See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/317394791

Image Segmentation by Combining the Global

and Local Properties

Article in Expert Systems with Applications June 2017

DOI: 10.1016/j.eswa.2017.06.008

CITATIONS READS

0 17

1 author:

Zhenzhou Wang
Chinese Academy of Sciences
47 PUBLICATIONS 180 CITATIONS

SEE PROFILE

All content following this page was uploaded by Zhenzhou Wang on 08 June 2017.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
Accepted Manuscript

Image Segmentation by Combining the Global and Local Properties

Zhenzhou Wang

PII: S0957-4174(17)30416-5
DOI: 10.1016/j.eswa.2017.06.008
Reference: ESWA 11378

To appear in: Expert Systems With Applications

Received date: 8 March 2017

Revised date: 5 June 2017
Accepted date: 6 June 2017

Please cite this article as: Zhenzhou Wang , Image Segmentation by Combining the Global and Local
Properties, Expert Systems With Applications (2017), doi: 10.1016/j.eswa.2017.06.008

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT
1

Highlights
We calculate the means of different pixel classes from slope difference distribution.
The image is segmented by clustering the pixels to its nearest mean.
The segmentation is further refined by minimizing the energy of its Gibbs distribution.
The boundary of the refined segmentation is deformed to the nearest detected canny edges.

T
IP
CR
US
AN
M
ED
PT
CE
AC
 1
ACCEPTED MANUSCRIPT

Expert Systems With Applications

journal homepage: www.elsevier.com

Image Segmentation by Combining the Global and Local Properties

Zhenzhou Wang
State key Lab of Robotics, Shenyang Institute of Automation, Chinese academy of science (CAS), Shenyang, China

T
IP
ABSTRACT

Image segmentation plays a fundamental role in many computer vision applications. It is challenging because of

CR
the vast variety of images involved and the diverse segmentation requirements in different applications. As a
result, it remains an open problem after so many years of study by researchers all over the world. In this paper, we
propose to segment the image by combing its global and local properties. The global properties of the image are
characterized by the mean values of different pixel classes and the continuous boundary of the object or region.

US
The local properties are characterized by the interactions of neighboring pixels and the image edge. The proposed
approach consists of four basic parts corresponding to the global or local property of the image respectively: (1),
the slope difference distribution that is used to compute the global mean values of different pixel classes; (2),
Energy minimization to remove inhomogeneity based on Gibbs distribution that complies with local interactions of
neighboring pixels; (3), the Canny operator that is used to detect the local edge of the object or the region; (4), the
AN
polynomial spline that is used to smooth the boundary of the object or the region. These four basic parts are
applied one by one and each of them is indispensable for the achieved high accuracy. A large variety of images are
used to validate the proposed approach and the results are favorable.
Keywords:
Segmentation; Histogram; Gibbs distribution; Slope difference distribution
M

2017 Elsevier Ltd. All rights reserved.

ED
PT
CE
AC

Corresponding author. Tel.: +86-24-2397-0612; Fax. +86-024-23970021; e-mail: wangzhenzhou@sia.cn


