S. N. Sulaiman and N. A. M.

Isa: Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation 26

Adaptive Fuzzy-K-means Clustering Algorithm
for Image Segmentation
Siti Noraini Sulaiman and Nor Ashidi Mat Isa, Member, IEEE
Abstract Clustering algorithms have successfully been
applied as a digital image segmentation technique in various
fields and applications. However, those clustering algorithms
are only applicable for specific images such as medical
images, microscopic images etc. In this paper, we present a
new clustering algorithm called Adaptive Fuzzy-K-means
(AFKM) clustering for image segmentation which could be
applied on general images and/or specific images (i.e.,
medical and microscopic images), captured using different
consumer electronic products namely, for example, the
common digital cameras and CCD cameras. The algorithm
employs the concepts of fuzziness and belongingness to
provide a better and more adaptive clustering process as
compared to several conventional clustering algorithms. Both
qualitative and quantitative analyses favour the proposed
AFKM algorithm in terms of providing a better segmentation
performance for various types of images and various number
of segmented regions. Based on the results obtained, the
proposed algorithm gives better visual quality as compared to
several other clustering methods.
1
.
Index Terms Adaptive Fuzzy-K-means Clustering (AFKM),
clustering, image segmentation, digital image processing.
I. INTRODUCTION
Clustering is a process of grouping a set of objects into
classes of similar characteristics. It has been extensively used
in many areas, including in the statistics [1], [2], machine
learning [3]-[5], pattern recognition [6]-[8], data mining [9]-
[14], and image processing [15], [16].
In digital image processing, segmentation is essential for
image description and classification. The technique is
commonly used by many consumer electronic products (i.e.,
conventional digital image) or in a specific application field
such as the medical digital image. The algorithms are
normally based on similarity and particularity, which can be
divided into different categories; thresholding [17], template
1
This work was partially supported by Majlis Kanser Nasional (MAKNA),
Kuala Lumpur, Malaysia, under the Grant Development of an Intelligent
Screening System for Cervical Cancer.
Siti Noraini Sulaiman is a PhD student at Imaging and Intelligent System
Research Team (ISRT), School of Electrical & Electronics Engineering,
Universiti Sains Malaysia, Engineering Campus, Penang, Malaysia (e-mail:
siti.noraini.sulaiman@hotmail.com).
Nor Ashidi Mat Isa is an Associate Professor, Imaging and Intelligent
System Research Team (ISRT), School of Electrical & Electronics
Engineering, Universiti Sains Malaysia, Engineering Campus, Penang,
Malaysia (phone: +604 5996051; fax: +604 5941023; e-mail:
ashidi@eng.usm.my).
matching [18], [19], region growing [20], [21], edge detection
[22], [23], and clustering [24].
Clustering algorithm has been applied as a digital image
segmentation technique in various fields such as engineering,
computer, and mathematics. Recently, the application of
clustering algorithms has been further applied to the medical
field, specifically in the biomedical image analysis wherein
images are produced by medical imaging devices. Previous
studies proved that clustering algorithms are capable in
segmenting and determining certain regions of interest in
medical images [25]-[32]. In a biomedical image segmentation
task, clustering algorithm is often deemed suitable since the
number of cluster for the structure of interest is usually known
from its anatomical information [30].
Among the clustering formulations based on minimizing
formal objective functions, the most widely used and studied
is the K-means (KM) clustering. KM is an exclusive clustering
algorithm, (i.e., data which belongs to a definite cluster could
not be included in another cluster). Although it is the most
favourable technique, it does have some weaknesses [13],
[33]:
1. It is dependent on initialization.
2. It is sensitive to outliers and skewed distributions.
3. It may converge to a local minimum.
4. It may miss a small cluster.
As a result, it may lead to poor or wrong representation of
data.
There are several clustering algorithms proposed to
overcome the aforementioned weaknesses. Fuzzy C-means
(FCM), an overlapping clustering that employs yet another
fuzzy concept, allows each data to belong to two or more
clusters at different degrees of memberships. In the FCM,
there is no clear, significant boundary between the elements if
they do, or do not belong to a certain class. In 2002, [34]
successfully proposed a modified version of K-means
clustering, namely, Moving K-Means (MKM) clustering. The
study proved that MKM possesses a great ability in
overcoming common problems in clustering, such as dead
centres and centre redundancy. Furthermore, the MKM was
proven to be effective in avoiding the centre from being
trapped in local minima. A number of studies have also
provided evidence that the MKM produced better performance
as compared to the conventional KM and FCM [26], [27].
In this paper, we introduce a new version of clustering
algorithm called Adaptive Fuzzy-K-means (AFKM) clustering
algorithm. As mentioned, clustering is the process of
organizing objects into groups wherein members are similar
on certain aspects. In most clustering models, the concept of
similarity is based on distances, such as the Euclidean
distance. Agarwal and Mustaffa claimed that simply looking at
Contributed Paper
Manuscript received 10/08/10
Current version published 12/23/10
Electronic version published 12/30/10. 0098 3063/10/$20.00 2010 IEEE
IEEE Transactions on Consumer Electronics, Vol. 56, No. 4, November 2010 26
the Euclidean distance between two points was not sufficient
[33]. Therefore, in the proposed algorithm, we establish the
fuzzy concept to be applied right after the member is assigned
to its respective centre via Euclidean distance. Next, we
introduce the concept of belongingness to measure the
relationship between the centre and to ensure that its
membership meets certain criteria. The degree of membership
is updated based on its degree of belongingness. Hence, the
positions of the centres are recalculated based upon the
updated membership function. These features are introduced
in order to provide a better and more adaptive clustering
process.
