You are on page 1of 14

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

A Multi-Region Segmentation Method for SAR


Images based on the Multi-Texture Model with
Level Sets
Shiyu Luo, Ling Tong, and Yan Chen

 a good performance, whereas directly applying them to


Abstract—Synthetic Aperture Radar (SAR) image Synthetic Aperture Radar (SAR) images leads to incorrect
segmentation is a difficult problem due to the presence of strong segmentation. Because SAR imaging system is quite different
multiplicative noise. To attain multi-region segmentation for SAR from optical imaging system that SAR is side-looking imaging.
images, this paper presents a parametric segmentation method
based on the multi-texture model with level sets. Segmentation is
This signal acquisition mode produces an inevitably strong
achieved by solving level set functions obtained from minimizing speckle noise, which can be modeled as multiplicative noise in
the proposed energy functional. To fully utilize image information, terms of empirical observation. For SAR images, high noise is
edge feature and region information are both included in the more likely to occur in high grey level areas while low noise
energy functional. For the need of level set evolution, the Ratio of almost existing in low grey level areas. The traditional
Exponentially Weighted Averages (ROEWA) operator is segmentation methods aim at images with additive noise and
modified to obtain edge feature. Region information is obtained
by the Improved Edgeworth Series Expansion (IESE), which can
thus are not suited to SAR images.
adaptively model a SAR image distribution with respect to Due to the difficulties in distinguishing edges from noise for
various kinds of regions. The performance of the proposed low grey level areas in SAR images, a segmentation method
method is verified by three high resolution SAR images. The using only edge feature cannot achieve a good result. Therefore,
experimental results demonstrate that SAR images can be statistical parameters are commonly used for segmentation.
segmented into multiple regions accurately without any speckle Generally, the distribution of a SAR image is considered to
pre-processing steps by the proposed method.
obey some empirically statistical distribution models [9], such
Index Terms—SAR image segmentation, level sets, Edgeworth as the Gamma model [10], the log-normal model [11], the
series expansion, multi-texture model, multi-region segmentation. Weibull model [12], and the G0 model [13]. In the literatures
[14-19], all segmentation methods based on level set are
implemented under the assumption that SAR images follow
I. INTRODUCTION these specific distribution models. However, statistical
characteristics of various types of regions in a SAR image
I mage segmentation has a significantly influence on image
automatic analysis and interpretation. In recent years, a
segmentation method based on level sets under Partial
cannot be completely modelled by a specific distribution. For
example, the Weibull distribution is well suited to model
Differential Equation (PDE) has emerged and developed homogeneous regions such as grassland, whereas it is not
rapidly. This method has many advantages, such as its dedicated to extremely heterogeneous regions such as
simplicity at expansion, robustness to noise, and good mountainous areas [20]. Therefore, the good performance of
performance at segmentation of complex images. SAR image segmentation by using these methods is difficult to
The traditional segmentation methods based on level set achieve without knowing the types of ground objects. Besides,
basically are divided into two categories. One is built on edge segmentation performance decreases sharply with the
feature, such as the typical “snake” model [1], the Geometric increasing number of segmented regions in a SAR image.
Active Contour (GAC) [2], and the Geodesic Active Contour Consequently, most of current methods for SAR image
(GAC) [3, 4]. The other one is based on region information, the segmentation using level set can segment an image only into
representative method is the active contours without edges two regions [14-18, 21-23].
model (C-V model) [5]. The above methods can segment an As far as multi-region segmentation for SAR images
image only into two regions. To extend the methods applied to concerned, a more flexibly and precisely statistical distribution
multi-region segmentation, some methods have been developed, is required to guarantee the good segmentation performance. A
which encountered in the literatures [6-8]. These methods attain few methods have been proposed for multi-region
segmentation of SAR images [19, 24-26]. Similarly, they used
one or more empirically statistical distributions in their models
This work was supported by the National Natural Science Foundation of
China under Grant No.41371340 and No. 41571333.
to attain segmentation. In their experiments [19, 24-26], the
The authors are with the School of Automation Engineering, University of tested images were all acquired from agricultural areas
Electronic Science and Technology of China, Chengdu, 611731 (email: composed of various crops, however, in fact, different crops
shiyu_luo@126.com).

