A Novel Image Segmentation Algorithm Based On Neutrosophicsimilarity Clustering
A Novel Image Segmentation Algorithm Based On Neutrosophicsimilarity Clustering
a r t i c l e i n f o a b s t r a c t
Article history: Segmentation is an important research area in image processing, which has been used to extract objects
Received 17 June 2014 in images. A variety of algorithms have been proposed in this area. However, these methods perform well
Received in revised form 22 August 2014 on the images without noise, and their results on the noisy images are not good. Neutrosophic set (NS) is
Accepted 29 August 2014
a general formal framework to study the neutralities origin, nature, and scope. It has an inherent ability
Available online 8 September 2014
to handle the indeterminant information. Noise is one kind of indeterminant information on images.
Therefore, NS has been successfully applied into image processing algorithms. This paper proposed a novel
Keywords:
algorithm based on neutrosophic similarity clustering (NSC) to segment gray level images. We utilize the
Image segmentation
Clustering analysis
neutrosophic set in image processing eld and dene a new similarity function for clustering. At rst,
Neutrosophic set an image is represented in the neutrosophic set domain via three membership sets: T, I and F. Then, a
Similarity function neutrosophic similarity function (NSF) is dened and employed in the objective function of the clustering
analysis. Finally, the new dened clustering algorithm classies the pixels on the image into different
groups. Experiments have been conducted on a variety of articial and real images. Several measurements
are used to evaluate the proposed methods performance. The experimental results demonstrate that the
NSC method segment the images effectively and accurately. It can process both images without noise and
noisy images having different levels of noises well. It will be helpful to applications in image processing
and computer vision.
2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2014.08.066
1568-4946/ 2014 Elsevier B.V. All rights reserved.
392 Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398
-mean operator was replaced with -median operator to alle- function of the clustering analysis. Finally, the new dened clus-
viate the blurring effect of the mean operator. A fully automatic tering algorithm segments the pixels on the image into different
NS and wavelet transform based color texture image segmenta- groups. Experiments have been conducted on a variety of arti-
tion approach was proposed in Ref. [9]. The wavelet transform was cial and real images. Several measurements are used to evaluate
applied to each channel of the color image and vertical and hori- and compare the proposed methods performance. The experimen-
zontal details were extracted for subsequent processes. Then, the tal results demonstrate that the NSC method segment the images
NS based segmentation approach was applied to the detail images. effectively and accurately. It is able to process both images without
Energy features were calculated and concatenated for forming the noise and noisy images having different levels of noises well.
feature matrix. Finally, -K-means clustering algorithm was used The paper is organized as follows. Section 2 describes the
for segmentation. Zhang et al. [10] used NS and watershed method proposed method which contains clustering analysis and image
for image segmentation. A T membership of the NS was dened segmentation method. Section 3 discusses the experimental results
via S-function. The input image was transformed to neutrosophic and comparisons, and the conclusions are drawn in Section 4.
domain and neutrosophic logic was applied for obtaining a binary
image. Then the watershed technique was applied to obtain the
segmentation results. An unsupervised method, which synthesized 2. Proposed method
the NS and mean-shift, was proposed in Ref. [11]. The proposed
algorithm adopted the mean-shift clustering in NS domain to seg- 2.1. Clustering analysis
ment images, which makes it possible to detect constructions with
a consistent threshold. Ling et al. [12] proposed an unsupervised Clustering can classify similar samples into the same group [15].
color image segmentation algorithms based on NS. The centers Let X = {Xi , i = 1, 2, . . ., n} be a data set, and xi be a sample. The
of image clusters were determined by using color information in goal of clustering is to nd a partition C = {C1 , C2 , . . ., Cm }, which
m
RGB color space. The neutrosophic indeterminacy was dened by satises: X = Ci , Ci =
/ for i = 1, 2, . . ., m, Ci Cj = for i, j = 1, 2,
using spatial information in CIE (L*u*v*) color space. The color and i=1
spatial information were integrated by the neutrosophy approach. . . ., m ; i =
/ j.
Recently, Guo and Sengur [13] proposed ltering in NS domain Among clustering methods, the K-means algorithm is widely
and level set theory for image segmentation. A newly dened lter used and more efcient [15]. It is important to dene the object
was employed to reduce the indeterminacy of the image in the NS function for a clustering analysis method. Each cluster should be as
domain and the level set algorithm was used to extract the objects compact as possible. The objective function of K-means is dened
boundaries automatically. as:
Based on the reviewed literature, it is evident that NS is an
nj
open area for further image segmentation applications. Therefore,
m
Fig. 2. The relation between SNR and ME. *: NSC method, O: NS method.
