2021 International Conference on Information and Communication Technology

for Sustainable Development (ICICT4SD), 27-28 February, Dhaka

Image Segmentation Based on Fuzzy C Means

Clustering Algorithm and Morphological
2021 International Conference on Information and Communication Technology for Sustainable Development (ICICT4SD) | 978-1-6654-1460-9/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICICT4SD50815.2021.9396873

Reconstruction
Takowa Rahman ∗, Md. Saiful Islam ∗
∗Dept. of Electronics and Telecommunication Engineering, Chittagong University of Engineering and Technology, Chattogram-4349, Bangladesh
Email: [email protected], saiful05eee @gmail.com

Abstract—The purpose of segmentation is to depict an used without regard for spatial details [8], [9], [10]. To address
original picture in something easier to interpret. Generally, in the problem, an algorithm is proposed to deal with this
image processing watershed algorithm is used essentially for problem which provides strong segmentation findings for
segmentation purposes which is fast and simple method and different digital images with low computation costs. In order
requires low computation time. But, it has disadvantages causing to improve noise immunity, the proposed procedure employs
excessive segmentation and this method is sensitive of falsifying morphological reconstruction [11] to smooth images.
edges. The fuzzy c means (FCM) technique is extremely successful
when segmenting images. Fuzzy c means clustering’s biggest The remainder of this article is structured as follows. The
advantage is the high identification rate and the lower false literature review of this research is defined in section Ⅱ.
location rate. Nevertheless, the fuzzy c means algorithm is noise- Proposed algorithm is suggested in section Ⅲ. Section Ⅳ
sensitive. To overcome these problems, an improved image describes the findings of the implementation by the
segmentation algorithm based on morphological reconstruction MATLAB of the suggested approach for image
and fuzzy c means algorithm is presented in order to improve the
performance of the segmentation. Firstly, principle component
segmentation.
analysis method is applied to reduce number of variables in data II. LITERATURE REVIEW
by extracting important one from large pool. Secondly,
morphological reconstruction operation is introduced which A. Watershed Algorithm
guarantees the immunity to noise. Thirdly, fuzzy c means
A watershed is a grayscale visual transition. Usually, when
algorithm is applied. Finally, digital images are segmented by
using this proposed method. Segmented findings indicate that
two regions of interest are close together – this is to say their
better segmentation efficiency than watershed algorithm and edge touches, this technique attempts to segment the image.
fuzzy c means algorithm were obtained with proposed approach. The transformation in the watershed processes the picture on
which it works as a topography map, the luminosity of each
Keywords—Image segmentation, Watershed algorithm, Fuzzy c spot is its height and finds the lines [3]. A complete division
means (FCM) clustering algorithm, Morphological reconstruction of the image can be generated by the help of the watershed
(MR). algorithm. However, it has disadvantages causing excessive
segmentation and falsified edge sensitivity [10], [4].
I. INTRODUCTION
B. Fuzzy C Means Clustring Algorithm
The segmentation of images is a way to separate an image
into various parts. This is a key to image processing and The FCM algorithm is one of the most common used
pattern recognition and it is essential for the identification of clustering algorithms. In 1981, Professor Jim Bezdek
objects, image recovery. Several researchers recently implemented this method. The FCM algorithm proposes to
proposed different segmentation methods based on modern divide into c fuzzy clusters a finite X= {...} entity based on a
mathematical theories, including threshold technology, edge particular criterion [12]. This algorithm is based on the
detection, region growth, neural network, watershed following objective function being minimized.