ACCEPTED MANUSCRIPT
1. Introduction
Image segmentation algorithms are usually only effective for
some specific types of images while they might make great
mistakes for other types of images due to four major problems.
(1), the histogram distributions of different pixel classes overlap
with each other, which makes the estimation of the global
parameters inaccurate. (2), the noise that modifies the histogram
distribution might cause great errors during global parameter
estimation. (3), the local properties of different regions in the
image are usually not homogeneous at the same spatial scale,
which makes it difficult for modeling-based segmentation
methods to estimate the local parameters with satisfactory
accuracy. (4), the overlap of different grayscale distributions
makes the boundaries of the labeled regions inaccurate.
The first two problems persist in the famous expectation
maximization (EM) algorithm (Dempster et al., 1977; Besag,
1986) that has been widely used to estimate the means and
variances of different pixel classes during image segmentation
although its effectiveness has been widely accepted (Carson et
al., 2002; Won and Gray, 2004). The other popular parameter
estimation algorithm, k-means clustering method (Jain, 2010)
could not effectively deal with the first problem and the second
problem either. The estimated global parameters by EM or K-
means are frequently inaccurate in image segmentation
applications. For the popular thresholding methods (Osuna-
Encisa et al., 2013), their accuracies are also affected by the first
two problems. Some researchers have tried to solve the second
problem by estimating the noise separately and apply the de-
noising technique to suppress the estimated noise (Sagar et al.,
2014, Manjon et al., 2015, Chan et al., 2006), which achieved
US
AN
significant improvement in segmentation accuracy. However, it
is also very difficult to estimate the parameters of the noise
model accurately. The overlap of the gray-scale distributions of
different pixel classes, the different levels of noise magnitudes
and the different forms of noise distributions make it difficult to
M
estimate the parameters with adequate accuracy in most cases.
To address the third problem, the image had been assumed as
a Markov Random Field (MRF) that is subject to the Gibbs
distribution (Derin and Elliott, 1987; Caillol et al., 1993; Wang
ED
and Zhang, 2012). Based on this assumption, a lot of efforts have
been put into the segmentation research work and many MRF-
based segmentation methods have been proposed. For instance,
some researchers proposed to cluster the texture features with a
PT
mixture of Gaussian models for better performance (Yang et al.,
2008). Some researchers combined the Maximum a posteriori
(MAP), maximum likelihood (ML) and graph cut to achieve the
satisfactory segmentation results (Chen et al., 2010). One merit
CE
of the MRF-based method is that it could distinguish different
textured regions, especially for the textures that could be
modeled and defined with parameters. The major drawback of
MRF-based methods is that they could not estimate the
AC
parameters of the textures with adequate accuracy in most cases,
especially for the natural images. On the other hand, its
segmentation accuracy is determined by the accuracy of the
parameter estimation. Consequently, the accuracy of MRF-based
methods is only good for a specific class of images whose texture
parameters could be estimated robustly. Although the parameter
of the Gibbs distribution might not be estimated accurately in
most cases, it characterizes the interaction of neighboring pixels
well and thus could be utilized in the segmentation process.
Instead of estimating its parameters, we use the iterative gradient
descent method to minimize the energy of the Gibbs distribution
until it converges to a steady state.
For the fourth problem, we did not find any literature that is
specially devoted to addressing it though we have found the
segmentation artifacts caused by it in some popular methods
(Carson et al., 2002; Felzenszwalb and Huttenlocher, 2004). The
methods based on variational formulations (Chan and Vese,
2001; Li et al., 2010; Lankton and Tannenbaum, 2008; Arbelaez
et al., 2011) are not affected by the fourth problem. However,
2
none of these methods is good at segmenting a variety of images.
They usually perform well in a class of images while performs
badly for another class of images. Based on all these facts, it is
necessary to explore new ways or directions for image
segmentation to achieve better overall performance.
In this paper, we propose a new image segmentation approach
that could deal with these four problems effectively. (1), to
reduce the interactions between different distributions and to
estimate the global parameters of the image robustly, we estimate
the means of different pixel classes by the global peaks of the
histogram distribution. These peaks are isolated from each other
and are not affected by the overlap of the grayscale distributions
of different pixel classes. (2) to reduce the noise, we smooth the
histogram distribution with a Fourier transformation based filter
that is proven to be more effective than other filters. (3), to make
use of the local property of the image, we propose an iterative
gradient descent method to minimize the global Gibbs energy
T
until it converges. (4), to improve the accuracy of the extracted
boundaries from the labeled regions, we deform them to the
IP
nearest detected Canny edges and then use the polynomial
interpolation to smooth them.
To detect the global peaks from the smoothed histogram
CR
distribution robustly, we utilize the slope difference distribution
that is defined as the variation rate of the histogram distribution
(Wang, 2016). The peaks and valleys of the histogram
distribution usually have much greater variation rates compared
to other parts of the histogram. In addition, the global peaks of
the histogram distribution correspond to the cluster centers of the
pixel classes when they are Gaussian distributed, which is true in
most image segmentation cases. The slope difference distribution
is computed by calculating the difference of the slope values on
the left side and the right side of each point on the smoothed
histogram. The slope values are computed by fitting two straight
lines on both sides of the point on the smoothed histogram with N
points. The proper selection of N could suppress smaller local
peaks while retain larger global peaks. The means of different
pixel classes correspond to the global peaks of the slope
difference distribution distinctly and are resistant to the problem
caused by the distribution overlap of different pixel classes.
This paper is organized as follows. In section 2, we analyzed
the problems of image segmentation and described the proposed
approach. In section 3, both synthesized images and natural
images are segmented to validate the proposed approach.
Discussions are given in section 4. Section 5 concludes the paper.
2. The proposed approach
The proposed approach deals with these four described
challenging problems one by one. Firstly, the means of different
pixel classes are computed from the global peaks of the
histogram distribution. The advantages of computing the mean
values of the pixel classes based on the global peaks of the
histogram distribution include: (1), If we calculate the cluster
centers of the pixel classes from the global peaks in the
histogram distribution, the overlapping parts of different pixel
classes at the valley parts of the histogram will not affect the
calculation accuracy. (2), Since the peaks of the histogram
distribution are isolated from each other, they will not affect each
others parameter calculation accuracy either.
As the estimation of the probability distribution of a
continuous variable, the histogram distribution of the image plots
the number of pixels for each gray-scale value. It was utilized by
different researchers in image quality evaluation, pixel clustering
and image segmentation. However, its functionality is usually
limited by the complexity of the image and the noise that
modifies the histogram. As a result, it is usually difficult to
compute the global peaks from the histogram distribution directly.
In this paper, we compute the slope difference distribution
(Wang, 2016) from the histogram distribution first and then
calculate the global peaks of the histogram distribution from the
peaks of the slope difference distribution. The calculation of the
slope difference distribution is summarized as follows.
 3
ACCEPTED MANUSCRIPT
The histogram distribution P x is formulated as: 0 50, 0 20; 1 100,1 20; 2 150, 2 20 . The
Ni process of finding the global peaks by the slope difference
P x i ; i G1 ,..., G2 (1) distribution for each synthesized image is demonstrated in Fig. 1
Nj
(c) and (d) respectively. The histogram distribution P x
j argmax N k (2)
k G1 ,G2
computed by Eq. (1) is drawn in green, the slope difference
distribution computed by Eq. (7) is drawn in blue. The blue
where N i denotes the frequency of the gray-level i and N j crosses corresponding to the global peaks are computed by Eq.
denotes the maximum frequency which occurs at j in the (8). In Fig. 1 (c), the slope difference distribution has three peaks
interval [G1 , G2 ] . G1 is the minimum gray-level of the image and that correspond to the mean values of the background, the dark
object and the bright object respectively. Because of the
G2 is the maximum gray-level of the image. relatively smaller noise magnitude, the computed slope
For any point, i on the histogram distribution, there are two difference distribution is smooth and the global peaks could be
slopes, one on the left and the other on the right. We compute the detected by Eq. (8). On the contrary, the noise magnitude in Fig.
slope by fitting a straight line with N adjacent points on each side 1 (b) is relatively larger. Consequently, the computed slope
of the point, i . The linear model of the straight line is formulated difference distribution is not smooth enough for Eq. (8) to detect
as: the global peaks correctly.

T
yi axi b (3)
BT B BT Y
1
a, b
T

IP
(4)
x1 1

CR
x 1
B 2
(a) (b)
(5)


xN 1

y1 , y2 ,, yN
Y

US
T
(6)
Two slopes at point i , a1 i and a2 i are obtained from Eq.
(4). The slope difference, s i , at point i is computed:
AN
s i a2 i a1 i ; i G1 N ,..., G2 N (7)
(c) (d)
Fig. 1. Demonstration of finding the global peaks by the slope difference
The number N of the fitting points in the line model is critical distribution. (a) Synthesized image with noise level 0, 10 ; (b)
for magnifying the global peaks while suppressing the local Synthesized image with noise level 0, 20 ; (c) Finding the global
peaks. Thus, its selection is very important and needs some peaks by slope difference distribution computed for the first synthezied image;
M

calibration process to determine its value for a specific class of
images. Fortunately, there is a wide range of N values that could
meet the requirement and its default value is chosen as 15 for the
histogram in the interval of [0, 255].
ED
(d) Finding the global peaks by slope difference distribution computed for the
second synthezied image.
From Fig. 1 (d), we see the second problem that the noise
modifies the histogram distribution and causes great errors during

