 (0.503) 40.223 (1.945) 34.833 (0.461)
DBI 0.1548 (0.0044) 0.1554 (0.0045) 0.1629 (0.0073) 0.154 (0.001)
XBI 0.2297 (0.0047) 0.2361 (0.0083) 0.2585 (0.0287) 0.2266 (5.9E-5)
MorroBay Je 9.696 (0.137) 9,763 (0.248) 10,004 (0.285) 9,998 (0.076)
dmax 11.710 (0.167) 12,09 (0.873) 12,451 (1.173) 14,032 (1.263)
dmin 46,955 (0.466) 47,972 (0.999) 47,132 (1.733) 41,565 (3,863)
DBI 0.1352 (0.0013) 0.1360 (0.0047) 0.1412 (0.0100) 0.148 (0.005)
XBI 0.1536 (0.0019) 0.1559 (0.0093) 0.1673 (0.0207) 0.1666 (0.0071)
MRI Je 8.049 (0.102) 8.106 (0.175) 8.484 (0.327) 8,142 (0.056)
dmax 10.601 (0.373) 10.693 (0.282) 11.323 (0.619) 17,198 (0.280)
dmin 41.483 (1.194) 41.578 (0.922) 40.369 (2.763) 21,698 (0.492)
DBI 0.1338 (0.0024) 0.1327 (0.0014) 0.1442 (0.0096) 0.154 (0.001)
XBI 0.2648 (0.0051) 0.2661 (0.0096) 0.2685 (0.0386) 0.1704 (0.0001)
42049 Je 8.699 (0.053) 8.744 (0.139) 9.158 (0.718) 8.989 (0.244)
dmax 9.900 (0.158) 10.026 (0.254) 10.521 (0.990) 12.759 (1.470)
dmin 38.024 (0.766) 38.930 (0.863) 37.543 (2.878) 19.838 (7.271)
DBI 0.1558 (0.0025) 0.1558 (0.0022) 0.1655 (0.0173) 0.2088 (0.2088)
XBI 0.1281 (0.0025) 0.1286 (0.0036) 0.1486 (0.0314) 0.1298 (0.0089)
48025 Je 10.895 (0.103) 10.872 (0.098) 11.277 (0.701) 10.526 (0.021)
dmax 12.276 (0.302) 12.079 (0.272) 13.082 (1.274) 12.512 (0.232)
dmin 44.716 (1.914) 41.917 (2.750) 45.798 (4.553) 31.350 (0.249)
DBI 0.1504 (0.0023) 0.1552 (0.0052) 0.1594 (0.0089) 0.1683 (0.0012)
XBI 0.2781 (0.0100) 0.2699 (0.0141) 0.2871 (0.0294) 0.2440 (0.0002)
Table 4 The Friedman test
results of ABC, PSO and GA
over Pfit objective function
Images Je dmax DBI XBI
v2 p val v2 p val v2 p val v2 p val
Airplane 19.04 6.1e2005 23.15 9.4e2006 29.87 3.2e2007 1.67 0.4346
House 6.07 0.0482 40.78 1.3e2009 11.04 0.0033 5.60 0.0608
Lena 24.47 4.8e2006 3.58 0.1668 24.27 5.3e2006 29.60 3.7e2007
Morro Bay 21.07 2.6e2005 7.2 0.0273 2.07 0.3558 18.47 9.7e2005
MRI 27.80 9.1e2007 19.66 5.3e2005 5.34 0.0692 4.27 0.1184
42049 24.87 3.9e2006 11.37 0.0034 13.07 0.0015 20.60 3.3e2005
48025 5.4 0.0672 19.76 5.1e2005 16.47 0.0003 8.87 0.0119
Pattern Anal Applic
123
on Pfit are demonstrated in the Table 3. In addition, the
statistical significance of the proposed objective function
results over the algorithms: ABC, PSO and GA; summa-
rized in Table 4. Last, the statistical significance of the
employed and proposed objective functions on the evalu-
ation criterions through the ABC algorithm is presented in
Table 5.