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2

can be considered as being of similarly statistical II. EDGE FEATURE AND REGION INFORMATION FOR SAR
characteristics. Therefore, the data corresponded to different IMAGES
regions in their tested images is theoretically fitted well by one In our previous work [32], we found that it is an effective
specific distribution. The limitation of their experiments was way to enhance the performance of the two-region
that the performances of these segmentation methods for a SAR segmentation of SAR images by integrating edge and region
image composed of entirely different types of land covers e.g. information. On one hand, the performance of methods using
river, urban areas, and vegetation were not tested. Since only edge feature is not satisfactory in case that edge features
different empirical distributions focus on different land covers are not obvious or an image contains many false edge features.
[9], we infer that the methods for multi-region segmentation Moreover, it cannot be applied to multi-region segmentation.
using empirical distribution models are possibly not flexible for On the other hand, segmentation methods using only region
various types of SAR images. information are not suited to images presenting similarly
To this end, we proposed a multi-region segmentation statistical information with respect to different regions, since
method for SAR images based on the multi-texture model with region boundaries are only determined by region information.
level sets. In order to make the most of image information, this Consequently, to fully utilize image information, we combine
method combines the Ratio of Exponentially Weighted edge feature and region information to the level set framework.
Averages (ROEWA) operator [27] and the Edgeworth Series
Expansion (ESE) [28-30] to the level set framework, which is A. The modified ROEWA operator
called multi-texture model here. The contributions of this paper Edge feature is one of the most significant image
are summarized as follows. First, to employ the ROEWA characteristics. The commonly used edge detection operators
operator to obtain edge feature in the level set framework, we for SAR images are the Ratio Of Average (ROA) operator [33],
modify the operator and it is thus not only suited to level set the Generalized Likelihood Ratio (GLR) operator [34], and the
evolution but also enhancing the discrimination of edge ROEWA operator [27]. ROA and GLR operators are based on
intensities. Second, considering the applicable condition of the the monoedge models, they are not appropriate for some actual
ESE fitting data, the Improved Edgeworth Series Expansion SAR images [35]. The ROEWA operator designed for the
(IESE) was developed in terms of the central limit theorem, multiedge case, which computes a ROEWA on opposite sides
which improves the capability of modeling various land covers of the central pixel in the horizontal (X) and vertical (Y)
and is theoretically more flexible for any SAR images. Third, directions, is therefore employed in our study.
aimed at the need of multi-region segmentation, we defined the Edge intensity r  x, y  obtained by the ROEWA operator is
energy functional integrating the modified ROEWA operator expressed by [27],
and the IESE. This energy functional has the general form and
can be generalized for arbitrary number of regions r ( x, y )  rX2 max ( x, y )  rY2 max ( x, y )  0,   (1)
segmentation. We employed the similar segmentation strategy where the two components are calculated as follows,
from [6] to implement segmentation, the segmentation with L   ˆ X1 ( x  1, y ) ˆ X2 ( x  1, y ) 
regions thus only requires log 2  L  level sets. Besides, to rX max ( x, y )  max  , 
  ˆ X2 ( x  1, y ) ˆ X1 ( x  1, y ) 
reduce ambiguities caused by multiple curves overlap, we  (2)
  ˆ Y1 ( x  1, y ) ˆ Y2 ( x  1, y ) 
introduced prior probabilities to the level set framework. rY max ( x, y )  max  ˆ ( x  1, y ) , ˆ ( x  1, y ) 
Different regions can be segmented simultaneously by equal   Y2 Y1 
weight level sets and thereby can attain a better segmentation In (2), ̂X1 , ̂X2 , ̂Y1 , and ̂Y2 are exponentially weighted
for a SAR image. Consequently, the segmentation is achieved averages that can be obtained in terms of a filter f  x 
by solving the PDE obtained from minimizing the energy
calculated as follows,
functional. Since speckle pre-processing step result in the loss
of some edge features [31], this step is not required in this  ˆ X1 ( x, y )  f1 ( x)  ( f ( y )  I ( x, y ))
 ˆ ( x, y )  f 2 ( x)  ( f ( y )  I ( x, y ))
method, thus simplifying the whole process.  X2 (3)