=∑ ∑ − ||²
transformation [1], spatial clustering, etc. [2]. Watershed
algorithm is a fast, easy way to divide the image into regions || (1)
in a complete way. This technique provides low calculation
time but it causes excessive segmentation [3], [4]. Among Here, m defines fuzziness exponent, N = Number of data,
these technologies, because of its effectiveness and speed, one C = Number of clusters, uij defines the degree of xi
of the most common approaches is clustering which is used to membership in j cluster, xi= ith of measured data in d
segment images [5].In this paper, clustering methods by dimensions, cj = Cluster dimension center, and ||*||= The
minimizing objective function image segmentation will be correlation of every assessed data with center. The total
focused on. One of the clustering methods classifications can number of memberships in each data point must be equal to
be determined whether the subsets are fuzzy or hard. Hard one [8].
clustering approaches are based on classical set theory.
Nevertheless, fuzzy clustering methods allow objects to be FCM Algorithm Parameters: The following parameters
part of many clusters at the same time with different must be defined before using the FCM algorithm. (1) The
membership levels [6]. FCM has more space for uncertainty number of clusters, c and (2) The value of the fuzziness
and preserves more original image details. It is superior to exponent, m, must be greater than 1[6].
hard clustering [7]. While in images with a clear texture and Steps of the FCM Algorithm:
backdrop, FCM is successful, this does not segment dynamic
texture and context imagery, since only gray information is • Step 1. Select at least 2 random centroid.

978-1-6654-1460-9/21/$31.00 ©2021 IEEE

259

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on July 04,2021 at 05:12:00 UTC from IEEE Xplore. Restrictions apply.
 • Step 2. Count membership matrix.
(2)
∑ !

• Step 3. Calculate the c cluster centers.

∑%! " . $ ⅰ) Image 1 ⅱ) Image 2 ⅲ) Image 3
∑%! "
(3)
Fig. 2. The tested images [15]

III. PROPOSED SEGMENTATION METHODOLOGY A. Principle Component Analysis

An improved image segmentation algorithm based on Principle component analysis (PCA) is a dimensionality-
morphological reconstruction and FCM algorithm is presented reduction approach that often reduces the dimension of large
in order to improve the performance of the segmentation. One data sets by translating a wide variety of variables into a
experiment is conducted to segment digital images. The smaller variable which yet contains most data in the larger set.
method consists in few processing steps including analysis of Here, this technique is used to reduce number of variables in
principle component, morphological reconstruction and fuzzy data by extracting important one from large pool. Fig. 3
c means clustering. The programming part is carried out with indicates the key feature of the checked images.
MATLAB. The methodology for this analysis Fig. 1
demonstrates step by step.

Input Image

Representing Principle Component which Determines

Special Feature
(ⅰ) (ⅱ) (ⅲ)
Fig. 3. Representing principle component of tested images
Applying Morphological Reconstruction to Guarantee
Immunity from Noise B. Morphological Reconstruction Operation
The convergence rate of the FCM algorithm is determined
Set Number of Cluster c, Fuzziness Exponent m, by the data delivery features. If the data delivery features are
Window Size s and Minimal Threshold for Error η satisfactory to the clustering, there are a small number of
iterations, otherwise there are large numbers of iterations.
Within this article, morphological reconstruction is applied to
Randomly Initialize the Matrix of Membership
Partition * +,-
the FCM algorithm for optimizing data distribution before
clustering is used. Without understanding the noise type in
advance, morphological reconstruction can maintain object
Set the Counter of the Loop t=0 contour and minimize noise which is good for optimizing data
delivery. Two basic reconstruction operations, morphological
Update the Centers of Clustering dilation and erosion are performed.
If A and B are two sets in & ' , the dilation of B by A which
Update Membership Partition Matrix * +./ - is indicated by B⊕A and can be defined as:

B ⊕ A = {Z | (Â) Ζ Ⴖ B≠ Ø }
max {* +.- − * +./ - }
(4)
Otherwise
<η?
If A and B are two sets in & ' , the erosion of B by A which
Yes is indicated by B⊖A and can be defined as:

Iteration Stop Β ⊖ Α = {Ζ |(Α) Ζ ⊆ Β} (5)

Segmentation Result The opening and closing are two other important
morphological reconstruction operations. Opening normally
smoothens an object contour by removing thin protrusions.
Detecting the Segmented Area Opening deletes smaller objects and positions them back in
the context of a picture while closing eliminates tiny holes in
Fig. 1. Block diagram for proposed method the front and turns tiny backdrop islands into the forefront
[16], [17]. If B is structuring element, the opening of set A by
The Input Image: Here randomly three different digital B can be denoted as B ◦ A = (B ⊖ A) ⊕ A. Similarly, if B
images have been used which are shown in Fig. 2. defines structuring element then the closing of set A by B can
be denoted as B● A = (B ⊕ A) ⊖ A.