The most important evaluation criterion for comparing
evolutionary algorithms is to analyse obtained fitness
values. As clearly seen from the Table 1, the ABC algo-
rithm is almost superior to the PSO and GA algorithms in
terms of minimizing the objective functions. Table 1 also
reveals that the ABC algorithm gets the minimum stan-
dard deviation values in most cases. Indeed, ABC out-
performs the others in minimizing the Je and dmax
criterions which can be indicated from the Table 2.
Although the PSO and GA algorithms perform better than
the ABC algorithm in maximizing the dmin criterion, the
Table 2 demonstrates that the ABC algorithm almost
outperforms the others in terms of minimizing the DB and
XBI values which are accepted as the well-known crite-
rions of measuring the quality of clustering process. That
makes the ABC algorithm is the most convenient algo-
rithm for clustering process. Therefore, this disadvantage
of the ABC algorithm in maximizing the dmin criterion can
be ignored. Moreover, the GA algorithm is always
arranged in the last position in minimizing the Je, dmax
criterions, objective functions, DBI and XBI values
against the ABC and PSO algorithms. It can be apparently
inferred from this situation that the GA algorithm is not
very suitable to the clustering process against the ABC
and PSO algorithms. In addition, the statistically signifi-
cant difference between the ABC and other algorithms
over Pfit in terms of the criterions except dmin can be
observed in Table 4, where the results clearly show that
almost in all cases ABC obtains much better performances
against the others. As for the K-means algorithm, the
performance of the K-means in minimizing XBI is suc-
cessful for almost all cases and in minimizing Je is suc-
cessful in some situations against the ABC and PSO
algorithms. However, the K-means algorithm actually
obtains the most awful performance in minimizing the
dmax criterion and in maximizing the dmin criterion. In
addition, the performance of the K-means algorithm in
minimizing the DBI values is also not sufficient when
compared with the other algorithms.
Considering the comparison of the objective functions,
it can be observed from the Table 2 that the proposed
objective function Pfit is almost superior to the other
objective functions in minimizing the Je and dmax criterions
and obtaining the best standard deviation values. Also, the
implementations of the ABC, PSO and GA algorithms with
Pfit almost get better results than the implementations of
the algorithms with the other objective functions. More-
over, it should be notified that the best performance using
Pfit is obtained by the ABC algorithm. In this way, the
ABC algorithm with the help of Pfit also gets advantages
versus the K-means algorithm in terms of the Je criterion.
Moreover, Pfit also performs better than fit1, fit2 and CSfit
in terms of obtaining the best DBI values except some
situations and Pfit gets the best XBI values in all cases
which can be extracted from the Table 2. As for CSfit, it
exhibits the best dmin values, yet the performance of this
objective function is not very sufficient for the other cri-
terions. The superiority of Pfit over the other criterions
through the ABC algorithm can be also inferred from
Table 5. It can therefore be concluded that Pfit is the most
convenient objective function to get well-separated and
well-compacted clusters despite its disadvantage in mini-
mizing the dmin criterion.