The outline of this paper is as follows. In Section Ⅱ, the  ˆ Y1 ( x, y )  f1 ( y )  ( f ( x)  I ( x, y ))
modified ROEWA operator is briefly formulated and the IESE 
 ˆ Y2 ( x, y )  f 2 ( y )  ( f ( x)  I ( x, y ))
is introduced, respectively. In Section Ⅲ, the energy functional where symbol ∗ denotes the convolution in X-scale, symbol ×
based on the multi-texture model is defined, followed by the denotes the convolution in Y-scale. I  x, y  is the image. In the
solution and its numerical computation method, as well as the
discrete case, the filter f  x  is defined as,
termination condition and the example for the four-region
model. In Section Ⅳ, experiments are carried out to test the 1 a
f ( x)  f1 ( x)  f 2 ( x  1) (4)
segmentation performance of the proposed method, the effect 1 a 1 a
of each term in the proposed level set framework, and the where f1 ( x)  1  a  a x H ( x) , f 2 ( x)  1  a  a  x H ( x) , 0  a  1 .
segmentation performance with respect to various number of
segmented regions. Finally, the conclusion is given in Section H  x  is the Heaviside function.
Ⅴ. Here r=0 denotes that the corresponding pixel is in uniform
areas. The greater r represents the corresponding pixel closer to

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

edge. Nevertheless, learning from the methods proposed in


[36-38], the solving method for the level set framework is
minimizing energy functional. Therefore, in order to guarantee
the solvability of energy functional, edge intensity tending to
zero should indicate that the corresponding pixel is close to
edge, whereas a pixel should be in homogeneous areas when its
corresponding edge intensity value is one. Consequently, we
modified the ROEWA operator to be applicable to level set
evolution. Besides, with the purpose of enhancing the
discrimination of edge intensities, each edge value is
homogenized.
Based on the above analysis, the modified ROEWA operator
is formulated as follows,
 r ( x, y )  min  r( x, y )  
R ( x, y )  histeq 1   (5)
 max  r ( x, y )   min  r( x, y )   (a)
where histeq   denotes histogram equalization. R  x, y   0,1 .
In (5), a small value leads to the decrease of the energy
functional, thus forcing the segment curve to move to the edge.
B. The Improved Edgeworth Series Expansion
In general, SAR data is considered to follow some
empirically statistical distributions. However, one specific
distribution cannot completely describe statistical properties
with respect to homogenous, heterogeneous, and extremely
heterogeneous regions in a SAR image, thereby resulting in the
difficulties in multi-region segmentation. SAR image is better
described by a highly parametric model and it is thus
theoretically more flexible.
The Edgeworth series expansion law indicates that a density,
which is assumed not to be too far from a Gaussian Probability
(b)
Distribution Function (PDF), can be approximated by the
standard normal distribution and the Chebyshev-Herimite
polynomial [39, 40]. However, not all distributions of SAR
images satisfy this assumption. In fact, Gaussian model can be
used for SAR images based on the fact that electromagnetic
waves are scattered from a large number of targets and
randomly distributed throughout a cell, the speckle is fully
developed, and homogeneous surfaces appear as stationary
fields [41]. In some cases, these hypotheses are possibly invalid,
especially for SAR images at high resolution [20]. Therefore,
we transform data distribution in terms of the central limit
theorem, and then fit the transformed distribution using ESE.
This approximation method is called IESE in this paper.
The central limit theorem (Levy-Lindeberg) says that the
mean of a sufficiently large number of independent random
variables can be approximately normally distributed. Hence, (c)
supposing that V  V1 ,V2 , ,Vn ,  is a sequence of Fig. 1. The tested SAR images are of (a) the homogenous region, (b) the
heterogeneous region, and (c) the extremely heterogeneous region.
independent random variables and obeys one same distribution,
t2
 is the mean value of Vk ,  2 is the variance of lim P Sn  v  
v 1
e