The positions where the global peaks occur correspond to the
mean values of the pixel classes in the image. To compute the
peaks positions from the slope difference distribution, we set the
derivative of the slope difference distribution s x to zero and
PT
calculation of the pixel mean values. To eliminate or reduce the
bad impact of noise, we propose the following low-pass filtering
method in the frequency domain. The histogram distribution,
P x is transformed into the frequency domain by the Discrete

select the solutions corresponding to the peak parts. Fourier Transformation (DFT):
G2 2 kx
ds i

dx
0; s( x) 0 (8) F k P x e
x G1
T
; k G1 ,..., G2 (9)
CE

We obtain the candidate mean values of different pixel classes where

by solving the above equations. We define the peaks whose T G2 G1 1 (10)
magnitudes are greater than a threshold (0.05 in this study) as the
valid peaks. The candidate means that correspond to the valid In the frequency domain, the low frequency parts of F k are
AC

peaks are sorted. The number of pixel classes is determined by kept while the high frequency parts of it are eliminated by the
the total number of the sorted means or manual input. When the following equation.
F k ; k G1 , G1 1,..., G1 B 1
number of pixel classes is determined manually, the means are
selected one by one based on their magnitudes. The mean value
that corresponds to the peak with the greatest magnitude is F k F k ; k G2 B 1,..., G2 1, G2 (11)
selected as the first mean. 0; k G1 B,..., G2 B

To demonstrate the process of finding the global peaks by
slope difference distribution, we synthesized two images with where B denotes the bandwidth of the low pass filter and we
two objects and one background in Fig. 1 (a) and (b) respectively. set its default value as 10 based on testing a great variety of
The synthesized image in Fig. 1 (a) is with the pixel mean value images. After transforming the histogram from the frequency
of the background equal 50, the pixel mean value of the first domain back into the time domain by the following equation, we
object equal 100 and the pixel mean value of the second object get the smoothed histogram, P x .
equal 150. For the first synthesized image, we add the Gaussian 2 xk
1 G2
P x F k e T ; x G1 ,, G2
i
noise with the magnitude 0, 10 and the image model (12)
T k G1
is 0 50, 0 10; 1 100,1 10; 2 150, 2 10 . For
the second synthesized image, we add the Gaussian noise with
the magnitude 0, 20 and the image model is

ACCEPTED MANUSCRIPT
Fig. 2. Demonstration of the effectiveness of the proposed filter and the
correctness of finding the global peaks by the computed slope difference
distribution.
To demonstrate the effectiveness of the proposed low-pass
filtering method, we show the process of finding global peaks by
the slope difference distribution computed from the smoothed
histogram distribution for the synthesized image (Fig. 1 (b)) in
Fig. 2. As can be seen, the proposed filter could remove the noise
from the histogram distribution effectively and the global peaks
could be detected correctly after the slope difference distribution
was computed from the smoothed histogram distribution. The
reason we use the Fourier transform based low pass filter instead
of other filters is that it is capable of preserving the shape of the
low frequency part well while removing the high frequency parts
effectively.
After the mean values of all the pixel classes, l ; l 1,.., L are
computed, the image is labeled by the following equation.
Ilabel (u, v) l; u 1,..,U ; v 1,..,V (13)
where (u, v) denotes the position of the pixel in the image and
U V denotes the resolution of the image. L denotes the total
number of pixel classes in the image and l denotes the pixel class
that is computed as:
l arg min I (u, v) l
(14)
(a) (b)
Fig. 3. Demonstration of labelling the image. (a) The labelled image by
Eqs. (13-14); (b) The labelled image after Gibbs energy minimization.
ED
We computed the labeled result of the synthesized image (Fig.
1 (b)) by Eqs. (13-14) and showed it in Fig. 3 (a). It is seen that
the labeled regions are inhomogeneous, which is caused by the
third challenging problem that the local properties of different
PT
regions in the image are usually not homogeneous at the same
spatial scale. This local property could be formulated by the
Gibbs distribution that characterizes the image as a Markov
random field X = {Xs} with a joint distribution (Derin and Elliott,
CE
1987; Caillol et al., 1993; Wang and Zhang, 2012).
e cC c
1 V ( x)
P( X x)
V ( x)
e cC c
xL
AC
where Vc ( x) is the potential function associated with clique c
and it is defined as follows:
; all values in c are equal
Vc ( x) (16)
; else
where is a constant and its default value is 1.
To eliminate the inhomogeneity in the labelled image, we
utilize the Gibbs distribution to minimize its total energy over the
whole image.
L U V
I o arg min E( I
Ilabel ( u , v )[1, L ] l 1 v 1 v 1
label (u, v)) (17)
E ( Ilabel (u, v)) cC Vc ( x) (18)
where I o is the labeled image with Gibbs energy
minimization and Fig. 3 (b) shows the labelled image with Gibbs
energy minimization for the synthesized image in Fig. 1 (b). As
can be seen, the inhomogeneity inside each region has been
4
eliminated successfully. For the practical implementation of the
Gibbs energy minimization, please refer Wang and Zhang, 2012.
After the image is labeled and its inside inhomogeneity is
eliminated, the boundaries of the labeled objects are extracted as
shown in Fig. 4 (a). It is seen that the Gibbs energy minimization
causes the boundary to be noisy and the extracted boundary does
not match the objects true boundary well. In addition, there are
two different boundaries extracted for the brightest object, which
is caused by the fourth challenging problem that the boundary of
the labeled region is affected greatly by the overlap the grayscale
distributions of different pixel classes. We define the problem
that multiple boundaries are extracted from the segmentation
result for the same object as the pixel distribution overlapping
problem. This problem could not be avoided for the labelling
method based on the Euclidean distance (Eqs. (13-14)) and it
frequently appears in other literatures (Carson et al., 2002;
Felzenszwalb and Huttenlocher, 2004).
T
IP
CR
(a) (b)
Fig. 4. Demonstration of boundary delineation. (a) Extracted boundary
from the labelled image with Gibbs energy minimization; (b) Delineated
boundary after deforming the bounday in (a) to the nearest Canny edges.
US
AN
(a) (b)
Fig. 5. Comparison of boundary with detected canny edges. (a) Extracted
boundary overlaying on the detected canny edges; (b) Final smooth boundary
overlaying on the detected canny edges.
To refine the extracted boundary B , we propose the
following boundary deformation method.
M
Step 1: the edge of the image is detected with the canny
operator (Canny, 1986).
Step 2: the extracted boundaries are deformed to the nearest
detected edges by minimizing the following energy function.
R
arg min B(r ) E(q)
r 1 q[1,2...,Q ]
(19)
where B denotes the boundary extracted from the segmented
object by Eqs (1)-(18) . R and r denote the total number of
points and the index of the points on it respectively. E denotes
the detected canny edges. Q and q denote the total number of
points and the index of the points on it respectively. The total
number of points R on the canny edges are usually much greater
than the total number of the points Q on the extracted boundary.
(15)
There exist R points on the extracted canny edges that make
Eq. (19) minimum and they form the deformed boundary that is
denoted as E . As can be seen, the points on the deformed
boundary are all from the detected canny edges.
The final smooth boundary Bs is computed by minimizing the
energy between the deformed boundary E and the fitted
polynomial spline.
2
2 d 2 Bs (t )
E ( Bs ) (1 ) Bs ( j ) E ( j ) dj dt (20)
dt 2
where is the smoothing factor and its default value is 0.5.
Fig. 4 (b) shows the final smooth boundary obtained from the
extracted boundary in Fig. 4 (a) by Eqs. (19-20). As can be seen,
the final smooth boundary match the objects true boundary well
and the accuracy of the extracted boundary is improved
significantly. To show the matching degree of the detected Canny
edges with the extracted boundaries from the labelled objects and
the final smooth boundaries respectively, we show the overlay of
them in Fig. 5 (a) and Fig. 5 (b) respectively. In (a), most of the
extracted boundaries from the labelled objects do not match the
 5
ACCEPTED MANUSCRIPT
detected Canny edges while most of the final smooth boundaries acceptable. This demonstration indicates that there is a wide
match the detected Canny edges well in (b). range of the bandwidth for the Fourier transform based low-pass
filter to achieve an acceptable performance (at least the range [5
3. Experimental results 40] meets the requirement).
3.1. Performance analysis
The proposed method depends on the fitting number N, the
Fourier transform based low-pass filter and the Canny edge to
achieve the optimal segmentation performance. In this section,
we will analyze the performance of the proposed method in detail
with demonstrations. Firstly, we demonstrate how the fitting
number N affects the segmentation performance in Fig. 6. We
change the magnitude of the Gaussian noise of the synthesized
image to 0, 15 and segment it by the proposed (a) (b)
method with different values of N. The fitting number N is
chosen as 5, 15, 25 and 50 respectively. Their corresponding
slope difference distributions are plotted in Fig. 6 (a), (b), (c) and

