Professional Documents
Culture Documents
An Image SegmentationA lgorithm SegmentationA Based on F uzzy C -Means Clustering uzzyC -Means
Xin-bo ZHANG College of Information and Electronic Engineering,ZheJiang Gongshang University Hangzhou,zhejiang,china Li JIANG College of Information and Electronic Engineering ,ZheJiang Gongshang University Hangzhou,zhejiang,china jl321123@163.com
lots of work have been done, there is still no a general segmentation algorithm to and a impersonality criterion estimate the
Abstract
Image segmentation algorithm based on fuzzy cmeans clustering is an important algorithm in the image segmentation field. It has been used widely. However, it is not successfully to segment the noise image because the algorithm disregards of special constraint information. It only considers the gray information. Therefore , we proposed a weighed fcm algorithm based on gaussian kernel function for image segmentation . The original euclidean distance is replaced by a kernel-induced distance in the algorithm . Then, a bound term is added to the objective function to compensate the influence of the spatial information. The experimental results illustrate that the proposed method is more effective to image segmentation.
segmentation. According to concretely instances, many segmentation algorithms have been put forward. Among them, image segmentation algorithm field[10] [12]. based It on fuzzy c-means could retain much is an more important algorithm in the image segmentation information from the original image than hard segmentation methods . Fuzzy c-means clustering algorithm is a representative clustering algorithm . However, FCM algorithm can't realiz e feature optimiz ation of the sample dates. Meanwhile, with the original FCM algorithm, we can't get satisfactory results because of the influence of the noise. Document[2][4][7][9] put forward a FCM algorithm based on kernel density function. The original euclidean distance is replaced by a kernel-induced distance in the algorithm . This new algorithm can effectively analyz e many data structures such as non-hyperspherical structure, mix ed structure composed of heterogeneous cluster prototypes , and data with noise and so on . In this paper, using kernel function instead of euclidean distance, considering spatial information, initializ ing weight and increasing constraint item, we proposed a weighed fcm algorithm based on gaussian kernel function for image segmentation . The proposed method can
22
1. Introduction
Image segmentation algorithm is a crucial step
separate an image into some regions of feature and to pick up the interesting objects.
978-0-7695-3565-4/09 $25.00 2009 IEEE DOI 10.1109/ICDIP.2009.15
fully consider the influence of the noise. By it the target can be ex tracted from the background effectively in noisy images.
In the literature [4], the author brought forward a fuzzy c-means clustering algorithm based on kernel density function. The Euclidean distance in the standar d FCM algorithm was replaced by kernel model. The objective function [3][5][9]was:
C N 2 m Jm = u ij ( x j ) ( v i ) i=1 j =1
Where
(3)
gray. C was the number of classification. was the menbership of cluster . center. degree
i
ij
th object in the
i th
parameter estimation. However, experiences and theories show that this presumption disagrees with actual physial models. And these methods can't always achieve satisfactory results. Due to the above defects, people brought forward the kernel density estimation, one kind of the nonparameter estimations. We can map the datas from highdimensional space to low-dimensional space by kernel density estimation. That realiz ed feature optimiz ation. That is to say, a set of sample date as (
{v , i = 1,2,..., C}
m
of degree. classification
was clustering
menbership
to
the
and
v were:
i
2
ij
( x j ) (v i )
c
1 m
k=1
( x j ) (vk )
m ij j
1 m
x , i = 1,2,..., N
i
could
be
ex pressed
(u
j =1 N
x ).
i
) x (u )
m ij j =1
(4)
= 1,then:
k ( xi , x j ) = ( ( xi ), ( x j ))
is G aussian kernel [11]. radial basis function
(1)
(x)(v) =k(x,x)+k(v,v) j i j j i i
2 (x,v)=2(1k(x,v)) k j i j i
C N m J m = 2 uij (1 k( xj ,vi)) i =1 j =1
(5) And,obtained by substituting (5) into (3)and(4):
k ( xi , x j ) = exp(
xi x j
2
2
(2)
(6)
(x , v )] u = [1 k (x , v )]
[1 k
j i ij c j k k =1
1 1 m 1 1 m
(7)
23
by using the original image and its smoothed image, the target and background can be more easily distinguished than in a 1D histogram. So the implement method of initializing weight
( ) ( ) )
1 1m
m 1m
f ( m, n ) , we get the
Then
g ( m, n) .