2
dt (7)
Vk  k  1, 2,  , a random variables S n is defined as,
n  
2
n
Equation (7) indicates that the distribution of S n approaches
Vk  n to the standard normal distribution when n is sufficiently large.
Sn  k 1
(6) For a SAR image, the number of the pixels can be
n considered as being sufficient. Therefore, assuming that the
Thus the distribution function of S n meets, pixel intensity values of a SAR image are modeled by

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

j-order, which is expressed by [40],


C3 ( s )  s 3  3s
C4 ( s )  s 4  6 s 2  3 (9)
C6 ( s )  s  15s  45s  15
6 4 2

To solve  r , supposing that V ;i denotes ith-original


moment of variable V, then the ith-order cumulants  i  1, 2,3, 4 
of variable V KV ;i is given by [42],
KV ;1  V ;1
KV ;2  V ;2  V2 ;1
(10)
KV ;3  V ;3  3V ;2 V ;1  2V3 ;1
KV ;4  V ;4  4 V ;3 V ;1  3V2 ;2  12 V ;2 V2 ;1  6 V4 ;1

(a) Thus,  r  r  3,4  is expressed by,


KV ;3 KV ;4
3  3 , 4  (11)
n n 4 3

Consequently, the distributions of SAR image intensities


within each region can be modeled in terms of (6) and (8).
We used three real high resolution SAR images to test the
performance of the IESE. The compared statistical models are
the ESE, the Weibull model, the log-normal model, and the G0
model. All images are single-look and intensity images at
resolution of 5.2 m × 7.7 m. They were acquired by Radarsat-2
(C-band) at VV polarization and were not processed by the
multi-look or speckle processing. The three images are of (a)
the homogenous region (grassland and a lake), (b) the
heterogeneous region (urban areas), and (c) the extremely
(b)
heterogeneous region (mountainous areas and a town),
respectively as shown in Fig. 1. Fig. 2 shows the results of
fitting the histograms of the three images by the IESE and the
compared models. In order to exhibit a clear result, the intensity
in Fig. 2 is only shown the range from 0 to 150. It should be
noted that because the tested images are all single-look
intensity data, all the densities look like an exponential [43].
We can observe from Fig. 2 that the fitting accuracy of each
model changes for different images.
The Kullback-Leibler (KL) distance [44, 45] is employed to
quantitatively analyze the fitting results. The shorter KL
distance implies the better performance of fitting distribution.
The results are shown in Table 1. It shows that the ESE
performs worse than the IESE for the three tested images. The
fitting capability of the ESE reduces sharply with the increasing
(c) image complexity, which agrees with the conclusion in [20].
Fig. 2. (a)(b)(c) are the corresponding histograms and the estimated PDFs Besides, the G0 model fits the data best with respect to the
of Fig. 1(a)(b)(c), respectively. heterogeneous and the extremely heterogeneous regions,
V  V1 ,V2 , ,Vn  , where n denotes the number of the pixels, the whereas it is less accurate than other three models with respect
to the homogenous region. The Weibull model is suitable for
distribution of the variable S derived by (6) thus can be
the homogenous region, whereas the fitting accuracy decreases
regarded not to be too far from a PDF. Based on ESE law [28],
such distribution can be approximated by,
Table 1. The KL distances for different fitted models with respect to
 1 1 1 
P( s)  G ( s) 1  3C3 ( s)  4C4 ( s)  32C6 ( s)  (8) region (a), (b), and (c).
 6 24 72  Region a b c
IESE 0.1122 0.2311 0.2126
where G(s) denotes the standard normal distribution.  r is ESE 0.1579 0.2973 0.5259
the rth-order cumulants  r  3, 4  of variable S, C j ( s ) G0 model 0.3048 0.2071 0.1707
Weibull model 0.1028 0.2715 0.4426
 j  3,4,6 is known as the Chebyshev-Herimite polynomial of Log-normal model 0.5270 0.4539 0.2761