Fig. 3. The relation between SNR and FOM. *: NSC method, O: NS method.
where Zj is the center of the jth cluster, m is the total number of An image is dened in the NS as: let U be a universe, BP be
clusters and nj is the number of pixels in the jth cluster. a bright pixel set in U, and an image Im described using NS is
called neutrosophic image INS . A neutrosophic image INS is depicted
2.2. Neutrosophic similarity function using subsets T, I and F. According to the denition of neutrosophic
image, a pixel P(x, y) is interpreted in the neutrosophic set domain:
A neutrosophic set can be dened as: let A = {A1 , A2 , . . ., Am } be PNS (x, y) = {T(x, y), I(x, y), F(x, y). T(x, y), I(x, y) and F(x, y) represent
a set of alternatives in neutrosophic set, and C = {C1 , C2 , . . ., Cn } be a memberships belonging to bright pixel set, indeterminate set and
set of conditions. The alternative Ai at the Cj condition is denoted as non-bright pixel set, respectively. Using the intensity criterion, they
{TCj (Ai ), ICj (Ai ), FCj (Ai )}/Ai , where TCj (Ai ), ICj (Ai ) and FCj (Ai ) are the are dened as:
membership values to the true, indeterminace and false set at the
g(x, y) gmin
Cj condition. TCg (x, y) = (5)
A similarity function is proposed to evaluate the similarity gmax gmin
degree between two elements in neutrosophic set [16]:
TCj (Am )TCj (An ) + ICj (Am )ICj (An ) + FCj (Am )FCj (An )
Gd(i, j) Gdmin
ICg (x, y) = 1 (6)
SCj (Am , An ) = (3) Gdmax Gdmin
TC2 (Am ) + IC2 (Am ) + FC2 (Am ) TC2 (An ) + IC2 (An ) + FC2 (An )
j j j j j j
The ideal element in NS can identify the best alternative. The FCg (x, y) = 1 TCg (x, y) (7)
ideal alternative A* is denoted as: {TC (Ai ), IC (Ai ), FC (Ai )}/Ai . The
j j j
where g(x, y) and Gd(x, y) are the intensity
similarity to the ideal alternative is computed as: value and gradient value at the position of (x, y) on the image.
TCj (Ai )TCj (A ) + ICj (Ai )ICj (A ) + FCj (Ai )FCj (A ) Then, a similarity value is calculated to identify the degree to the
SCj (Ai , A ) = (4) ideal object under intensity condition. gmin and gmax are minimum
TC2 (Ai ) + IC2 (Ai ) + FC2 (Ai ) TC2 (A ) + IC2 (A ) + FC2 (A )
j j j j j j and maximum intensity values, respectively. Gdmin and Gdmax are
394 Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398
minimum and maximum of the gradient value in the image. FCm (x, y) = 1 TCm (x, y) (12)
TCg (x, y)TCg (A ) + ICg (x, y)ICg (A ) + FCg (x, y)FCg (A ) where gm (x, y) and Gdm (x, y) are the intensity value and gradient
SCg (P(x, y), A ) = magnitude value at the position of (x, y) on the image after mean
TC2 (x, y) + IC2 (x, y) + FC2 (x, y) TC2 (A ) + IC2 (A ) + FC2 (A )
g g g g g g
(8) lter processing. gm min and gm max are the minimum and maxi-
mum of the intensity on the image after mean lter processing, and
Gdm min and Gdm max are the minimum and maximum of the gra-
To make the segmentation results robust to noise, we inte- dient magnitude value on the image after mean lter processing,
grate two new conditions, local mean intensity criterion Cm and respectively.
local homogeneity criterion Ch , into the neutrsophic similarity The neutrosophic set under the local homogeneity condition Ch
function. is also dened as:
The neutrosophic set using the local mean intensity Cm is dened H(x, y) Hmin
TCh (x, y) = (13)
as: Hmax Hmin
gm (x, y) gm min Gdh (i, j) Gdh min
TCm (x, y) = (9) ICh (x, y) = 1 (14)
gm max gm min Gdh max Gdh min
x+w/2
y+w/2 FCh (x, y) = 1 TCh (x, y) (15)
1
gm (x, y) = g(m, n) (10)
ww H(x, y) = TEM(g(x, y)) (16)
m=xw/2n=yw/2
where H(x, y) is the homogeneity value at (x, y), which is depicted as
Gdm (i, j) Gdm min the ltering result with the texture energy measures (TEM) lters
ICm (x, y) = 1 (11)
Gdm max Gdm min [14]. Gdh (x, y) is the gradient value on H(x, y).
Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398 395
2.3. Image segmentation based on neutrosophic similarity Step 1: Convert the image into NS domain;
clustering Step 2: Compute the NSF values under three conditions;
Step 3: Calculate the object function using the NSF;
A novel clustering method, neutrosophic similarity clustering Step 4: Select the partition with minimized the object function
(NSC) is proposed to segment the pixels in image by using an opti- value;
mum partition with minimizing the object function. Step 5: Segment the image using the optimum partition result.