T
(d) respectively. The corresponding segmentation results are
shown in Fig. (e), (f), (g) and (h) respectively. As can be seen,

IP
the number of the peaks and valleys of the slope difference
distributions decrease with the increase of the value of N. The
automatic computed clustering centers are correct when N equals

CR
5, 15 and 25. However, when N equals 45, the computed (c) (d)
clustering center is incorrect. This demonstration indicates that
there is a wide range of N that could meet the requirement to
make the proposed method to achieve an acceptable performance
(at least the range [5 25] meets the requirement).

US (e) (f) (g) (h)

Fig. 7. Demonstration of how the bandwidth of the Fourier transform
based low-pass filter affects the segmentation performance with the
AN
systhesize image (its noise level equals 0, 15 ). (a) clustering
results of the slope difference distribution when the bandwidth equals 5; (b)
clustering results of the slope difference distribution when the bandwidth
equals 20; (c) clustering results of the slope difference distribution when the
bandwidth equals 40; (d) clustering results of the slope difference distribution
M

(a) (b) when the bandwidth equals 50; (e) the segmentation result when the
bandwidth equals 5; (f) the segmentation result when the bandwidth equals 20;
(g) the segmentation result when the bandwidth equals 40; (h) the
segmentation result when the bandwidth equals 50.
ED

Thirdly, we demonstrate how the Gaussian noise level affects

the segmentation performance in Fig. 8. The Gaussian noise level
will affect the detected edges by Canny operator and the
clustering accuracy of slope difference distribution. We select the
PT

noise level of the synthesized image as 20, 30, 40 and 50

(c) (d) respectively. The detected Canny edges and the final
segmentation results for these four synthesized images are shown
in Fig. 8 (a)-(h) respectively. As can be seen, the quality of the
CE

detected Canny edges decreases with the increase of the noise

level, so does the segmentation performance. However, the
increase of the noise level has more adverse effect on the
clustering accuracy of slope difference distribution than on the
AC

(e) (f) (g) (h)

detected Canny edges. Hence, we conclude that the Canny
Fig. 6. Demonstration of how the fitting number N affects the
segmentation performance with the systhesize image (its noise level equals detection could withstand the Gaussian noise level as long as the
0, 15 ). (a) clustering results of the slope difference distribution slope difference distribution could withstand. In addition, noise
when N equals 5; (b) clustering results of the slope difference distribution
will cause burrs on the detected Canny edges, which will
when N equals 15; (c) clustering results of the slope difference distribution decrease the segmentation performance.
when N equals 25; (d) clustering results of the slope difference distribution
when N equals 45; (e) the segmentation result when N equals 5; (f) the
segmentation result when N equals 15; (g) the segmentation result when N
equals 25; (h) the segmentation result when N equals 45.