This paper is organized as follow: Section II describes in
detail the proposed AFKM clustering algorithm. Section III
presents the data used and also discusses the type of analyses
applied to test the capability of the proposed algorithm.
Section IV presents the segmentation results obtained by the
proposed algorithm. In addition, a comparison of performance
comparison with several selected conventional clustering
algorithms is also presented. The comparison is done based on
both qualitative and quantitative analyses. Finally, Section V
concludes the work focused on of this paper
II. THE PROPOSED CLUSTERING ALGORITHM
As mentioned, the KM clustering is the most widely used
clustering algorithm due to its simplicity. The objective of the
algorithm is to minimize an objective function in order to
assign a group of data to its centre. However, the performance
of KM is still limited due to the weaknesses it displays, as
indicated in Section I. Thus, the MKM introduced a fitness
concept together with the basic concept of KM to improve the
segmentation performance. The FCM is also introduced to
overcome the KM weaknesses by incorporating fuzzy concept
in its implementation.
In this section a new version of clustering algorithm namely
AFKM is introduced. The proposed AFKM is specifically
design to incorporate both the fundamental theories of the
conventional K-means and MKM clustering algorithms (i.e.,
assigning each data to its closest centre or cluster) and the
conventional Fuzzy C-means (FCM) clustering algorithm (i.e.,
allows the data to belong to two or more clusters or centres).
For details on the implementation of the proposed AFKM,
consider a digital image with RS pixels (i.e., R represents
number of columns and S represents number of rows) to be
clustered into n
c
regions or clusters. Let p(x,y) be the
considered pixel and c
j
as the j-th centre, where x = 1, 2, ...., R,
y=1, 2, ...., S and j = 1, 2, ., n
c
. Based on the aforementioned
consideration, for the conventional KM and MKM, all data
will be assigned to the nearest centre based on Euclidean
distance. The new position for each centre is calculated using
[34]:
( )
_ _
e e
=
j j j
c x c y c
j
y x p
n
c ,
1
(1)
For the conventional FCM, the process of allocating each
data member to be assigned simultaneously to more than one
class is based on the following membership function [34]:
( )
( )
( )
;
1
1
) 1 /( 2
,
,
,
_
=

|
|
.
|

\
|
=
c
n
k
m
y x kp
y x jp
y x jp
m
d
d
M if
( )
( ) y x p j d
y x kp
, , , 0
,
> (2)
( ) 1 , = y x kp
m
M if
( )
; 0
,
=
y x kp
d (3)
( ) ; 0 , = y x jp
m
M for ( ) k y x p = ,
where d
jp(x,y)
is distance from point (x,y) to the current cluster
centre j, d
kp(x,y)
is distance from point (x,y) to other cluster
centres k, n
c
is number of centres and m is an integer, m >1
which determine the degree of fuzziness.
In the conventional FCM, the class membership is assigned
based on the distance between two points. Meanwhile, in the
MKM, the basic concept introduced suggests that each cluster
should have a significant number of members or final fitness
value and the differences among the clusters should be
minimized to ensure a good clustering process.
For the proposed AFKM algorithm, all centres are firstly
initialized to a certain value. In order to ensure a better
clustering process, (1) is no longer employed to update the
centre, hence, fuzziness and belongingness concepts are
introduced in the proposed AFKM algorithm. The concept of
fuzzy partitioning concept is applied to allow each data to be
assigned to more than one class simultaneously by different
degrees of membership, ( ) y x jp
m
M , . The membership function
( ) y x jp
m
M , is determined by using (2) and (3). In AFKM, the
membership function measurement is performed within the
data members without any influence from outsiders (non-
members). Usually, the closer the connection of the data, the
bigger the significance of impact on the clustering results.
Thus, this provides a significant impact on the degree of
membership, which is then used in determining the new
position of the centre.
For a good clustering process, the modification concept
introduced in the AFKM algorithm suggests that each cluster
should have a significant value of belongingness which
measures the relationship strength between the centre and its
members. Therefore in the proposed AFKM, after specifying
the membership for each data, the degree of belongingness, B
j
for each cluster is calculated. This relationship provides a
significant impact on the clustering results through the degree
of belongingness. Higher degree of belongingness shows a
stronger relationship between the centre and its members,
which will ensure a better data clustering. The degree of
belongingness, B
j
is calculated based on:
( ) y x jp
m
j
j
M
c
,
= (4)
In order to improve the clustering process, it is necessary
for the degree of membership is needed to always be updated.