5 Conclusion
Clustering is a very comprehensive research area and has
been used in many applications. In this paper, the ABC-
based clustering algorithm with the improved function is
proposed, in which the mean square error (MSE) is effi-
ciently used. The algorithms such as ABC, PSO, GA and
K-means with the handled objective functions are com-
pared in terms of many criterions on seven benchmark
images. The results indicate that ABC-based clustering
Table 5 The Friedman test
results of objective functions
(FitCS, Fit1, Fit2 and Pfit)
through ABC
Images Je dmax DBI XBI
v2 p val v2 p val v2 p val v2 p val
Airplane 66.16 2.8e2014 55.92 4.4e2012 58.48 1.2e2012 66.04 3.0e2014
House 81 1.8e2017 75.96 2.2e2016 57.72 1.8e2012 81.04 1.8e2017
Lena 66.16 2.8e2014 69.96 4.3e2015 59.16 8.8e2012 75.28 3.1e2016
Morro Bay 66.64 2.3e2014 64.08 7.8e2014 54.00 1.1e2011 66.96 1.9e2014
MRI 71.56 1.9e2015 69.96 4.3e2015 56.92 2.6e2012 54.12 1.1e2011
42049 70.68 3.1e2015 69.96 4.3e2015 54.52 8.7e2012 58.28 1.2e2012
48025 66.64 2.2e2014 75.96 2.2e2016 59.16 8.9e2012 74.32 5.1e2016
Pattern Anal Applic
123
algorithm with the proposed objective function outper-
forms the others. Moreover, it is also indicated from the
results that all the criterions satisfying separateness and
compactness are needed to be handled together to obtain
well-separated clusters. In other words, one criterion is not
sufficient to satisfy well-separated and well-compacted
clusters. In conclusion, it can be suggested that the ABC-
based clustering algorithm with the proposed fitness func-
tion can be efficiently employed for clustering problems.
References
1. Samet R, Hancer E (2012) A new approach to the reconstruction
of contour lines extracted from topographic maps. J Vis Commun
Image Represent 23(4):642–647. doi:10.1016/j.jvcir.2012.02.005
2. Hancer E, Samet R (2011) Advanced contour reconnection in
scanned topographic maps. In: 2011 5th International conference
on application of information and communication technologies
(AICT), 12–14 Oct. 2011, pp 1–5
3. Jafar OAM, Sivakumar R (2010) Ant-based clustering algo-
rithms: a brief survey. Int J Comput Theory Eng 2:787–796
4. Jain AK, Murty MN, Flynn PJ (1999) Data clustering: a review.
ACM Comput Surv 31(3):264–323
5. Carpineto C, Romano G (1996) A lattice conceptual clustering
system and its application to browsing retrieval. Mach Learn
24(2):95–122. doi:10.1023/a:1018050230279
6. Leung Y, Zhang J-S, Xu Z-B (2000) Clustering by scale-space
filtering. IEEE Trans Pattern Anal Mach Intell 22(12):
1396–1410. doi:10.1109/34.895974
7. Manning CD, Schutze H (1999) Foundations of statistical natural
language processing
8. Frigui H, Krishnapuram R (1999) A robust competitive clustering
algorithm with applications in computer vision. IEEE Trans
Pattern Anal Mach Intell 21(5):450–465
9. Omran MGH, Salman A, Engelbrecht AP (2006) Dynamic
clustering using particle swarm optimization with application in
image segmentation. Pattern Anal Appl 8(4):332–344. doi:10.
1007/s10044-005-0015-5
10. Hamerly G, Elkan C (2002) Alternatives to the K-means algo-
rithm that find better clusterings. In: Proceedings of the ACM
conference on information and knowledge management (CIKM-
2002)
11. Omran M (2004) Particle swarm optimization methods for pattern
recognition and image processing. University of Pretoria, Envi-
ronment and Information Technology
12. Jain AK, Duin RPW, Mao JC (2000) Statistical pattern recogni-
tion: a review. IEEE Trans Pattern Anal Mach Intell 22(1):4–37
13. MacQueen J (1967) Some methods for classification and analysis
of multivariate observations. In: 5th Berkeley Symp. Math. Stat.