Secondly, we demonstrate how the Fourier transform based

low-pass filter affects the segmentation performance. From Eq. (a) (b) (c) (d)
(11), it is seen that the bandwidth of the filter determines the
output of the filter. Hence, we use different bandwidth values to
compare the segmentation performance. We select the bandwidth
as 5, 20, 40 and 50 respectively. The clustering results by slope
difference distribution and the final segmentation results are
shown in Fig. 7 (a)-(h) respectively. The final segmentation (e) (f) (g) (h)
Fig. 8. Demonstration of how the Gaussian noise level affects the
results are acceptable when the bandwidth equals 5, 20 and 40.
segmentation performance with the systhesize image. (a) canny edges of the
However, when the bandwidth equals 50, the segmentation is not
 6
ACCEPTED MANUSCRIPT
synthesized image with Gaussian noise magnitude 0, 20 ; (b) final approach depends on the shape of the filtered histogram, thus we
segmentation result of the synthesized image with Gaussian noise magnitude say that the shape of histogram is kept well by the DFT low filter.
0, 20 ; (c) canny edges of the synthesized image with Gaussian Table 1. Comparison of the accuracy of computing the clustering centers
by the proposed approach with different filters
noise magnitude 0, 30 ; (d) final segmentation result of the
Filters\Ground Truths 50 100 150
synthesized image with Gaussian noise magnitude 0, 30 ; (e)
canny edges of the synthesized image with Gaussian noise magnitude butter-worth filter 60 108 157
0, 40 ; (f) final segmentation result of the synthesized image with
Chebyshev Type I filter 59 107 156
Gaussian noise magnitude 0, 40 ; (g) canny edges of the
synthesized image with Gaussian noise magnitude 0, 50 ; (h) final Chebyshev Type II filter 55 104 153
segmentation result of the synthesized image with Gaussian noise magnitude Elliptic filter 56 105 153
0, 50 .
FIR low pass filter 51 99 149
We also demonstrate how the Salt and Pepper noise affects
moving average filter 51 99 149
the segmentation performance in Fig. 9. We select the Salt and
Pepper noise density of the synthesized image as 0.8, 1.6, 2.4 and DFT filter 50 100 150
3.2 respectively. The detected Canny edges and the final

T
segmentation results for these four synthesized images are shown
in Fig. 9 (a)-(h) respectively. As can be seen, the quality of the

IP
detected Canny edges decreases with the increase of the noise 3.2. Results of segmenting synthesized images
level, so does the segmentation performance. However, the To control the image model precisely, we synthesized several
increase of the noise level has more adverse effect on the images to demonstrate the advantage of the proposed approach

detected Canny edges than on the clustering accuracy of slope
difference distribution. Hence, we conclude that the canny
detection could withstand less Salt and Pepper noise level than
the slope difference distribution could withstand.
CR
over state of the art methods. We compared the performance of
the proposed approach with eleven methods with available
MATLAB codes: (1), Kernel Graph Cuts (Salah et al., 2011); (2),
Normalized Graph Cuts (Shi and Malik, 2000); (3) Graph

US(Felzenszwalb and Huttenlocher, 2004); (4), EM (Dempster et al.,

1977); (5), K-means (Jain, 2010); (6), Active contours without
edge (Chan and Vese, 2001); (7), GM (Chan et al., 2006); (8),
Distance regularized level set (Li et al., 2010); (9), Region based
AN
active contour (Lankton and Tannenbaum, 2008); (10), GMM
(a) (b) (c) (d) (Nguyen et al., 2013); (11) FFCM (Nguyen and Wu, 2013). We
set the iteration of active contour and level set methods as 2000
and the number of N in fitting straight lines during computing the
slope difference distribution as 15 in all the experiments
M

conducted in this research work. We select the same number of

the pixel classes for the proposed approach, K-means method and
(e) (f) (g) (h) EM method in segmenting the tested images. The parameters of
Fig. 9. Demonstration of how the Salt and Pepper noise level affects the other state of the art methods are kept as their default values.
ED

segmentation performance with the systhesize image. (a) canny edges of the
synthesized image with Salt and Pepper noise density 0.8; (b) final
segmentation result of the synthesized image with Salt and Pepper noise
density 0.8; (c) canny edges of the synthesized image with Salt and Pepper
noise density 1.6; (d) final segmentation result of the synthesized image with
Salt and Pepper noise density 1.6; (e) canny edges of the synthesized image
PT

(a) (b) (c) (d) (e) (f)

with Salt and Pepper noise density 2.4; (f) final segmentation result of the
synthesized image with Salt and Pepper noise density 2.4; (g) canny edges of
the synthesized image with Salt and Pepper noise density 3.2; (h) final
segmentation result of the synthesized image with Salt and Pepper noise
CE

density 3.2.
To prove the DFT low pass filter is good at preserving the (g) (h) (i) (j) (k) (l)
shape of the low frequency part well, we design the following Fig.10. Results of segmenting synthesized noisy images; (a) Proposed; (b)
experiments. We compare it with other well-known filters Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph Cuts (Shi and
quantitatively in maintaining the means of the pixel classes. We Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher, 2004); (e) EM
AC

vary the magnitude of the Gaussian noise from 0, 10

to 0, 20 and get 10 synthesized images. We then use
different low pass filters to filter the histogram for the proposed
approach and compute the clustering centers for these 10
synthesized images. For any low pass filter, the computed
clustering centers almost keep the same for these 10 synthesized
images. We use their averages to compare the performances of
different low pass filters. The comparison is shown in Table 1.
The compared filters include Infinite Impulse Response (IIR)
butter-worth low pass filter, IIR Chebyshev Type I low pass
filter, IIR Chebyshev Type II low pass filter, IIR Elliptic low
pass filter, FIR low pass filter and the time domain moving
average filter. Only when the histogram is filtered by the
designed DFT low pass filter, the proposed approach could find
the clustering centers accurately, which indicates that the centers
of the pixel classes are kept well by the DFT low pass filter.
Because the calculation of the clustering centers by the proposed
(Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active contours
without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i) Distance
regularized level set (Li et al., 2010); (j) Region based active contour
(Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013); (l) FFCM
(Nguyen and Wu, 2013).
We synthesize an image with three ellipse objects on the
background in Fig. 10 and their means and variances are {200,
30}, {150, 30}, {100, 30} and {50, 30} respectively. As can be
seen, the proposed approach, Normalized graph cuts and k-means
perform better than other methods for this type of images.
Distance regularized level set and region based active contour are
affected by the noises greatly. Compared to them, GMM, FFCM,
active contour without edge and GM are more accurate and
efficient. However, they could only separate one ellipse
successfully. Kernel graph cuts and EM fail to separate the
smallest ellipse from the background. Graph segmented the
image into too much small regions while the boundaries of the
segmented ellipses are not accurate.
 7
ACCEPTED MANUSCRIPT
Fig.13. Results of segmenting another synthesized fuzzy images; (a)
Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph
Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher,
2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i)
(a) (b) (c) (d) (e) (f) Distance regularized level set (Li et al., 2010); (j) Region based active
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013);
(l) FFCM (Nguyen and Wu, 2013)..