The degree of membership is optimized based on the degree of
belongingness to ensure that the procedure of the reassigning
member places the data to its appropriate centre or cluster. By
introducing the concept of belongingness, the AFKM suggests
S. N. Sulaiman and N. A. M. Isa: Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation 26
that each centre or cluster should have members with a strong
relationship formed between them and the difference of
belongingness among the clusters should be minimized. This
criterion could be achieved using the following processes. The
value of ( ) y x jp
m
M , is updated in the iteration according to:
( ) ( ) ( ) ( ) y x jp
m
y x jp
m
y x jp
m
M M M , , ,
'
A + = (5)
where ( ) ( )
'
, y x jp
m
M is the new membership. ( ) y x jp
m
M , A is
defined as:
( ) ( )( )
j j
y x jp
m
e c M o = A ,

(6)
where is a designed constant with value between 0 and 1 and
is typically set to 0.1. Subsequently, the value of e
j
is
calculated according to:
j j j
e B B =

(7)
where
j
B

is the normalized value for degree of belongingness.

Finally the new centre positions of the entire existing
clusters are calculated based on the new (optimized)
membership function according to:
( )
( ) ( )
( )
( )
_ _
_ _
e e
e e
=
j j
j j
c x c y
m
y x jp
c x c y
m
y x jp
j
M
y x p M
c
'
,
'
,
,
(8)
(Note: for this work, degree of fuzziness, m is set to 2)
All processes are repeated until the values of all centres are
no longer change.
III. DATA SAMPLES AND ANALYSIS
The proposed AFKM has been tested on numerous outdoor
and indoor images taken using the most commonly used
consumer electronic product namely the digital camera. In this
paper, ten out of those tested images are chosen to be used to
visualize the performance of the proposed algorithm. These
images are called Air Force, Peppers, Elaine, Smarties1,
Pirate, Smarties2, Golden Gate, Butterfly, River Bridge and
Boat, as shown in Figs. 1(a) to 1(j) respectively. In order to
further prove the capability and suitability of the proposed
AFKM as a segmentation technique for a conventional
consumer electronics field (i.e., conventional digital image),
two conventional clustering algorithms namely MKM and
FCM are used as a comparison. The FCM clustering is chosen
for comparison as both of the proposed AFKM and the
conventional FCM algorithms employ a similar fundamental
concept for clustering process. The MKM clustering is chosen
because it has been proven to produce better segmentation
performance as compared to the conventional KM clustering
algorithm [25-27]. For all clustering algorithms, centre
initializations are set according to:
( )
|
|
.
|

\
|
+ + =
c
y x p y x p
y x p j
n
min max
j min c
2
1 2
) , ( ) , (
) , (
(9)
where min
p(x,y)
and max
p(x,y)
are the minimum and the
maximum value of gray level in the image respectively.
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
(j)
Fig. 1. The original images of (a) Air Force (b) Peppers (c) Elaine (d)
Smarties1 (e) Pirate (f) Smarties2 (g) Golden Gate (h) Butterfly (i) River
Bridge (j) Boat.
The performance analysis is carried out based on qualitative
and quantitative analyses. The qualitative analysis refers to the
visual interpretation that is observed via unaided human visual
perception. In this case, it determines the capability of the
segmentation algorithms in detecting and distinguishing the
regions of interest from unwanted backgrounds. In our studies,
the qualitative analysis is a judged by a panel of experts from
Universiti Sains Malaysia and Universiti Teknologi MARA
Malaysia. Nonetheless, the quantitative analysis enables the
performance of the segmentation results produced by the
proposed algorithm and the other algorithms to be evaluated
statistically. Here, three evaluation functions are used as the
quantitative benchmarks. The three functions are;
F(I) proposed by Liu and Yang [ 35],
( )
_
=
=
R
i
i
i
A
e
R I F
1
2
(10)
F(I) proposed by Borsotti et al. [36],
( )
( )
( ) | |
_ _
= =
+

=
R
i
i
i
Max
A
A
A
e
A R
M N
I F
1
2
1
1
1
1000
1
(11)
And Q(I) the improved version from F(I) proposed by Borsotti
et al. [36],
( )
( )
( )
_
=
(
(

|
|
.
|

\
|
+
+
=
R
i
i
i
i
i
A
A R
A
e
R
M N
I Q
1
2
2
log 1 1000
1

(12)
IEEE Transactions on Consumer Electronics, Vol. 56, No. 4, November 2010 26
For the above three formulae, I is the resultant image to
be evaluated, N M is the image size, R is the number of
regions found, A
i
is the size of the i-th region, and R(A
i
) is
the number of regions having area A
i
. e
i
, is defined as the
sum of Euclidean distances between the features of pixels
of region I and the corresponding region in the segmented
image. With reference to (11), R(A) is the number of
regions having exactly area A, and Max denotes the area
of the largest region in the segmented image. These
functions allow the segmentation to be evaluated without
labelling the image and without the requirement of user-
set parameters. Moreover, these functions correspond
more closely to visual judgement. Smaller values of F(I),
F(I), and Q(I), show better segmentation results.
IV. RESULT AND DISCUSSION
As mentioned in section III, two analyses are applied
namely the qualitative and quantitative analyses. The
results for both these analyses will be presented in Section
IV (A) and IV (B) respectively.