The new objective function is dened as:
3. Experimental results and discussions
K
H W
JNKC = ||NSF(x, y) Zl ||2 (20) We have tested the proposed algorithm using different images,
l=1 x=1 y=1 and compared its performance with those of newly developed
396 Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398
algorithms. In the experiments, we compare the NSC method with foreground set, and object points wrongly grouped to background
a published method based on NS [5]. set [17,18]:
|BO BT | + |FO FT |
3.1. Performance on articial images ME = 1 (22)
|BO | + |FO |
We use articial images to compare the NSC method and NS where Fo and Bo are the object and background sets on the ground
method visually, and then evaluate their results quantitatively. truth image. FT and BT are the object and background sets on the
In Ref. [5], several articial images having two intensities and result image. | | denotes the elements number in the set. The ME
impaired by different levels of noise were employed to evaluation denes a metric of the mis-grouped points between the ground
the performance of the NS method. truth image and the test image.
To identify the performance difference of the NS and NSC meth- Another metric gure of merit (FOM) proposed by Pratt [19]
ods, we employ an articial image having two gray levels (64 and is also utilized to evaluate the difference between the methods
128) impaired by different levels of Gaussian noise to test the results with the ideal segmentation result quantitatively:
performance of two segmentation methods. Fig. 1(a) shows the
synthetic images having Gaussian noise, whose mean values are 0 1 NA
1
FOM = (23)
and standard variance values are 15 and 30, respectively. Fig. 1(b) max(NI , NA ) 1 + d2 (k)
and (c) are the results by the NS and NSC methods, respectively. k=1
From the results in Fig. 1, the NSC method achieved the better where NI and NA are the numbers of the object points and the ideal
performances than those of the NS algorithm. object pixels, respectively. d(k) is the distance from the kth actual
The results on the second synthetic image with low contrast point to the nearest segmented result point. is a constant, and
show the NSC method performs better than NS method on the set as 1/9 in Pratts paper [19]. The greater the FOM, the better the
synthetic images with two different noise levels. A lot of pixels in segmentation results are.
Fig. 1(b) are identied in wrong classes, while they are classied The quality of the noisy image is measured using the signal to
correctly by NSC method in Fig. 1(c). noise ratio (SNR):
To evaluate the segmentation results for articial images quan- H1 W 1
titatively, we utilize a metric: misclassication error measure Im(r, c)
r=0 c=0
(ME) [17,18], to measure the segmentation performances. The ME SNR = 10 log H1 W 1
(24)
(Im(r, c) In(r, c))2
depicts the percentage of background points wrongly grouped to r=0 c=0
Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398 397
where In(r, c)and Im(r, c) are the intensities of pixel (r, c) in the method. The results of the NSC method eliminate noise effect and
noisy and original images, respectively. most pixels are segmented into the right groups, while the NS meth-
The values of ME are plotted at different SNR levels in Fig. 2. It ods are affected severely by the noise on the real images.
demonstrates the NSC method archives lower MEs at all SNRs. All Fig. 4 is a scene image of a lake. In the original image, both NS
ME values of the NSC method are smaller than 0.065, and all values and NSC methods obtain similar segmentation results. The ground
of NS method are higher than those of the NSC method. The NSC and the lake surface are correctly segmented. For the second and
algorithm achieves the optimum performance with ME = 0.0057 third rows, the image is corrupted by noises and the NS method
which is nearly equal to zero when SNR is 20.67 dB, while NS obtains yields many misclassied pixels in the ground and water surface.
the optimum value with ME = 0.016. Meanwhile, the values of FOM There are also several numbers of misclassied pixels on the sky
of NS are bigger than those of NS at most SNR levels. The average region. The superiority of the proposed NSC method can be seen
value of FOM of NSC and NS are 0.9724 and 0.7698, respectively. in the result of the Peppers image which is depicted in Fig. 5.
