0% found this document useful (0 votes)
13 views11 pages

NG 2008

The document presents an improved fuzzy C-means (FCM) clustering algorithm that integrates Otsu thresholding to enhance accuracy and reduce susceptibility to local minima. The proposed method, FCM-LT, is evaluated against conventional FCM on various gray-scale images, demonstrating superior performance in terms of threshold values and region entropy. Results indicate that FCM-LT effectively differentiates image components better than traditional FCM, particularly in medical imaging applications.

Uploaded by

Ab Ir
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
13 views11 pages

NG 2008

The document presents an improved fuzzy C-means (FCM) clustering algorithm that integrates Otsu thresholding to enhance accuracy and reduce susceptibility to local minima. The proposed method, FCM-LT, is evaluated against conventional FCM on various gray-scale images, demonstrating superior performance in terms of threshold values and region entropy. Results indicate that FCM-LT effectively differentiates image components better than traditional FCM, particularly in medical imaging applications.

Uploaded by

Ab Ir
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

International Journal on Artificial Intelligence Tools

Vol. 17, No. 4 (2008) 765–775


 World Scientific Publishing Company

FUZZY C-MEANS ALGORITHM WITH LOCAL THRESHOLDING FOR


GRAY-SCALE IMAGES

HSIAO PIAU NG
Biomedical Imaging Lab, Agency for Science Technology and Research,
30 Biopolis Street, #07-01 Matrix, Singapore 138671
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

[email protected]
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

SIM HENG ONG


Department of Electrical and Computer Engineering, National University of Singapore,
21 Lower Kent Ridge Road, Singapore 119077

KELVIN WENG CHIONG FOONG


Department of Preventive Dentistry, National University of Singapore,
21 Lower Kent Ridge Road, Singapore 119077

POH SUN GOH


Department of Diagnostic Imaging, National University Hospital,
5 Lower Kent Ridge Road, Singapore 119074

WIESLAW L. NOWINSKI
Biomedical Imaging Lab, Agency for Science Technology and Research,
30 Biopolis Street, #07-01 Matrix, Singapore 138671

Received 20 September 2007


Accepted 10 April 2008

An improved fuzzy C-means (FCM) clustering method is proposed. It incorporates Otsu


thresholding with conventional FCM to reduce FCM’s susceptibility to local minima, as well as its
tendency to derive a threshold that is biased towards the component with larger probability, and
derive threshold values with greater accuracy. Thresholding is performed at the cluster boundary
region in feature space. A comparison of the results produced by improved and conventional
algorithms confirms the superior performance of the former.

Keywords: Fuzzy c-means; clustering; thresholding; Otsu.

1. Introduction
Clustering is a process for grouping a set of objects into classes or clusters so that objects
within a cluster have high similarity. Fuzzy clustering, which produces overlapping
cluster partitions based on a membership function, has been widely studied and applied in
various areas. In fuzzy clustering, the fuzzy c-means (FCM) algorithm1 is the best known

765
766 H. P. Ng et al.

method used in cluster analysis. FCM combines K-means clustering with fuzzy logic and
is useful for dealing with nontrivial data and uncertainties. It is an unsupervised
clustering technique that is commonly used in the pattern-recognition field as well as in
medical applications.2 Among its many medical applications, it has been used to segment
breast lesions from dynamic contrast-enhanced MR images3 and for brain tissue clustering.4,5
Despite its utility, there are some drawbacks with FCM. A common problem
addressed in literature is that its performance suffers from perturbation such as tissue
inhomogeneity and partial volume artifact. Context-based methods, such as regularized
FCM clustering,6,7 have been proposed to tackle this problem by taking into consideration
the spatial relationship among voxels to provide stable and reliable results. Another
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

drawback with FCM, which has received lesser attention, is that FCM may be susceptible
to local minima, depending on the complexity of the feature space as well as the initial
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

conditions. If the initial state is located near a local minimum, FCM will converge to this
point. Similar to most variance-based thresholding methods, FCM also tends to select a
threshold that is biased towards the component with the larger probability or larger
variance.8 In our work here, we show that the results produced by FCM can be improved
by performing Otsu thresholding9 in the region of the cluster boundary (FCM-LT).
The paper is organized as follows. In Section 2, the theory of conventional FCM
algorithm is first briefly reviewed, followed by a description of the proposed method
FCM-LT. Section 3 presents the results when FCM-LT and conventional FCM algorithm
are applied to test images. Discussion and comparisons between these results are in
Section 4 and finally, the paper is concluded in Section 5.