A. Qualitative Analysis
We have tested the AFKM on various images with
different numbers of clusters, some of which are
illustrated in Figs. 2 to 4. In this study, the main criterion
used to evaluate the segmentation performance of the
proposed algorithms is based on the capability to outline
the desired important regions in the image.
First, the proposed AFKM, the FCM and the MKM
clustering algorithms are tested on the aforementioned ten
standard images with the number of clusters being set to
3. The resultant images are illustrated in Fig. 2. Then we
compare the capability of those clustering algorithms in
providing good image segmentation performance as the
number of clusters increases. Figs. 3 and 4 illustrate the
image segmentation performance as the number of
clusters is set to 4 and 5 respectively. Significant changes
on the images are highlighted with arrows and/or circles.
Let us consider the segmentation results on the three
standard images (i.e., Air Force, Elaine, and Golden
Gate) when the number of cluster is set to 3. From the
first row of Fig. 2, (i.e., the Air Force image) it can be
seen that the conventional algorithms and the proposed
AFKM algorithm produce almost very similar results; but
as compared with the FCM, the AFKM shows a better
shading result particularly on the mountain area.
Moreover, the label U.S Air Force on the body of plane
appears clearer and sharper in the resultant image from
the proposed AFKM as compared to other images
produced by the conventional clustering algorithms. The
picture becomes clearer as we increase the number of
cluster to 4 and then 5 (as depicted in Figs 3, and 4,
respectively), with the U.S Air Force label and the
mountain area become sharper and crispier. In Fig. 5
(i.e., the cropped image of Air Force on its tail area) it
can be seen that the details on the tail of the plane are
remain unaffected by the proposed AFKM as compared to
the FCM, and the MKM. The serial number, highlighted
with a circle and arrow is sharper and clearer for the
image after the application of the proposed AFKM, while
in the other resultant images produced by the FCM, and
the MKM, the serial number is corrupted with noise (i.e.,
occurrence of white spots on the serial numbers). Overall,
the proposed AFKM algorithm has outperformed other
clustering algorithms. The segmented images produced by
the proposed AFKM are sharper and crispier with less
noisy pixels.
AirForce
(512 512)
Pepper
(512 512)
Elaine
(512 512)
Smarties1
(256256)
Pirate
(512 512)
Smarties2
(256256)
Golden
Gate
(512 512)
Butterfly
(512 512)
River
Bridge
(512 512)
Boat
(512 512)
Fig. 2. Resultant segmented images with 3 clusters after applying: First
column: Original images. Second column: FCM. Third column: MKM.
Forth column: AFKM.
S. N. Sulaiman and N. A. M. Isa: Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation 26
AirForce
(512 512)
Pepper
(512 512)
Elaine
(512 512)
Smarties1
(256256)
Pirate
(512 512)
Smarties2
(256256)
Golden
Gate
(512 512)
Butterfly
(512 512)
River
Bridge
(512 512)
Boat
(512 512)
Fig. 3. Resultant segmented images with 4 clusters after applying: First
column: Original images. Second column: FCM. Third column: MKM.
Forth column: AFKM.
For the Elaine image as shown in the third row of Fig. 2, as
we cluster the image into three regions, the folded line on the
Elaines towel is seen to be unclear for the resultant images
produced by the FCM and the MKM. Nonetheless, the
proposed algorithm has lined the folded towel much better
than the conventional techniques. Elaines hair and eyebrow
are segmented clearly by the proposed AFKM algorithm as
compared to the FCM, and the MKM, which segmented the
aforementioned areas rather poorly. The segmentation regions
of hair and eyebrow of Elaine, and the folded line on the
Elaines towel produced by the proposed AFKM are much
clearer as we increase the number of cluster from 3 to 4 and 5.
The resultant images for number of clusters set to 4 and 5 are
shown in Figs. 3, and 4, respectively. A more significant
comparison between the clustering algorithms can be observed
from the cropped image of Elaine on her towel as shown in
Fig. 6.
AirForce
(512 512)
Pepper
(512 512)
Elaine
(512 512)
Smarties1
(256256)
Pirate
(512 512)
Smarties2
(256256)
Golden
Gate
(512 512)
Butterfly
(512 512)
River
Bridge
(512 512)
Boat
(512 512)
Fig. 4. Resultant segmented images with 5 clusters after applying: First
column: Original images. Second column: FCM. Third column: MKM.
Forth column: AFKM.
IEEE Transactions on Consumer Electronics, Vol. 56, No. 4, November 2010 26
Fig. 5. The segmentation results (3 clusters) for cropped image of
AirForce. First column: Original images. Second column: FCM. Third
column: MKM. Forth column: AFKM.
Fig. 6. The segmentation results (4 clusters) for cropped image of Elaine.
First column: Original images. Second column: FCM. Third column:
MKM. Forth column: AFKM.
Fig. 7. The segmentation results (5 clusters) for cropped image of Golden
Gate. First column: Original images. Second column: FCM. Third
column: MKM. Forth column: AFKM.
With reference to the seventh row of Fig.2, (i.e., the
Golden Gate image) it can be seen that again, the
proposed AFKM algorithm has successfully produced a
better segmentation performance. The main structure of
the bridge image namely the beams and the pillars are
significantly detected and segmented while the
conventional FCM clustering algorithm have evidently
produced poor segmentation performance. The same
observation can also be seen when the number of cluster
is set to 4 and 5 as shown in Figs. 3 and 4 respectively.