260

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on July 04,2021 at 05:12:00 UTC from IEEE Xplore. Restrictions apply.
 In this paper, morphological closing reconstruction false sensitivity of the edge. Therefore the FCM algorithm has
operation is used. been used to produce better segmentation results than this
method.
C. Fuzzy C Means Clustering
Categorizing clustering methods into hierarchical,
theoretical graph, density decomposition and objective
function minimization can be classified. In this paper,
clustering method by minimizing objective function image
segmentation is focused on. Clustering is helpful in data
exploration. Clustering algorithms can be used to find natural
groups if there are several instances, and there is no obvious
group. Fuzzy c means segmentation technique is used in this (ⅰ) Output Image 1 (ⅱ) Output Image 2 (ⅲ) Output Image 3
article, since the efficiency of this approach to achieve an Fig. 6. Segmentation results of the test images obtained by using FCM
optimal solution depends on the starting position of the cluster algorithm.
centers, the membership degree calculation of each data. Here,
the number of clusters, c=3, the value of m must be greater From Fig. 6, using fuzzy c means clustering output images
than 1, window size s = 3 and minimal error threshold η have are more effective than watershed algorithm. Although FCM
been setup. Randomly matrix of the membership partition succeeds in clear texture and background images, dynamic
* +,- is initiated and counter of the loop is started with t=0. texture and context imagery are never segmented and FCM is
Clustering centers are updated to find the overlapped dataset also noise- sensitive.
and membership partition matrix * +./ - is also updated. If the
difference between the value of initial and updated
membership partition matrix is less than minimal error
threshold η, then the iteration will be stopped otherwise, the
loop will be continued. Finally, the outcome is discovered.
IV. RESULTS AND ANALYSIS
The implementation of the method to obtain significant
(ⅰ) when cluster (ⅱ) when cluster (ⅲ) when
results for digital images segmentation will be further C= 3 C= 5 cluster C= 9
discussed in this section. By comparing the proposed
methodology with the conservative watershed algorithm and (a) Output of the input image 1 using proposed algorithm
FCM clustering algorithm, the performance is evaluated.
Using Fuzzy C-means clustering with morphological
reconstruction, the issue of computing cost can be reduced and
produces more effective performance. The results of
segmentation are described below.
A. Performance of Proposed Method

(ⅰ) when cluster (ⅱ) when cluster (ⅲ) when

C= 3 C= 4 cluster C= 6

(b) Output of the input image 2 using proposed algorithm

(ⅰ) Original Image 1 (ⅱ) Original Image 2 (ⅲ) Original Image 3

Fig. 4. Tested Images [15]

Fig. 4 displays original images.

(ⅰ) when cluster (ⅱ) when cluster (ⅲ) when
C= 3 C= 5 cluster C= 15

(c) Output of the input image 3 using proposed algorithm

Fig. 7. Segmentation results of the test images obtained by using proposed
algorithm.