Probability
14. Karaboga D, Ozturk C (2011) A novel clustering approach:
artificial bee colony (ABC) algorithm. Appl Soft Comput
11(1):652–657
15. Forgy EW (1965) Cluster analysis of multivariate data: efficiency
versus interpretability of classification. Biometrics 21(3):768–769
16. Dunn JC (1973) A fuzzy relative of the ISODATA process and its
use in detecting compact well-separated clusters. J Cybern 3:32–57
17. Bezdek JC (1981) Pattern recognition with fuzzy objective
function algorithms. Plenum Press, New York
18. Liew AWC, Leung SH, Lau WH (2000) Fuzzy image clustering
incorporating spatial continuity. IEE Proc Vis Image Signal
Process 147(2):185–192
19. Forouzanfar M, Forghani N, Teshnehlab M (2010) Parameter
optimization of improved fuzzy c-means clustering algorithm for
brain MR image segmentation. Eng Appl Artif Intell
23(2):160–168. doi:10.1016/j.engappai.2009.10.002
20. Krishnapuram R, Keller JM (1993) A possibilistic approach to
clustering. IEEE Trans Fuzzy Syst 1(2):98–110
21. Pham DL (2001) Spatial models for fuzzy clustering. Comput Vis
Image Underst 84(2):285–297. doi:10.1006/cviu 2001.0951
22. Jian Y, Miin-Shen Y (2007) A generalized fuzzy clustering
regularization model with optimality tests and model complexity
analysis. IEEE Trans Fuzzy Syst 15(5):904–915
23. Frackiewicz M, Palus H (2008) Clustering with K-harmonic
means applied to colour image quantization. In: IEEE Interna-
tional Symposium on signal processing and information tech-
nology, 2008. ISSPIT 2008, 16–19 Dec. 2008, pp 52–57
24. Li Q, Mitianoudis N, Stathaki T (2007) Spatial kernel K-har-
monic means clustering for multi-spectral image segmentation.
IET Image Proc 1(2):156–167
25. Kennedy J, Eberhart R (1995) Particle swarm optimization. In:
IEEE international joint conference on neural networks, Australia
26. Eberhart R, Kennedy J (1995) A new optimizer using particle
swarm theory. In: Paper presented at the 6th international sym-
posium on micro machine and human science
27. Bharne PK, Gulhane VS, Yewale SK (2011) Data clustering
algorithms based on swarm intelligence. In: 3rd International
conference on electronics computer technology (ICECT), 8–10
April 2011, pp 407–411
28. Salman A, Omran M, Engelbrecht A (2002) Image classification
using particle swarm optimization. In: Paper presented at the
conference on simulated evolution and learning, Singapore
29. Omran M, Engelbrecht A, Salman A (2005) Particle swarm
optimization method for image clustering. Int J Pattern Recognit
Artif Intell 19(3):297–322
30. Omran M, Engelbrecht A, Salman A (2006) Particle swarm opti-
mization for pattern recognition and image processing. In: Abra-
ham A (ed) CGaVR Swarm intelligence and data mining, SCI series
studies in computational intelligence, vol 1. Springer, Berlin
31. Omran MGH, Engelbrecht AP (2006) Self-adaptive differential
evolution methods for unsupervised image classification. In: 2006
IEEE conference on cybernetics and intelligent systems, vols 1, 2.
IEEE, New York
32. Dorigo M (1992) Optimization Learning And Natural Algo-
rithms. PHD Thesis, Politecnico Di Milano, Italy
33. Piatrik T, Izquierdo E (2008) An application of ant colony opti-
mization to image clustering. In: Paper presented at the proc.
K-Space Jamboree Workshop
34. Das S, Abraham A, Konar A (2008) Automatic clustering using
an improved differential evolution algorithm. IEEE Trans Syst
Man Cybern Part A Syst Hum 38(1):218–237. doi:10.1109/tsmca.
2007.909595
35. Das S, Abraham A, Konar A (2009) Metaheuristic clustering. In:
Studies in computational intelligence, vol 178. Springer, Berlin
36. Das S, Konar A (2009) Automatic image pixel clustering with an
improved differential evolution. Appl Soft Comput 9(1):226–236.