The aforementioned advantages can be clearly seen from
the cropped Golden Gate image as shown in Fig. 7.
For the remaining seven images, the results obtained in
Figs. 2 to 4 further favour the proposed AFKM as the best
segmentation technique compared to the FCM and the
MKM clustering algorithms. Significant differences
between those algorithms are highlighted with arrows.
Based on these results, the proposed algorithm can be
suggested to be used as post processing tools for digital
images; an added-value to existing digital image
processing techniques in recent consumer electronics
products (e.g., CCD camera).
B. Quantitative Analysis
As mentioned in Section III, three benchmark functions
are used to evaluate the proposed AFKM quantitatively.
Table 1, Table 2, and Table 3 summarise the quantitative
evaluations of the segmentation results for the
conventional clustering methods, and the proposed
AFKM. The proposed AFKM produces the smallest
values of F(I), F(I) and Q(I) for almost all images as well
as for almost all number of clusters used (i.e., 3, 4, and 5).
These promising quantitative results verify and support
the superiority shown by the proposed AFKM in the
qualitative analysis. Additionally, the proposed AFKM
manages to segment the image successfully with less
noisy pixels. In general, these observations indicate that
the AFKM might be a better approach in terms of image
segmentation application.
This study further investigates the capability of the
proposed AFKM algorithm with sixty four standard test
images frequently used in literature and real medical
images. Using the aforementioned quantitative evaluation
functions, the average performance of all algorithms on
these tested images is tabulated in Table 4. From the table
it can be seen the proposed AFKM algorithm provides a
relatively lower value of each evaluation function which
demonstrate its quality in improving segmentation
capability.
In addition, the processing time analysis is also
performed to measure the algorithm efficiency. The
analysis is performed using Intel

Core i3 CPU
3.07GHz Processor, 2.00GB of RAM, and 500GB of disk
drive space. The results are tabulated in Table 5. For
almost all images, the proposed AFKM clustering
algorithm acquired a slightly higher processing time as
compared to the conventional clustering MKM clustering
algorithm. However, the processing time for the FCM and
AFKM are almost similar. Overall, with significant better
performances although with a slightly higher processing
time, the results obtained have successfully proven that
the proposed AFKM is highly capable to be applied as a
segmentation tool of digital images for consumers
electronic applications.
V. CONCLUSION
This paper presents a new clustering algorithm named
the AFKM algorithm for segmentation purposes. It
employs an adaptive and iterative fuzzy and
belongingness concept to obtain the optimum value of
clusters centre for a better segmentation process. The
proposed clustering algorithm is applicable to numerous
outdoor and indoor images whereby ten images have been
presented as case studies. The conclusion of this paper
sees the proposed algorithm outperforming the
conventional FCM and MKM algorithms by successfully
producing better segmented images. The proposed AFKM
also successfully preserves important features on digital
images. Both qualitative and quantitative analyses have
justified the conclusion that the proposed approach has
been able to illustrate good segmentation results
efficiently. Thus, it is recommendable for this algorithm
to be applied in the post image processing in consumer
electronic products such as the digital camera for general
applications and the CCD camera which is extensively
used with the microscope in capturing microscopic
images, especially in segmenting medical images.
S. N. Sulaiman and N. A. M. Isa: Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation 26
TABLE 1.
QUANTITATIVE EVALUATION F(I) ON SEGMENTATION OF SELECTED
IMAGES
F(I) For Cluster Number (*1.0e+12)
Algorithm Images
3 4 5
AirForce 36.6533 54.0284 76.0762
Pepper 10.7947 21.8951 28.6708
Elaine 9.0628 20.8674 37.4632
Smarties1 3.3909 6.9606 9.8462
Pirate 4.7425 7.1192 755.6790
Smarties2 2.9503 5.4771 7.9805
GoldenGate 4.0661 11.3159 15.2695
Butterfly 3.8466 7.3766 12.4678
River Bridge 6.7996 10.2375 17.8349
FCM
Boat 16.2922 24.3458 33.6430
AirForce 13.3609 30.9955 52.2924
Pepper 5.8701 9.6614 14.1086
Elaine 4.0949 8.0236 15.3663
Smarties1 2.1560 3.6655 5.3134
Pirate 7.4972 7.6087 19.5192
Smarties2 1.7981 2.9323 4.3565
GoldenGate 3.4956 2.9058 7.2228
Butterfly 3.1593 4.3336 6.0057
River Bridge 5.8340 7.3061 9.8613
MKM
Boat 6.2771 10.8656 17.3629
AirForce 8.8518 21.6530 42.4109
Pepper 6.2680 9.3014 14.3828
Elaine 3.8346 5.8647 10.0399
Smarties1 2.1071 3.1866 4.2387
Pirate 5.0881 6.5205 12.1597
Smarties2 1.7435 2.8110 3.4876
GoldenGate 2.4505 4.4111 5.6799
Butterfly 3.1770 4.1881 5.6341
River Bridge 5.5767 6.9262 9.3933
AFKM
Boat 6.4890 10.4207 15.3653
Note: Bolded numbers show the best result obtained for each analysis. The
same process is used in Table 2 to 4.