The values of FOM are plotted at different SNR levels in Fig. 3. The surface of peppers and the boundaries between peppers are
segmented correctly by NS method, while the NSC method obtains
3.2. Performance on real world images better results. In Fig. 5, it is obvious that our proposal obtains better
results for all noise variance levels than the NS algorithm. Similar
A great number of real world images are employed to measure results can be seen in the Woman image in Fig. 6. The face, hair,
the NSC methods performance. Eight representative images are background regions are segmented correctly by the NSC method
selected to demonstrate the effectiveness and robustness of the for the rst and second noise levels. In addition, the proposed NSC
NSC method against the various noise levels. These images are method can obtain reasonable segmentations even for the third
Lake, Peppers, Woman, and Lena, respectively. In Figs. 47, noise level. For the second and third noise levels the NS based
the rst row lists the original images without noise and the results method yields many misclassications in the face and background
of NS and NSC method on them. The second and third rows demon- of the Woman image. The segmentation results of the Lena
strate the results on the noisy images with different Gaussian noise image are given in the second and the third columns of the Fig. 7. It
variances. The experimental results demonstrate that the results is obvious that the results obtained under the rst and second noise
by NSC method exhibits visually better quality than those of NS variance are better than the NS method. The shoulder, face and the
398 Y. Guo, A. Sengr / Applied Soft Computing 25 (2014) 391398
hair regions are segmented correctly. The NS based method obtains [2] R.C. Gonzalez, R.E. Woods, Digital Image Processing, 2nd ed., Prentice Hall, 2002.
several wrong segmentations in the hair and the face regions. Both [3] H.D. Cheng, X.H. Jiang, Y. Sun, J.J. Wang, Color image segmentation: advance
and prospects, Pattern Recognit. 34 (2001) 22592281.
NSC and NS have several wrong segmentations for the third noise [4] S.K. Pal, A. Rosenfeld, Image enhancement and thresholding by optimization of
level. However, by visual inspection, the NSC method has better fuzzy compactness, Pattern Recognit. Lett. 7 (1988) 7786.
segmentation than the NS method when we consider the overall [5] Y. Guo, H.D. Cheng, A new neutrosophic approach to image segmentation,
Pattern Recognit. 42 (2009) 587595.
results. [6] F. Smarandache, A Unifying Field in Logics Neutrosophic Logic. Neutrosophy,
From the experiments on the articial and real world images, we Neutrosophic Set, Neutrosophic Probability, 3rd ed., American Research Press,
can draw a conclusion that the NSC method is feasible to segment 2003.
[7] H. Wang, R. Sunderraman, F. Smarandache, Y.Q. Zhang, Interval Neutro-
both the clear images without noise and the noisy images with
sophic Sets and Logic: Theory and Applications in Computing: Innite Study,
different noise levels. 2005.
[8] E. Karabatak, Y. Guo, A. Sengur, Modied neutrosophic approach to color image
segmentation, J. Electron. Imaging 22 (1) (2013) (art. no. 13005).
4. Conclusions
[9] A. Sengur, Y. Guo, Color texture image segmentation based on neutrosophic
set and wavelet transformation, Comput. Vis. Image Underst. 115 (8) (2011)
This paper presents a new image segmentation algorithm using 11341144.
[10] M. Zhang, L. Zhang, H.D. Cheng, A neutrosophic approach to image segmenta-
neutrosophic similarity clustering algorithm. The image is depicted
tion based on watershed method, Signal Process. 90 (5) (2010) 15101517.
in neutrosophic set. A neutrosophic similarity function is dened [11] B. Yu, Z. Niu, L. Wang, Mean shift based clustering of neutrosophic domain for
to measure the membership to the object pixels on the image, and unsupervised constructions detection, Optik 124 (21) (2013) 46974706.
used in the clustering algorithm. Finally, an optimized clustering [12] L. Zhang, M. Zhang, H.-D. Cheng, Color image segmentation based on neutros-
ophy, Optical Eng. 51 (3) (2012).
partition is obtained using the NSC algorithm. The experimental [13] Y. Guo, A. Sengr, A novel image segmentation algorithm based on neutro-
results show that the NSC method can segment the images properly sophic ltering and level set, Neutrosophic Sets Syst 46 (2013).
and effectively. It is able to process both images without noise and [14] K. Laws, Textured Image Segmentation (Ph.D. Dissertation), University of
Southern California, January 1980.
noisy images with different levels of noises. This advantage will [15] J.A. Hartigan, Clustering Algorithms, John Wiley & Sons, Inc., 1975.
bring the proposed method into more applications in the research [16] J. Ye, Multicriteria decision-making method using the correlation coefcient
areas such as image processing and computer vision. under single-valued neutrosophic environment, Int. J. Gen. Syst. 42 (4) (2013)
386394.
[17] S.T. Wang, F.L. Chung, F.S. Xiong, A novel image thresholding method based on
References Parzen window estimate, Pattern Recognit. 41 (1) (2008) 117129.
[18] W.A. Yasnoff, J.K. Mui, J.W. Bacus, Error measures for scene segmentation, Pat-
[1] N.R. Pal, S.K. Pal, A review on image segmentation techniques, Pattern Recognit. tern Recognit. 9 (4) (1977) 217231.
26 (9) (1993) 12771294. [19] W.K. Pratt, Digital Image Processing, John Wiley & Sons, 1978, pp. 429432.