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

eight-region segmentation with three level sets.


To achieve multi-region segmentation, we define
characteristic functions Ai  x, y  to distinguish different
regions as follows,
1,  x, y   i
A i  x, y    ,  i  1, 2, , L  (13)
0, else
In the most of cases, the computational stability is poor using
integers, thereby we use the Heaviside function H  z  instead of
0 and 1 [48], where H  z  is defined as H  z   1 when z  0 ,
and H  z   0 when z  0 . H  z  has many forms. The
differential of H  z  is the Dirac function   z  . In this method,
Fig. 3. An eight-region segmentation using three level sets. we use one form of the regularization Heaviside functions.
Their definitions encountered in [32].
with the increase of regional complexity. On the contrary, with Furthermore, we introduce a combined vector
the increase of regional complexity, the fitting accuracy of the
, aNi  ,  i  1, 2, , L  indicating the ith combination
T
ai   a1i , a2i ,
log-normal model increases as well. Although the performance
of the IESE fitting the data is not the best with respect to the for all level set functions uk  k  1,2, , N  , where aki  0 or 1.
three different regions, respectively, it performs well at fitting Thus, Ai  x, y  are re-written as follows,
data for various types of images, indicating that the IESE is
N
more flexible modeling different regions than other specific Ai  x, y    i  uk ,  i  1, 2, , L (14)
models. Consequently, the IESE is more suited to multi-region k 1

where i  uk     k 
segmentation. H u , aki  0
. The differential of i  uk 
1  H  uk  ,
 aki  1
III. SEGMENTATION BASED ON THE MULTI-TEXTURE MODEL
is  i  uk     k 
 u , aki  0
WITH LEVEL SETS .
  uk  ,
 aki  1
Theoretically, same regions have the similarly statistical
properties in a SAR image. The objective of SAR image Here, R  x, y  in (5) denotes edge intensity (edge feature)
segmentation is to identify a finite number of statistically obtained by the modified ROEWA operator, Pi  x, y 
homogenous regions. To make full use of image information,
 i  1, 2, , L  in (8) denotes statistical distribution functions
edge feature and region information are both used for
segmentation. In this section, we combine the modified (region information) with respect to each region obtained by the
ROEWA operator and IESE to the level set framework. IESE. This model utilizing edge feature and region information
is called multi-texture model in this paper.
A. Defined Functional and Solutions Referring to the C-V model and the Region Competition (RC)
Supposing that   D2 is the domain, and an intensity SAR model [5, 49], the negative log likelihood of energy functional
image I :   D consists of L regions. The region  i E needs to be minimized for achieving segmentation.
Consequently, the energy functional E based on the
 i  1, , L  satisfies  j  , i, j  i  j  .
L
i 1
i   and i
multi-texture model is defined as follows,
Generally, level set function u  x, y  is the signed distance  N L
 L
E  uk ; Pi      ln i Pi ( x, y )  A i  x, y  dxdy
function [46, 47]. Thus, two regions can be represented by one  k 1 i 1  i 1

level set function. For example, N

u  x, y   0  x, y   1 ,  x, y    2       uk  R ( x, y )dxdy

 k 1 (15)
u  x, y   0  x, y   1 (12) N

u  x , y   0  x , y    ,  x , y         uk  uk dxdy
 1 2 k 1

where  x, y  denote pixel coordinates, 1 denotes the


N
1
 uk  1 dxdy .
  