TABLE 2.
QUANTITATIVE EVALUATION F(I) ON SEGMENTATION OF SELECTED
IMAGES
F(I) For Cluster Number (*1.0e+5)
Algorithm Images
3 4 5
AirForce 1.3982 2.0610 2.9021
Pepper 0.4118 0.8352 1.0937
Elaine 0.3457 0.7960 1.4291
Smarties1 0.5174 1.0621 1.5024
Pirate 0.1809 0.2716 28.8269
Smarties2 0.4502 0.8357 1.2177
GoldenGate 0.1551 0.4317 0.5825
Butterfly 0.1467 0.2814 0.4756
River Bridge 0.2594 0.3905 0.6803
FCM
Boat 0.6215 0.9287 1.2834
AirForce 0.5097 1.1824 1.9948
Pepper 0.2239 0.3686 0.5382
Elaine 0.1562 0.3061 0.5862
Smarties1 0.3290 0.5593 0.8108
Pirate 0.2860 0.2903 0.7446
Smarties2 0.2744 0.4474 0.6648
GoldenGate 0.1333 0.1108 0.2755
Butterfly 0.1205 0.1653 0.2291
River Bridge 0.2225 0.2787 0.3762
MKM
Boat 0.2395 0.4145 0.6623
AirForce 0.3377 0.8074 1.6178
Pepper 0.2391 0.3548 0.5487
Elaine 0.1463 0.2237 0.3830
Smarties1 0.3215 0.4862 0.6468
Pirate 0.1941 0.2487 0.4639
Smarties2 0.2660 0.4289 0.5322
GoldenGate 0.0935 0.1683 0.2167
Butterfly 0.1212 0.1598 0.2149
River Bridge 0.2127 0.2642 0.3583
AFKM
Boat 0.2475 0.3975 0.5861
TABLE 3.
QUANTITATIVE EVALUATION Q(I) ON SEGMENTATION OF SELECTED
IMAGES
Q(I) For Cluster Number (*1.0e+7)
Algorithm Images
3 4 5
AirForce 0.2728 0.3820 0.5103
Pepper 0.1466 0.2475 0.3241
Elaine 0.1142 0.1989 0.2916
Smarties1 0.0674 0.1120 0.1428
Pirate 0.0887 0.1432 0.4586
Smarties2 0.0640 0.1010 0.1332
GoldenGate 0.0467 0.0975 0.1276
Butterfly 0.0640 0.1142 0.1659
River Bridge 0.1086 0.1783 0.2733
FCM
Boat 0.1933 0.2630 0.3403
AirForce 0.1541 0.2796 0.4130
Pepper 0.0997 0.1619 0.2296
Elaine 0.0726 0.1317 0.2123
Smarties1 0.0474 0.0734 0.1010
Pirate 0.1239 0.2028 0.3294
Smarties2 0.0440 0.0664 0.0932
GoldenGate 0.0362 0.0666 0.1316
Butterfly 0.0576 0.0924 0.1304
River Bridge 0.1030 0.1724 0.2518
MKM
Boat 0.1016 0.1646 0.2434
AirForce 0.1177 0.2216 0.3620
Pepper 0.1035 0.1586 0.2287
Elaine 0.0693 0.1120 0.1708
Smarties1 0.0465 0.0662 0.0849
Pirate 0.0967 0.1462 0.2229
Smarties2 0.0430 0.0642 0.0791
GoldenGate 0.0330 0.0548 0.0697
Butterfly 0.0578 0.0902 0.1209
River Bridge 0.0986 0.1643 0.2404
AFKM
Boat 0.1039 0.1572 0.2187
TABLE 4.
AVERAGE SEGMENTATION EVALUATION FUNCTIONS ON 64 IMAGES
Quantitative evaluation functions
F(I) F'(I) Q(I)
Number of
Cluster
Algorithms
(*1.0e+
12
) (*1.0e+
5
) (*1.0e+
6
)
FCM 7.8030 0.3269 0.8374
MKM 10.9354 0.4441 1.0336 3
AFKM 6.1938 0.2609 0.8495
FCM 13.6391 0.5766 1.4149
MKM 6.7291 0.2902 1.4543 4
AFKM 5.6972 0.2448 1.2936
FCM 46.4756 1.8677 2.2439
MKM 9.5417 0.4205 2.1243 5
AFKM 8.7510 0.3744 1.9115
TABLE 5.