3.3. Results of segmenting images from Berkeley dataset

(g) (h) (i) (j) (k) (l)
Fig.11. Results of segmenting synthesized fuzzy images; (a) Proposed; (b)
Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph Cuts (Shi and
Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher, 2004); (e) EM
(Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active contours
without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i) Distance
regularized level set (Li et al., 2010); (j) Region based active contour
(Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013); (l) FFCM
(Nguyen and Wu, 2013).
In this section, we show the results of segmenting some typical
images from the Berkeley dataset (Martin, D. and Fowlkes, C.) to
verify the superiority of the proposed approach over state of the
art methods. The color image is converted to the grayscale image
by selecting the first RGB image channel directly and then the
proposed approach and state of the art methods are applied on the
grayscale image one by one.

T
We blur the synthesized image in Fig. 10 by minimizing the
Gibbs distribution of the image to get the fuzzy image. The

IP
segmentation results are shown in Fig. 11 (a)-(j). At this time, k-
means failed to segment the image correctly while the proposed (a) (b) (c) (d) (e) (f)
approach and normalized cuts still perform well. All the other

CR
tested methods perform unsatisfactorily.

(g) (h) (i) (j) (k) (l)

(a) (b) (c) (d) (e) (f)

US Fig.14. Results of segmenting a typical image from Berkeley dataset; (a)
Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph
Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher,
2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i)
AN
Distance regularized level set (Li et al., 2010); (j) Region based active
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013);
(l) FFCM (Nguyen and Wu, 2013)..
(g) (h) (i) (j) (k) (l)
Fig. 14 (a)-(j) showed the results of segmenting the plane from
Fig.12. Results of segmenting another synthesized noisy images; (a)
Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph the sky background by different methods. As can be seen, the
kernel graph cuts and normalized graph cuts methods tend to
M

Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher,
2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active segment the image into some regions unexpectedly. The Graph,
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i) EM and K-means methods yield too many regions. The active
Distance regularized level set (Li et al., 2010); (j) Region based active contour without edge, GM, GMM, FFCM, Distance regularized
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013);
ED

(l) FFCM (Nguyen and Wu, 2013)..
We synthesize a more complex image with four rectangular
objects on the background in Fig. 12 and their means and
variances are {230, 30}, {180, 30}, {140, 30}, {90, 30} and {40,
30} respectively. As can be seen, only the proposed approach
PT
level set and region based active contour methods failed to
delineate the boundaries of the plane accurately. The result of the
proposed approach appears to be better. We compare the
segmentation results on another 3 images in Figs. 15-17. The
visual results verified the superiority of the proposed approach

could segment the rectangles successfully while all the other over state of the art methods.
methods perform badly, but differently. In addition, some state of
the art methods performed relatively better than other state of the
art methods. For instance, FFCM performs the second best in this
CE

specific case.
We blur the synthesized image in Fig. 12 by minimizing the (a) (b) (c) (d) (e) (f)
Gibbs distribution of the image again to get next fuzzy image.
The segmentation results are shown in Fig. 13 (a)-(j). At this time,
AC

both the proposed approach and the EM method could segment

the four rectangles successfully while the proposed approach has
better accuracy than the EM method. All the other tested state of
(g) (h) (i) (j) (k) (l)
the art methods failed to give acceptable segmentation results in Fig.15. Results of segmenting a typical image from Berkeley dataset; (a)
this specific example. From all the synthesized images, we could Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph
see that the proposed approach is more robust than state of the art Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher,
methods. 2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i)
Distance regularized level set (Li et al., 2010); (j) Region based active
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013);
(l) FFCM (Nguyen and Wu, 2013)..

(a) (b) (c) (d) (e) (f)

(a) (b) (c) (d) (e) (f)

(g) (h) (i) (j) (k) (l)

 8
ACCEPTED MANUSCRIPT
called VoI and the lower the better. We use the quantitative
results from the literatures directly instead of applying these
state-of-the-art methods on the Berkeley datasets to avoid
abusing these methods. The comparison is also favorable and the
proposed approach achieves better accuracy in both measures.
(g) (h) (i) (j) (k) (l)
Fig.16. Results of segmenting a typical image from Berkeley dataset; (a) For those methods that we could not find the quantitative results
Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph from the literatures, we design the following experiments to
Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher, acquire the quantitative comparisons. Firstly, we vary the
2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active magnitude of the Gaussian noise of the synthesized image in Fig.
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i) 1 from 1 to 100. Then we de-noise the synthesized noisy images
Distance regularized level set (Li et al., 2010); (j) Region based active
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013); by minimizing the Gibbs energy and generate the fuzzy images.
(l) FFCM (Nguyen and Wu, 2013).. We segment the two objects from the background by different
methods and compare the computed F-measures. The computed
F-measures for the 100 synthesized images by different methods
are plotted in Fig. 20. The average F-measures computed by
different methods are shown in Table. 4. As can be seen, the
(a) (b) (c) (d) (e) (f)
proposed method achieves significantly better accuracy. We

T
synthesize another 100 images with only one object and segment
the object from the background by different methods. The

IP
computed F-measures are compared in Fig. 21 and the average F-
measures are compared in Table. 5. As can be seen, the proposed
method also achieves better accuracy. We show the ground truth

CR
(g) (h) (i) (j) (k) (l) boundaries for the first dataset and for the second dataset in Fig.
Fig.17. Results of segmenting a typical image from Berkeley dataset; (a) 22 (a) and (b) respectively for better understanding of the
Proposed; (b) Kernel Graph Cuts (Salah et al., 2011); (c) Normalized Graph
Cuts (Shi and Malik, 2000); (d) Graph (Felzenszwalb and Huttenlocher, designed experiments. The ground truth boundaries are computed
2004); (e) EM (Dempster et al., 1977); (f) K-means (Jain, 2010); (g) Active when the noise level is zero and they are overlaying on the
contours without edge (Chan and Vese, 2001); (h) GM (Chan et al., 2006); (i) synthesized fuzzy image with the magnitude of the Gaussian
contour (Lankton and Tannenbaum, 2008); (k) GMM (Nguyen et al., 2013);
(l) FFCM (Nguyen and Wu, 2013)..
3.4. Quantitative results
US
Distance regularized level set (Li et al., 2010); (j) Region based active noise equal 10.
AN
In this section, we evaluate the proposed approach
quantitatively. In this study, we first compare the accuracy of the
proposed approach with that of the EM and K-means methods in (a) (b)
finding the pixel mean values. We synthesize an image dataset Fig.18. Demonstration of the synthesized images in the dataset (a) The
with 100 images in total. In the synthesized dataset, one hundred synthesized noisy image; (b) The fuzzy image obtained by minimizing the
M