From Fig. 7, using the proposed method the segmentation

performance of output images improves than other two
(ⅰ) Output Image 1 (ⅱ) Output Image 2 (ⅲ) Output Image 3 methods. In this proposed algorithm, PCA is used to reduce
Fig. 5. Segmentation results of the input images obtained by using number of variables in data by extracting important one from
watershed algorithm. large pool. FCM is being implemented with morphological
reconstruction (MR) because MR cannot only produce good
Fig. 5 does not make images with watershed algorithm so results but also takes short period of time. In order to achieve
obvious. Although the watershed algorithm can produce a full better image segmentation, MR incorporates
division of the image, it causes excessive segmentation and spatialinformation into FCM. By increasing the number of

261

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on July 04,2021 at 05:12:00 UTC from IEEE Xplore. Restrictions apply.
cluster the segmentation quality of output images also TABLE II. COMPARISON OF RUN TIME (IN SECONDS) AND
PERFORMANCE OF THREE ALGORITHMS ON DIGITAL IMAGE 2
improve.
Algorithm Input Image 2 Execution
B. Comparison of Findings with Existing Method Output Image
[15] Time

Watershed 4.0033
Algorithm seconds

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

Fig. 8. Comparison of the results of the color image segmentation 1. (a) FCM 122.142
Input image [15]. (b) Using Watershed algorithm (Running time is Algorithm seconds
0.142603s). (c)Using FCM algorithm (Running time is 126.49s) (d) Result
of proposed method (Running time is 58.159 s).

Fig. 8 shows that execution time of watershed algorithm is

so short, 0.142603s but performance is poor. Though Proposed 53.3851
performance of segmentation result of FCM algorithm is Method seconds
better than watershed algorithm but execution time is so high,
126.49s. Now, using the proposed method the execution time
can be minimized which is 58.159s with better performance;
segmented areas are indicated by the circle.

TABLE I. COMPARISON OF RUN TIME (IN SECONDS) AND

PERFORMANCE OF THREE ALGORITHMS ON DIGITAL IMAGE 1 (a) (b) (c) (d)
Algorithm Input Image 1 Execution
Output Image
[15] Time
(a) (b) (c) (d)
Fig. 10. Comparison of the results of the color image segmentation 3. (a)
Watershed Input image [15]. (b) Using Watershed algorithm (Running time is
Algorithm 0.142603
10.5072s). (c)Using FCM algorithm (Running time is 28.0063s) (d) Result
seconds
of proposed method (Running time is 20.1267s).

Fig. 10 shows that execution time of watershed algorithm

is so small, 10.5072s but performance is poor. Though
performance of segmentation result of FCM algorithm is
FCM 126.49
Algorithm better than watershed algorithm but execution time is so high,
seconds
28.0063s. Now, using the proposed method the execution time
can be minimized which is 20.1267s with better performance.

TABLE III. COMPARISON OF RUN TIME (IN SECONDS) AND

PERFORMANCE OF THREE ALGORITHMS ON DIGITAL IMAGE 3
Proposed 58.159
Method seconds Input Image 3 Execution
Algorithm Output Image
[15] Time

Watershed 10.5072
Algorithm seconds

FCM 28.0063
(a) (b) (c) (d) seconds
Algorithm
Fig. 9. Comparison of the results of the color image segmentation 2. (a)
Input image [15]. (b) Using Watershed algorithm (Running time is 4.0033s).
(c)Using FCM algorithm (Running time is 122.14s) (d) Result of proposed
method (Running time is 53.3851s).

Fig. 9 shows that watershed algorithm execution time is so Proposed 20.1267

short, 4.0033s but performance is low. Although the FCM Method seconds
algorithm performs better than the watershed algorithm but
execution time is so high, 122.14s. Now, using the proposed
method the execution time can be minimized with improved
performance.