EXECUTION TIME (IN SECONDS)
Algorithms and Cluster Number
FCM MKM AFKM
Images
3 4 5 3 4 5 3 4 5
AirForce 4.9 4.65 4.40 2.46 2.76 2.55 4.74 5.47 6.36
Pepper 5.14 5.24 5.33 2.46 2.46 2.57 3.77 3.80 4.52
Elaine 3.27 4.20 4.55 2.46 2.64 2.59 4.03 5.05 5.40
Smarties1 1.47 1.68 1.89 0.97 1.35 1.31 1.37 1.93 1.97
Pirate 2.95 4.71 5.71 2.43 2.47 2.73 4.64 6.06 7.44
Smarties2 1.68 1.67 1.66 1.14 1.01 1.25 1.46 1.62 2.25
GoldenGate 3.16 5.12 4.45 2.39 2.55 2.51 4.50 4.93 6.08
Butterfly 3.20 4.52 5.13 2.56 2.61 2.49 3.39 3.58 4.71
River Bridge 3.46 4.46 6.34 2.42 2.44 2.60 3.51 3.65 3.87
Boat 4.73 4.72 5.68 2.56 2.71 2.54 3.83 4.33 4.73
IEEE Transactions on Consumer Electronics, Vol. 56, No. 4, November 2010 266
REFERENCES
[1] D. Auber, and M. Delest, A clustering algorithm for huge trees, Advances
in Applied Mathematics, vol. 31, no. 1 pp. 46-60 2003.
[2] A. S. Mahani, A. E. Carlsson, and R. Wessel, Motion repulsion arises from
stimulus statistics when analyzed with a clustering algorithm, Biological
Cybernetics, vol. 92, no. 4, pp. 288-291, 2005.
[3] T. Abeel, Y. V. d. Peer, and Y. Saeys, Java-ML: A machine learning
library, Journal of Machine Learning Research, vol. 10, pp. 931-934, 2009.
[4] A. Y. Al-Omary, and M. S. Jamil, A new approach of clustering based
machine-learning algorithm, Knowledge-Based Systems, vol.19, no.4, pp.
248-258, 2006.
[5] M. J. Rattigan, M. Maier, and D. Jensen. Graph clustering with network
structure indices, in Proceedings of the 24th International Conference on
Machine Learning. 2007. Corvallis, OR.
[6] P. C. Boutros, and A. B. Okey, Unsupervised pattern recognition: An
introduction to the whys and wherefores of clustering microarray data,
Briefings In Bioinformatics,vol. 6, no. 4, pp. 331-343, 2005.
[7] J. P. Jesan, The neural approach to pattern recognition, Ubiquity, vol.5,
no.7, 2004.
[8] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and
A. Y. Wu, An efficient k-means clustering algorithm: analysis and
implementation, IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 24, no. 7, pp. 881-892, 2002.
[9] H. Wang, W. Wang, J. Yang, and P. S. Yu, Clustering by pattern similarity
in large data sets, in Proceedings of the 2002 ACM SIGMOD International
Conference on Management of data. 2002, Madison, Wisconsin.
[10] M. Halkidi, and M. Vazirgiannis, Clustering validity assessment: finding the
optimal partitioning of a data set, Proceedings IEEE International
Conference on Data Mining. 2001.
[11] Z.-H. Zhou, Three perspectives of data mining, Artificial Intelligence, vol.
143, no. 1, pp. 139-146, 2003.
[12] T. Zhang, R. Ramakrishnan, and M. Livny, BIRCH: an efficient data
clustering method for very large databases, in Proceedings of the 1996 ACM
SIGMOD international conference on Management of data, 1996, Montreal,
Quebec, Canada.
[13] C. Ordonez, Clustering binary data streams with K-means, in Proceedings
of the 8th ACM SIGMOD workshop on Research issues in data mining and
knowledge discovery. 2003, San Diego, California.
[14] J. Handl, J. Knowles, and D. B. Kell, Computational cluster validation in
post-genomic data analysis, Bioinformatics, 2005
[15] S. K. Singh, K. Shishir, G. S. Tomar, K. Ravi, and G. K. A. Santhalia,
Modified framework of a clustering algorithm for image processing
applications, in First Asia International Conference on Modelling &
Simulation, AMS '07, 2007.
[16] F. Hoeppner, Fuzzy shell clustering algorithms in image processing: fuzzy
C-rectangular and 2-rectangular shells, IEEE Transactions on Fuzzy
Systems,vol. 5, no. 4, pp. 599-613, 1997.
[17] M. Cheriet, J. N. Said, and C. Y. Suen, A recursive thresholding technique
for image segmentation, IEEE Transactions on Image Processing, vol. 7,
no. 6, pp. 918-921, 1998.
[18] S. K. Warfield, K. Michael, F. A. Jolesz, and K. Ron, Adaptive, template
moderated, spatially varying statistical classification, Medical Image
Analysis, vol. 4, no. 1, pp. 43-55, 2000.
[19] M. Lalonde, M. Beaulieu, and L. Gagnon, Fast and robust optic disc
detection using pyramidal decomposition and Hausdorff-based template
matching, IEEE Transactions on Medical Imaging, vol. 21, no. 11, pp.
1193-1200, 2001.
[20] Y. Seunghwan, and P. Rae-Hong, Red-eye detection and correction using
inpainting in digital photographs, IEEE Transactions on Consumer
Electronics, vol. 55, no. 3, pp. 1006-1014, 2009.
[21] N. A. Mat-Isa, M. Y. Mashor, and N. H. Othman, Automatic seed based
region growing for pap smear image segmentation, in Kuala Lumpur
International Conference on Biomedical Engineering. 2002. Kuala Lumpur,
Malaysia.