Gibbs energy of the synthesized image in (a).

images were generated with the global parameter y =
0 50,V ; 1 100, V ; 2 150, V; 3 200, V and V varies
from 10 to 110. In the synthesized image, there are four pixel
ED

classes with mean values equal to 50, 100, 150 and 200
respectively and their variances are denoted by V . Fig. 18 (a)
shows the first synthesized image with V equal to 10. Before
computing the pixel mean values by the compared methods, we
PT

de-noise the synthesized noisy images by minimizing the Gibbs

energy first and produce the corresponding fuzzy images. Fig. 18
(b) shows a produced fuzzy image after de-noising the
synthesized image in (a). Then, the proposed approach, the K-
CE

means method and the EM method calculate the pixel mean

values of different pixel classes from the de-noised fuzzy images.
The errors are computed by the following formulation:
Er 0 0 1 1 2 2 3 3 (21)
AC

Fig.19. Comparison of the errors in finding the pixel means

where 0 , 1 , 2 and 3 are the computed pixel mean
values by the proposed approach, the EM method or the K-means
method.
We plot the average errors for the 100 synthesized images by
the EM method, K-means method and the proposed approach in
Fig. 19. As a whole, the plotted errors of EM and K-means
methods are above the plotted error of the proposed approach.
The quantitative comparison is shown in Table. 2. As can be seen,
the proposed approach is significantly more accurate than EM
and K-means in calculating means for different pixel classes
during image segmentation.
For the quantitative comparison of the segmentation accuracy,
we first compared the proposed approach, Kernel graph cuts,
Normalized graph cuts, Graph, Mean-shift and K-means in Table.
3 with two measures: the probabilistic rand index (Unnikrishnan
et al., 2005) and variation of information (Yang et al., 2008) Fig.20. Comparison of the F-measures computed by different methods for the
using the Berkeley dataset. The probabilistic rand index is called first synthesized image set (LR-AC represents localized region based active
contour; NE-AC represents active contour without edge; DR-LS represents
PRI and the higher the better. The variation of information is distance regularized level set.)
 9
ACCEPTED MANUSCRIPT

(a) (b)
Fig. 22. Demonstration of groundtruth boundary overlaying on the
synthesized fuzzy image with Gaussian noise magnitude equal 10. (a)
Groundtruth boundary for the first synthesized image set; (b) Groundtruth
boundary for the second synthesized image set.

Fig.21. Comparison of the F-measures computed by different methods for the

T
second synthesized image set (LR-AC represents localized region based
active contour; NE-AC represents active contour without edge; DR-LS

IP
represents distance regularized level set.)

Table. 2 Comparison of the proposed approach with EM and K-means

CR
Methods Average error Average time
EM 7.8360 1.07 s
K-means 7.8177 0.11 s
Proposed 5.7431 0.12 s

Measure
PRI
VoI
Proposed
0.7984
2.2247
Kernel graph cuts
0.7650
2.4091
US
Table. 3 Comparison of the proposed approach with state of the art methods
N- cuts
0.7229
2.9329
Graph
0.7841
2.6647
Mean-shift
0.7555
2. 477
K-means
0.7882
2. 304
AN
Table. 4 Comparison of the average F-measure computed by different methods from the first synthesized image
Measure Proposed Region based Active contours Distance regularized GM EM K-means FFCM GMM
active contour without edge level set
F-measure 0.9860 0.8937 0.9482 0.9385 0.7893 0.9277 0.8051 0.8001 0.7990
M

Table. 5 Comparison of the average F-measure computed by different methods from the second synthesized image
Measure Proposed Region based Active contours Distance regularized GM EM K-means FFCM GMM
active contour without edge level set
F-measure 0.9801 0.9369 0.9617 0.9015 0.4015 0.9438 0.7038 0.5080 0.5658
ED

4. Discussion histogram distribution and the noise distribution. In this paper,

we propose an approach that utilizes the slope difference
Image segmentation is important for many computer vision distribution of the image histogram to calculate the means for
applications and thus a lot of efforts have been put into designing different pixel classes. The proposed approach could effectively
PT

image segmentation algorithms to solve specific problems. After
decades of evolution, the segmentation methods converge to
several categories, e.g. graph based methods (Felzenszwalb and
Huttenlocher, 2004; Shi and Malik, 2000; Salah et al., 2011),
CE
deal with the noise and overlaps of multiple distributions. As it
turned out that the slope difference distribution method is more
accurate than EM and K-means methods. For the local property

of the image, we utilize the Gibbs distribution to characterize its

stochastic modeling based methods (Carson et al., 2002; Won
and Gray, 2004; Jain, 2010; Wang and Zhang, 2012) and
variational formulations based methods (Chan and Vese, 2001;
Li et al., 2010; Lankton and Tannenbaum, 2008; Arbelaez et al.,
AC
distribution. We utilize an iterative gradient descent method to
minimize the Gibbs energy in the image. To refine the boundary,
we deform the extracted boundary to the nearest Canny edges
and use polynomial splines to achieve the global smoothness. In