2. Method
2.1. FCM clustering algorithm
Given data D = {di, i = 1,2,…, N |di ∈ R}, FCM can be performed on it to obtain
c classes. Here, the feature space consists of the gray levels belonging to the pixels in
the image. The class prototypes are represented by a c by m matrix V, with the kth
(k = 1,2,…, c) row representing the prototype of the kth class and m the dimension of
data point d i . The cluster is represented by a c by N matrix U. The elements of matrix U,
uki, represent the membership of the ith data point to the kth class with values in the range
[0, 1]. With FCM clustering, uki ranges from 0 to 1 and characterizes the degree of
similarity between the ith data point and the kth class prototype:
c

∑u
k =1
ki = 1, ∀ i ; 0 ≤ uki ≤ 1 ∀ k , i (1)

The matrix U is found by minimizing the generalized least-squares within-group


square error function Jm:
c N

∑ ∑u b 2
Jm = ki d i − vk (2)
k =1 i =1
Fuzzy C-Means Algorithm with Local Thresholding for Gray-Scale Images 767

where b ∈ [1, ∝ ) is a weighting exponent on each fuzzy membership and || || denotes the
Euclidean distance. uki and vk are updated as follows:

1
uki = 2 / (b−1)
, k = 1,2,..., c ; i = 1,2,..., N (3)
c  d i − vk 
∑ 

l =1  d i − vl



N

∑u
i =1
b
ki d i
, k = 1,2,..., c
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

vk = N
(4)
∑ ukib
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

i =1

After convergence of the iteration process, interpretation of the membership function


is carried out and each voxel is classified to the cluster with the largest membership.

2.2. FCM with Local Thresholding (FCM-LT )


After obtaining the clusters formed by FCM, we compute the number of pixels belonging
to each of the clusters as a fraction fk of the size of the entire feature space, which in this
case is the total number of pixels in the image. For the kth class,
n( k )
fk = (5)
N
where n (k ) is the number of pixels belonging to kth class and N is the total number of
data points. We then perform Otsu thresholding on the pixels that lie in the region f k − r
to f k + r of the cumulative histogram, where r is a variable which we define to be the
tolerance to FCM. Illustrations for cases when k = 2 and k = 3 are in Figs. 1(a) and (b)
respectively.
Referring to Fig. 1(a), f1 is the fraction of the total number of pixels that the first
cluster occupies. Otsu thresholding is performed on the intensity values of the pixels
which fall within the region f1 – r to f1 + r in the cumulative histogram in order to arrive
at the final threshold value.
In a situation where there are 3 classes (Fig. 1(b)), Otsu thresholding is performed
on the intensity values of the pixels which fall within the regions 1 – f3 – r to 1 – f3 + r
and f1 – r to f1 + r in the cumulative histogram to derive the final upper and lower
threshold values respectively.

3. Results
The proposed method presented here is implemented using MATLAB 7.0 and applied to
20 different gray-scale images. The parameters and default values of the FCM function
are: weighting component b = 2.0, maximum number of iterations = 100, minimum
768 H. P. Ng et al.
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

(a)

(b)

Fig. 1. Regions in cumulative histogram where Otsu thresholding is performed when (a) k = 2 and (b) k = 3.

amount of improvement ε = 1 × 10–5. r = 0.1 is used in our work here. We have also
experimented with r = 0.05 and r = 0.15 and observed that there was little impact on the
results.
The results obtained using FCM-LT are compared against those obtained using the
conventional FCM algorithm by computing the entropy of the targeted region in each of
the test images. These, together with the derived threshold values are summarized in
Table 1. In addition, we provide some sample results to allow visual comparisons.
Fuzzy C-Means Algorithm with Local Thresholding for Gray-Scale Images 769

Table 1. Summary of results obtained using conventional FCM and FCM-LT.


Threshold Values Region Entropy
Test Image
FCM FCM-LT FCM FCM-LT
1 129 105 3.62 3.83
2 71 24 3.60 3.94
3 108 103 2.59 2.68
4 160 163 2.21 2.23
5 97 85 3.17 3.40
6 115 124 2.69 2.81
7 79 87 2.89 3.01
8 147 142 3.07 3.22
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

9 183 175 3.38 3.49


10 94 84 3.56 3.74
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

11 59, 149 38, 187 3.16 3.48


12 100, 175 94, 188 3.52 3.73
13 77, 158 67, 152 3.08 3.24
14 81, 157 67, 151 2.95 3.19
15 69,168 57, 173 3.37 3.54
16 82,172 75, 162 2.79 3.02
17 75,156 68, 160 3.17 3.24
18 89,175 78, 169 2.95 3.19
19 62,167 65, 170 3.04 3.11
20 92,182 83, 189 3.32 3.54