262

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on July 04,2021 at 05:12:00 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSIONS [7] H. R. Mohammed, H. H. Alnoamani, A. A. Jalil, “Improved Fuzzy C-
Mean Algorithm for Image Segmentation,” (IJARAI) International
The advantage of watershed algorithm is that the image is Journal of Advanced Research in Artificial Intelligence, vol. 5, 2016.
fully divided. Nevertheless, it causes excessive segmentation [8] R. BABUSKA, Fuzzy and Neural Control DISC Course Lecture Notes,
and is sensitive of falsifying edges. While the FCM algorithm October 2001.
is successful for images with basic texture and background [9] T. Lei, X. Jia, Y. Zhang, L. He, H. Meng, A. K. Nandi, “Significantly
and preserves more original image information, it does not Fast and Robust Fuzzy C-Means Clustering Algorithm Based on
segment images with complicated textures. This just Morphological Reconstruction and Membership Filtering,” IEEE
TRANSACTIONS ON FUZZY SYSTEMS, vol. XXX, 2017.
recognizes information on the gray level without taking spatial
[10] M.C.J. Christ, R.M.S.Parvathi, “MEDICAL IMAGE
knowledge into account. So an improved image segmentation SEGMENTATION USING FUZZY C-MEANS CLUSTERING AND
algorithm based on fuzzy C means clustering and MARKER CONTROLLED WATERSHED ALGORITHM,”
morphological reconstruction has been applied to improve the International Journal of Modern Engineering Research (IJMER), vol.2,
segmentation quality with minimizing the time for Jan-Feb 2012.
segmentation. Results from segmentation displays that the [11] R. C. Gonzalez, R. E. Woods, Digital Image Processing, 3rd ed., India:
proposed approach achieved improved segmentation Dorling Kindersley (India) Pvt. Ltd., 2012, pp. 630-636.
performances than other two methods. [12] R.Dhivya, R.Prakash, R.Thilepa, “An Approach on Edge Detection in
Images Using Fuzzy C Means Clustering Method,” Imperial Journal of
REFERENCES Interdisciplinary Research (IJIR), vol.2, 2016.
[13] S. Kumar, S. K. Ray, “A Hybrid Approach for Image Segmentation
[1] K Lalitha, R Amrutha, S. Michahial, Dr M Shivakumar, Using Fuzzy Clustering and Level Set Method,” International Journal
“Implementation of Watershed Segmentation,” International Journal of of Image, Graphics and Signal Processing, July 2012.
Advanced Research in Computer and Communication Engineering,
vol. 5, December 2016. [14] A. Kumari, S.Meenakshi, S. Abinaya, “Plant Leaf Disease Detection
Using Fuzzy C-Means Clustering Algorithm,” International Journal of
[2] X. Wang, J. Bu, “A fast and robust image segmentation using FCM Engineering Development and Research, vol. 6, 2018.
with spatial information,” Digital Signal Processing, vol. 20, pp. 1173-
1182, July 2010. [15] "Images of High Resolution Picture," [Online]. Available:
https://www.bing.com/search?q=high+resolution+picture+of+bird+an
[3] "Watershed Algorithm," [Online]. Available: Watershed (image d +flower&id=
processing) - Wikipedia [Accessed March 31,2020]. E6EC8478D4DE73F6F0772027536B6C6DFDCC463D
[4] W. L. Nowinski, H.P. Ng, S. H. Ong, K. W. C. Foong,“An improved &FORM=IQFRBA.
watershed algorithm for medical image segmentation,” 12th [16] P. Soille, Morphological Image Analysis: Principles and Applications,
International Conference on Biomedical Engineering, 2005. 2nd ed., Springer Science & Business Media, 14 March 2013.
[5] J. C. Bezdek, R. Ehrlich, W. Full, “FCM: The fuzzy c- means clustering [17] J. M. Sharif, M. F. Miswan, M. A. Ngadi, Md S. H. Salam, M. M.bin
algorithm,” Computers & Geosciences, vol. 10, pp. 191- 203,1984. A. Jamil, "Red Blood Cell Segmentation Using Masking and
[6] T. Saikumar, P. Yugander, P. S. Murthy, B. Smitha, “Image Watershed Algorithms: A Preliminary Study," 2012 International
Segmentation Algorithm Using Watershed Transform and Fuzzy C- Conference on Biomedical Engineering (ICoBE), 27-28 February
Means Clustering on Level Set Method,” International Journal of 2012.
Computer Theory and Engineering, vol. 5, April 2013.

263

Authorized licensed use limited to: BOURNEMOUTH UNIVERSITY. Downloaded on July 04,2021 at 05:12:00 UTC from IEEE Xplore. Restrictions apply.