2011). Unfortunately, these methods are only good at segmenting nature, the proposed approach is different from the traditional
some specific types of images while perform poorly in other categories of state of the art segmentation methods. In
types of images. As a result, it is necessary to exploit new theory performance, the proposed approach is more robust than state of
or directions to come up with more effective methods that could the art methods because it utilizes the properties of the image
segment more classes of images robustly. more generally.
Image segmentation remains challenging because the vast
variety of images are usually with different properties. For 5. Conclusion
instance, the histogram distributions of different types of images
In this paper, a new direction of image segmentation based on
might comply with different mathematic formulations, although
both the local property and the global property of the image is
they are assumed to be Gaussian distributed in most cases. The
exploited. The proposed approach calculates the means of
noise and the overlapping of different pixel classes in the
different pixel classes from the slope difference distribution and
histogram distribution make the estimation of the parameters
it is significantly more accurate than the benchmark EM and K-
challenging. Furthermore, the inhomogeneous spatial scale of
means methods. The image is labelled by clustering the pixel to
different regions makes the segmentation more difficult. In
the nearest means. To improve the accuracy of the labelled result,
general, the image has local properties and global properties that
we suppress the inhomogeneity inside the image by minimizing
should be utilized effectively during segmentation. EM and K-
the Gibbs energy function. To improve the accuracy of the
means have been served in estimating the global parameters in
extracted boundary, we deform the boundary to the nearest
many image segmentation applications. Unfortunately, their
Canny edges and then optimize the whole boundary with
accuracies are limited by the complexity and variety of the image

ACCEPTED MANUSCRIPT
polynomial splines. Experiments were conducted on a large
quantity of images and experimental results show that this new
segmentation approach is very promising.
Acknowledgments
The author would like to thank the Chinese Academy of
Sciences for the funding with the grant number Y5A1270101.
10
Wang, Z.Z., Zhang, Y.M., 2012 Segmentation of fuzzy images: a novel and
fast two-step pseudo MAP method, Machine Vision and Applications,
23(6), 1209-1218.
Won, C.S., Gray, R.M., 2004 Stochastic image processing. Information
technology: Transmission, Processing, and Storage.
Yang, A.Y., Wright, J., Ma, Y., Sastry, S.S., 2008 Unsupervised
segmentation of natural images via lossy data compression, Computer
Vision and Image Understanding, 110, 212225.

References

Arbelaez, P. , Maire, M., Fowlkes, C., Malik, J., 2011 Contour detection and
hierarchical image segmentation, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 33(5), 898916.
Besag, J., 1986 On the statistical analysis of dirty pictures. Journal of the
Royal Statistical Society. Series B., 48(3), 259-302.
Caillol, H., Hillon A., Pieczynski, W., 1993 Fuzzy random fields and

T
unsupervised image segmentation. IEEE Transactions on Geoscience
and Remote Sensing GE-31, 4, 801810.

IP
Canny, J., 1986 A computational approach to edge detection, IEEE T PAMI,
8(6), 679-698.
Carson, C., Belongie, S., Greenspan, H., Malik, J., 2002 Blobworld: image
segmentation using expectation-maximization and its application to

CR
image querying, IEEE Transactions on Pattern Analysis and Machine
Intelligence, 24(8), 10261038.
Chan, T.F., Esedoglu, S., Nikolova, M., 2006 Algorithms for finding global
minimizers of image segmentation and denoising models, SIAM J. Appl.
Math., 66(5), 1632-1648.
Chan, T.F., Vese, L.A., 2001 Active contours without edges, IEEE
Transactions on image processing, 10(2), 266277.
Chen, S.F., Cao, L.L., Wang, Y.M., Liu, J.Z., Tang, X.O., 2010 Image
Segmentation by MAP-ML Estimations. IEEE Transactions on Image
Processing, 19(9), 2254-2264.
US
AN
Comaniciu, D., Meer, P., 2002 Mean shift: a robust approach toward feature
space analysis. IEEE Transactions on pattern analysis and machine
intelligence, 24(5), 603-619.
Dempster, A.P., Laird, N.M., Rubin, D.B., 1977 Maximum likelihood from
incomplete data via the EM algorithm. Journal of the Royal Statistical
Society, Series B, 39(1), 1-38.
M

Derin, H., Elliott, H., 1987 Modeling and segmentation of noisy and textured
images using gibbs random fields. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 9(1), 3955.
Felzenszwalb, P.F., Huttenlocher, D.P., 2004 Efficient graph-based image
segmentation. International Journal of Computer Vision, 59(2), 167-181.
ED

Jain, A.K., 2010 Data clustering: 50 years beyond K-means, Pattern

Recognition Letters, 31(8), 651666.
Lankton, S., Tannenbaum, A., 2008 Localizing region-based active contours,
IEEE Transactions on image processing, 17(11), 20292039.
Li, C., Xu, C., Gui, C., Fox, M.D., 2010 Distance Regularized Level Set
PT

Evolution and Its Application to Image Segmentation, IEEE Trans.

Image Processing, 19(12), 3243-3254.
Manjon, J.V., Coupe, P., Buades, A., 2015 MRI noise estimation and
denoising using non-local PCA, Medical image analysis, 22(1), 3547.
CE

Martin, D. and Fowlkes, C., The Berkeley segmentation database and

benchmark, Image Database and source code publicly [online].
Available:http://www.eecs.berkeley.edu/reseaqrch/projects/CS/vision/gr
ouping/segbench/
Nguyen, T.M., Wu, Q.M., Zhang, H., 2013 Bounded generalized Gaussian
mixture model, Pattern recognition, 47, 3132-3142.
AC

Nguyen, T.M., Wu, Q.M., 2013 A fuzzy logic model based Markov random
field for medical image segmentation, Evolving systems, 4, 171-181.
Osuna-Encisa, V., Cuevas, E., Sossa, H, 2013 A comparison of nature
inspired algorithms for multi-threshold image segmentation, Expert
systems with applications, 40, 1213-1219.
Sagar, S., Brando, V.E., Sambridge, M., 2014 Noise estimation of remote
sensing reflectance using a segmentation approach suitable for optically
shallow waters, IEEE Trans. Geosci. Remote Sens., 52(12), 75047512.
Salah, M.B., Mitiche, A., Ayed, B., 2011 Multiregion image segmentation by
parametric kernel graph cuts, IEEE Trans. on Image Processing, 20(2),
545-557.
Shi, J., Malik, J., 2000 Normalized cuts and image segmentation, IEEE Trans.
on Pattern Anal. Mach. Intell., 22(8), 888905.
Unnikrishnan, R., Pantofaru, C., Hebert, M., 2005 A measure for objective
evaluation of image segmentation algorithms, in Proc. IEEE Int. Conf.
Comput. Vis. Pattern Recognit., 3, pp. 3441.
Wang, Z.Z., 2016 A new approach for segmentation and quantification of
cells or nanoparticles, IEEE T Ind. Inform., 12 (3), pp. 962-971.

View publication stats