doi:10.1016/j.asoc.2007.12.008
37. Sarkar S, Das S (2013) Multi-level image thresholding based on
two-dimensional histogram and maximum tsallis entropy—adifferential evolution approach. IEEE Trans Image Process
99:1–1. doi:10.1109/tip.2013.2277832
38. Karaboga D (2005) An idea based on honey bee swarm for
numerical optimization. Technical Report-TR06, Erciyes Uni-
versity, Engineering Faculty, Computer Engineering Department
Pattern Anal Applic
123
39. Karaboga D, Basturk B (2007) A powerful and efficient algorithm
for numerical function optimization: artificial bee colony (ABC)
algorithm. J Global Optim 39(3)
40. Karaboga D, Ozturk C (2009) Neural networks training by arti-
ficial bee colony algorithm on pattern classification. Neural Netw
World 19(3):279–292
41. Ma M, Liang J, Guo M, Fan Y, Yin Y (2011) SAR image seg-
mentation based on artificial bee colony algorithm. Appl Soft
Comput 11(8):5205–5214. doi:10.1016/j.asoc.2011.05.039
42. Ozturk C, Karaboga D, Gorkemli B (2011) Probabilistic dynamic
deployment of wireless sensor networks by artificial bee colony
algorithm. Sensors 11(6):6056–6065
43. Karaboga D, Okdem S, Ozturk C (2012) Cluster based wireless
sensor network routing using artificial bee colony algorithm.
Wirel Netw 18(7):847–860
44. Sag T, Cunkas M (2012) Development of image segmentation
techniques using swarm intelligence (ABC-based clustering
algorithm for image segmentation). In: International Conference
on Computing and Information Technology Al-Madinah Al-
Munawwarah, Saudi Arabia, pp 95–100
45. Manda K, Satapathy SC, Rao KR (2012) Artificial bee colony
based image clustering. In: International conference on infor-
mation systems design and intelligent applications (INDIA 2012),
Visakhapatnam, India, 2012, pp 29–37
46. Hancer E, Ozturk C, Karaboga D (2012) Artificial bee colony
based image clustering. In: IEEE congress on evolutionary
computation, CEC 2012, Brisbane, Australia
47. Karaboga D (2010) Artificial bee colony algorithm. Scholarpe-
dia:6915. doi:10.4249/scholarpedia.6915
48. Omran M, Al-Sharhan S (2007) Barebones particle swarm
methods for unsupervised image classification. In: IEEE congress
on evolutionary computation 2007 CEC 2007, 25–28 Sept. 2007,
pp 3247–3252
49. Chou CH, Su MC, Lai E (2004) A new cluster validity measure
and its application to image compression. Pattern Anal Appl
7(2):205–220. doi:10.1007/s10044-004-0218-1
50. Man To W, Xiangjian H, Wei-Chang Y (2011) Image clustering
using particle swarm optimization. In: IEEE congress on evolu-
tionary computation (CEC’2011), 5–8 June 2011, pp 262–268
51. Davies DL, Bouldin DW (1979) A cluster separation measure.
IEEE Trans Pattern Anal Mach Intell PAMI-1(2):224-227.
doi:10.1109/tpami.1979.4766909
52. Zhao Q, Xu M, Franti P (2009) Sum-of-squares based cluster
validity index and significance analysis adaptive and natural
computing algorithms. In: Kolehmainen M, Toivanen P, Beli-
czynski B (eds). Lecture notes in computer science, vol 5495.
Springer, Berlin, pp 313–322. doi:10.1007/978-3-642-04921-7_32
53. Calinski T, Harabasz J (1974) A dendrite method for cluster
analysis. Commun Stat 3:1–27
54. Ball G, Hubert L (1965) ISODATA, A novel method of data
analysis and pattern classification. Tech. Rep. NTIS No. AD
699616, Standford Research Institute, Menlo Park, CA
55. Hartigan J (1975) Clustering algorithms. Wiley, New York
56. Arbelaez P, Fowlkes C, Martin D The Berkeley Segmentation
Dataset and Benchmark.http://www.eecs.berkeley.edu/Research/
Projects/CS/vision/grouping/segbench/
57. Milton F (1940) A comparison of alternative tests of significance
for the problem of m rankings. Ann Math Stat 11(1):86–92.
doi:10.2307/2235971
Pattern Anal Applic
123