Figure 2 presents the bi-level thresholding results of the electronic circuit board
image (1st row, left), where in the 2nd row are the FCM result with threshold 129 (left)
and the FCM-LT result with threshold 105 (right). From the gray-level histogram
(1st row, right), the picture roughly consists of the electronic components and the circuit
board, with the probability of the latter being greater than that of the latter. Conventional
FCM produces a threshold (129) which is further away from the valley as compared to
that produced by FCM-LT (105). The entropy of the foreground obtained using FCM-LT
(3.83) is greater than that obtained using conventional FCM (3.62). Visually, FCM-LT
produces a result in which the electronic components are better differentiated from the
circuit board.
Figure 3 displays the bi-level thresholding results of the moon image (1st row, left),
where in the 2nd row are the FCM result with threshold 71 (left) and the FCM-LT result
with threshold 24 (right). Conventional FCM produces a threshold (71) which is unable
to extract the entire moon unlike the result obtained using threshold value produced by
FCM-LT (24). The entropy of the foreground obtained using FCM-LT (3.94) is greater
than that obtained using conventional FCM (3.60).
Figure 4 shows the tri-level thresholding results of a mammogram image (1st row,
left), where in the 2nd row are the FCM result with thresholds (59, 149) on the left and
the FCM-LT result with thresholds (38, 187) on the right. It can be easily observed that
conventional FCM produces thresholds which are unable to fully extract the targeted
770 H. P. Ng et al.
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

(a) (b)

(c) (d)
Fig. 2. (a) Electronic circuit board image (b) Histogram (c) Result using FCM (d) Result using FCM-LT.

object. This is unlike the threshold values produced by FCM-LT which extracts the
targeted object completely. The entropy of the targeted region obtained using FCM-LT
(3.48) is greater than that obtained using conventional FCM (3.16).
Figure 5 is another example of tri-level thresholding. The proposed method is
applied to a computed tomography image (1st row, left), where in the 2nd row are the
FCM result with thresholds (100, 175) on the left and the FCM-LT result with thresholds
(94, 188) on the right. It can be easily observed that there are many speckles in the gray
matter region of the result produced by conventional FCM. In contrast, the gray matter
region in the result produced by FCM-LT is more uniform. The entropy of the targeted
region obtained using FCM-LT (3.73) is greater than that obtained using conventional
FCM (3.52).
Fuzzy C-Means Algorithm with Local Thresholding for Gray-Scale Images 771
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

(a) (b)

(c) (d)
Fig. 3. (a) Moon image (b) Histogram (c) Result using FCM (d) Result using FCM-LT.

4. Discussion
4.1. Comparison of results obtained using conventional FCM and FCM-LT
From the results displayed in Fig. 2 to Fig. 5, it can be observed that FCM-LT is capable
of producing thresholds which lead to more accurate extractions as compared to those
produced by conventional FCM. From our experiments, it was also observed that
FCM-LT can produce better results in low-contrast images. An example is in Fig. 6. The
threshold obtained by FCM is 108. Using FCM-LT, the computed threshold is 103. We
note that in the right half of the image, FCM-LT is able to extract the object fully unlike
FCM.
772 H. P. Ng et al.
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

(a) (b)

(c) (d)

Fig. 4. (a) Mammogram image (b) Histogram (c) Result using FCM (d) Result using FCM-LT.

Through our experiments, it was observed in majority of them that there are
improvements in the results. In the remaining few cases, the thresholds obtained by FCM
and FCM-LT were similar. An example is in Fig. 7, where thresholds of 160 and 163
were produced by conventional FCM and FCM-LT respectively. Here we would like to
stress that though the FCM-LT does not always yield better results than the conventional
FCM algorithm, the former will not produce a result which is inferior to that produced by
the latter.
Fuzzy C-Means Algorithm with Local Thresholding for Gray-Scale Images 773
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

(a) (b)

(c) (d)

Fig. 5. (a) Computed tomography image (b) Histogram (c) Result using FCM (d) Result using FCM-LT.