( s , t) ,
is a point
initializing a binary
Substituting
k:
(8)
N ' J m = 2 [1 k j = 1 k =1
(x , v )]
j k
1 1 m
1m
In this case, the objective function relates to the gaussian kernel function merely. It not only optimiz e the clustering model, but optimiz e the validity function judging clustering effect.
( s , t ) is
L 1
=0
st
s t nst = 1 .Then
=0
L 1
density function is
st
st
M N
, in another
st
( s, s) .
f (m, n)
, m = 1,..., M ; .
n = 1,..., N .
. Then the
h = M N , i = 0,1,..., L 1
i
n (i )
and
h = 1 which
i =0 i
L 1
n (i )
(9)
represents the number of occurrence of the pix el. i is the gray of the pix el [6]. However, on a 1D gray histogram, the distribution of the target and background is overlapped due to the relativity between a pix el and its neighborhood pix els. So with the method of histogram we can't get satisfactory result. Therefore, image's spatial information can be fully used. For ex ample, building a 2D histogram
multiplier
ij
and
v are:
i
u=
ij
[1k xj , vi ]1m
[1 k(x , v )]
j k k=1
1 1m
24
m j (u ij ) x j j
=1 N
(10)
j (u ij ) j
=1
Figure 1(b). the image divided by FCM Those from figure 1(b) to figure 1(d) are the images after division. it can Contrasting be seen that the the the the FCM modified Figure 2(a). The primitiv e image ex perimental results,
background after diving includes massive noises and the target is not clear although algorithm background. algorithm neighbourhood performance. can separate However, with the goal from the
which considers the influence of pix els we get good division Figure 2(b). the image after division by FCM algorithm
25
IEEE
Transactions on Fuz z y System,May 1995, vol. 3, no. 2, pp.241-245. [6] Refael C.Gonz alez and Richard E.Woods, Digital
6. Conclusion
Image segmentation is an important aspect in image processing. The division quality directly effect the high-level image analysis and understanding. That,using FCM algorithm in
Image
proccessing(second edition),
Pearson
Education Asia LimitedAnd Publishing House of Electronics Industyt, 2007.8. [7] M. Girolami, Mercer kernel-based clustering in feature space , IEEE Trans. Neual Networks, May 2002, vol. 13, pp. 780-784. [8] Z hang DQ, Chen SC. Fuz z y clustering using kernel methods. In: Proceedings on of the
image segmentation, can reduce the artificial intervention effectively. It suits the images with uncertainty and fuzziness. Since the tradition FCM algorithm failed to consider the influence of neighbourhood pix els,its anti-noise ability is pretty bad. This paper proposed a weighed fcm algorithm based on gaussian kernel function for image segmentation . This method introduced the influence of neighbourhood pix els and kernel function, initializ ed weight by 2D histogram and perfected the division model.The ex perimental results indicated that the proposed method is quite effective. It has strong robustness to the noise and certain actual meaning.
International
Conference
Controland
Automation, Xiamen,China, June, 2002. [9] Songcan Chen and Daoqiang Z hang, Robust
image segmentation using FCM with spatial constraints based on new kernel-induced distance measure, Systems, Man and Cybernetics, Part B, IEEE Transactions on V olume 34, Issue 4, Aug. 2004 Page(s): 1907 - 1916. [10] Bezdek JC, Hall LO and Clarke LP, Review of MR image segmentation techniques using pattern recognition, MedPhys ,1993, 20:103348. [11]Muller KR and Mika Setal, An introduction to kernel-based learning algorithms, IEEE Trans
7. References
[1] Bezdek J C, Pattern Recognition with Fuzzy
[12] Pham DL and Prince JL, An adaptive fuz zyCmeans algorithm for image segmentation in the presence of intensity inhomogeneities, Pattern Recognit Lett , 1999, 20(1):5768.
application in medical image segmentation, Artificial PP. 3750. [3] Xin-bo GAO, Fuzzy Cluster Analysis and its
26
Intelligence
in
Medicine,
2004,