[22] J. K. Paik, Y. C. Park, and S. W. Park, An edge detection approach to digital
image stabilization based on tri-state adaptive linear neurons, IEEE
Transactions on Consumer Electronics, vol. 37, no. 3, pp. 521-530, 1991.
[23] Y. Siyoung, K. Donghyung, and J. Jechang, Fine edge-preserving
deinterlacing algorithm for progressive display, IEEE Transactions on
Consumer Electronics,vol. 53, no. 3, pp. 1654-1662, 2009.
[24] X. Yang, Z. Weidong, C. Yufei, and F. Xin., Image segmentation with a
fuzzy clustering algorithm based on Ant-Tree, Signal Processing, vol. 88,
no. 10, pp. 2453-2462, 2008.
[25] N. A. Mat-Isa, M. Y. Mashor, and N. H. Othman, Pap smear image
segmentation using modified moving k-mean clustering, in Kuala Lumpur
International Conference on Biomedical Engineering. 2002. Kuala Lumpur
Malaysia.
[26] N. A. Mat-Isa, M. Y. Mashor, and N. H. Othman, Comparison of
segmentation performance of clustering algorithms for pap smear images, in
Proceeding of International Conference on Robotics, Vision, Information and
Signal Processing. January 2003. Penang.
[27] N. A. Mat-Isa, M. Y. Mashor, N. H. Othman and S. N. Sulaiman,
Application of moving k-means clustering for pap smear image processing,
in Proceeding of International Conference on Robotics, Vision, Information
and Signal Processing. January 2002. Penang Malaysia.
[28] A. E. O. Boudraa, M. Arzi, J. Sau, J. Champier, S. Hadj-Moussa, J. E.
Besson, D. Sappey-Marinier, R. Itti, and J. J. Mallet, Automated detection of
the left ventricular region in gated nuclear cardiac imaging, IEEE
Transactions on Biomedical Engineering, vol. 43, no. 4, pp. 430-436, 1996.
[29] A. E. O. Boudraa, J. J. Mallet, J. E. Besson, S. E. Bouyoucef, and J.
Champier, Left ventricle automated detection method in gated isotopic
ventriculography using fuzzy clustering, IEEE Transactions on Medical
Imaging, vol. 12, no. 3, pp. 451-465, 1993.
[30] C. W. Chen, J. Luo, and K. J. Parker, Image segmentation via adaptive K-
mean clustering and knowledge-based morphological operations with
biomedical applications, IEEE Transactions on Image Processing, vol. 7,
no. 12, pp. 1673-1683, 1998.
[31] M. R. Rezaee, P. M. J. van der Zwet, B. P. E. Lelieveldt, R. J. van der Geest,
and J. H. C. Reiber, A multiresolution image segmentation technique based
on pyramidal segmentation and fuzzy clustering, IEEE Transactions on
Image Processing, vol. 9, no. 7, pp. 1238-1248, 2000.
[32] N. A. Mat-Isa, A. S. Samy, U. K. Ngah, Adaptive fuzzy moving K-means
algorithm for image segmentation, IEEE Transactions on Consumer
Electronics, vol. 55, no. 4, pp. 2145-2153, 2009
[33] P. K. Agarwal, and N. H. Mustafa, k-means projective clustering, in
Proceedings of the twenty-third ACM SIGMOD-SIGACT-SIGART
symposium on Principles of database systems, 2004, Paris, France.
[34] M. Y. Mashor, Hybrid Training Algorithm for RBF Network,
International Journal of Computer, Internet and Management, vol. 8, no. 2,
pp. 50-65, 2000.
[35] J. Liu, and Y. H. Yang, Multiresolution color image segmentation, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 16, no. 7,
pp. 689-700, 1994.
[36] M. Borsotti, P. Campadelli, and R. Schettini, Quantitative evaluation of
color image segmentation results, Pattern Recognition Letters, vol. 19, no.
8, pp. 741-747, 1998.
BIOGRAPHIES
Siti Noraini Sulaiman obtained her B.Eng Hons in
Electrical and Electronics Engineering from Universiti
Sains Malaysia in 2000 and MSc in Medical Imaging
from the same university in 2003. She is currently on
leave from her current post as lecturer at Faculty of
Electrical Engineering, Universiti Teknologi MARA,
Malaysia. She is now attached with Imaging and
Intelligent System Research Team (ISRT), School of Electrical and Electronic
Engineering, Universiti Sains Malaysia, Engineering Campus in Nibong
Tebal, Penang, Malaysia and doing her PhD in Biomedical Engineering,
focusing on Medical Imaging at the university.
Associate Professor Dr. Nor Ashidi Mat Isa obtained
his B.Eng Hons (First Class Honors) in Electronic
Engineering from Universiti Sains Malaysia in 1999 and
PhD in Image Processing and Neural Networks from the
same university in 2003. He is currently an Associate
Professor and heading Imaging and Intelligent System
Research Team (ISRT), at the School of Electrical and
Electronics Engineering, Universiti Sains Malaysia,
Engineering Campus in Nibong Tebal, Penang, Malaysia. He specializes in
the area of intelligent systems, image processing, neural networks for medical
applications, and algorithms. He has published numerous research articles in
international journals and conference proceedings.