(a) (b) (c)


Fig. 6. (a) Low-contrast image (b) Result using FCM (c) Result using FCM-LT.
774 H. P. Ng et al.
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

(a) (b) (c)

Fig. 7. (a) Thumbprint image (b) Result using FCM (c) Result using FCM-LT.
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

4.2. On the choice of Otsu thresholding method


Many thresholding techniques have been proposed earlier. We decided on the use of Otsu
thresholding method due to its simplicity, efficiency and satisfactory performance.10, 11, 12
Another popular thresholding method which we considered is the minimum-error
thresholding method13 which suggests to compute the threshold that optimize the average
pixel classification error assuming that the difference image histogram derives from a
mixture of two normal components. However, this method is susceptible to model bias
and may produce thresholds which are largely inaccurate under such circumstances. Otsu
method, on the other hand, is proven to produce more stable and reliable results

4.3. Computation load of proposed method


The conventional FCM algorithm and Otsu thresholding are incorporated in our proposed
method. They are histogram-based methods and hence have low computation cost as
compared to context-based methods which consider the spatial relationship among voxels
besides the intensity. Furthermore, through our conducted experiments, the additional
amount of time which FCM-LT took as compared to the conventional FCM algorithm is
insignificant.

5. Conclusions
We have designed an improved FCM algorithm (FCM-LT) that incorporates Otsu
thresholding with the conventional FCM algorithm. By performing Otsu thresholding in
the region of the cluster boundary which has been produced by FCM, the proposed
algorithm reduces the susceptibily of FCM to local minima and also its tendency to
derive a threshold that is biased towards the component with larger probability.
Experiments indicate that the proposed method is capable of producing better results
compared to the conventional FCM algorithm.
Fuzzy C-Means Algorithm with Local Thresholding for Gray-Scale Images 775

Acknowledgments
The first author will like to thank Agency for Science, Technology and Research
(A*Star), Singapore for funding his PhD studies. The authors thank the anonymous
reviewers for their insightful comments.

References
1. J.C. Bezdek, “Pattern recognition with fuzzy objective function algorithm” (Plenum, NY,
1981).
2. W.L. Hung, M.S. Yang, and D.H. Chen, “Parameter selection for suppressed fuzzy c-means
with an application to MRI segmentation”, Pattern Recognition Letters, Vol. 27, No. 5,
Int. J. Artif. Intell. Tools 2008.17:765-775. Downloaded from www.worldscientific.com

pp. 424-438, 2006.


3. W. Chen, M.L. Giger, and U. Bick, “A fuzzy c-means (FCM)-based approach for
by NORTHEASTERN UNIVERSITY on 02/09/15. For personal use only.

computerized segmentation of breast lesions in dynamic contrast-enhanced MR images”,


Academic Radiology, 2006, 13, (1), pp. 63-72.
4. M.E. Brandt, T.P. Bohan, L.A. Kramer and J.M. Fletcher, “Estimation of CSF, white and grey
matter volumes in hydrocephatic children using fuzzy clustering of MR images”,
Computerized Medical Imaging and Graphics, Vol. 18, pp. 25-34, 1994.
5. C. Zhu and T. Jiang, “Multicontext fuzzy clustering for separation of brain tissues in magnetic
resonance images”, Neuroimage, Vol. 18, No. 3, pp. 685-696, 2003.
6. D.L. Pham, “Spatial models for fuzzy clustering”, Computer Vision and Image
Understanding, Vol. 84, No. 2, pp. 285-297, 2001.
7. Z.J. Hou, W. Qian, S. Huang, Q. Hu and W.L. Nowinski, “Regularized fuzzy c-means method
for brain tissue clustering”, Pattern Recognition Letters, Vol. 28, No. 13, pp. 1788-1794,
2007.
8. Z.J. Hou, Q. Hu and W.L. Nowinski, “On minimum variance thresholding”, Pattern
Recognition Letters, 2006, 27, (14), pp. 1732-1743.
9. N. Otsu, “A threshold selection method from gray-level histograms”, IEEE Trans. Sys. Man
Cybern., 1979, 9, pp. 62-66.
10. P.K. Sahoo, S. Soltani and A.K.C. Wong, “A survey of thresholding techniques”, Computer
Vision, Graphics, and Image Processing, Vol. 41, pp. 233-260, 1988.
11. S.U. Lee and S.Y. Chung, “A comparative performance study of several global thresholding
techniques for segmentation”, Computer Vision, Graphics, and Image Processing, Vol. 52,
pp. 171-190, 1990.
12. M. Sezgin and B. Sankur, “Survey over image thresholding techniques and quantitative
performance evaluation”, Journal of Electronic Imaging, Vol. 13, No. 1, pp. 146-165. 2004.
13. J. Kittler and J. Illingworth, “Minimum error thresholding”, Pattern Recognition, Vol. 19,
pp. 41-47, 1986.

You might also like