2

2 
boundary of Ω1. k 1

Assuming that the number of level sets is N, the level set where α, β, γ, λ are the weighting parameters, i is the prior
functions are expressed by uk  k  1,2, , N  . Here we use the probability parameter, and  i  1 .  is the Hamiltonian
similar segmentation strategy in [6]. A level set is represented operator.
by a signed distance function, inside and outside of a level set In (15), the first term is to model regional information with
represent two different regions, thus, the relationship between respect to each region, Pi  x, y  is obtained by (8). The second
N and L is N  log 2  L  . Fig. 3 shows the example of an

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

term is to measure edge feature, R  x, y  is obtained by (5). In ith-region from (10) can be estimated by,
 I ( x, y)  (u )dxdy
j
addition to these two terms, we introduce another two terms to i k
the functional for improving the performance and effectiveness ˆ Ii ; j  
(19)
of the method. The third term is the boundary regularization to   (u )dxdy

i k
guarantee the smoothness of the curve. Although the third term
contains the same mean-curvature factor as in the second term, where i  1, 2, , L; j  1, 2,3; k  1, 2, N .
it is still necessary in the functional. Because divide boundaries Besides, based on the principle of the small time overhead,
here are determined by both edge information and region the termination criterion is defined as follows,
information. The fourth term is the penalty term. For the
 D u , uk , q   T
N

successful evolution, level sets must be kept as signed distance k , q 1 (20)


k 1
functions. Due to the inner effect of numerical calculation, the
where D  denotes the number of the pixels changing from a
procedure of the re-initialization of level sets in each iteration
therefore is required to preserve them as signed distance region to another region between two loops. The threshold T is
functions all the time. This procedure is expensive, complicated, set adaptively, we found that T  103 M is applicable to most of
and has slight side effect [50]. Consequently, we introduce a images.
penalty term proposed by Li to the energy functional [51]. This Considering that the small time step probably results in little
penalty term is able to force level set functions to be close to change between two loops, the program stops unless (20) is
signed distance functions adaptively and therefore avoids the satisfied in continuous five iterations, or the maximum iteration
need of the re-initialization procedure. is reached.
The solution of (15) obtained by minimizing the energy The algorithm process is as follows,
functional E in terms of the variation principle and gradient 1) Set the segment number L based on the prior knowledge.
descent flow theorem is as follows, 2) Set the parameters  ,  ,  ,  and t .
uk L  N  3) Calculate edge intensity by the modified ROEWA
   ln i Pi ( x, y )    i (u j )  i (uk ) operator by (5).
t i 1  j 1, j  k 
4) Initialize level sets uk , set uk  1 inside and uk  1
 u k 
  (uk )div  R ( x, y )  (16) outside  k  1, 2, , N  .
 uk 
 5) Estimate the PDF with respect to each region by (8). The
 u k    u k   necessary parameters are determined by (19).
 (uk )div     uk  div 
 u   u  
.
6) Update level set functions uk by (16).
 k    k 
where div   denotes the divergence,  is the Laplace 7) Check whether the termination condition is satisfied. If so,
go to the next step, otherwise go back to step 5).
operator, t is the time variable. When t   , the level set 8) Obtain the boundaries by setting uk  0 and implement
function uk keeps unchanged indicating that SAR image
the SAR image segmentation.
segmentation is implemented.
C. The Example for a Four-region segmentation Model
B. Numerical Computation and the Termination Condition To illustrate detailed formulas derived from (16) and (19)
Since it is difficult to solve (16) directly, the iteration method with a certain L and N, a four-region model with two level sets
is introduced. Hence, (16) can be solved by, is given as an example, which is applied to the following
uk , q 1  uk , q uk experiments. The level set functions derived from (16) are
 ,  k  1, 2, N  (17)
t t formulated as follows,
where t is the time step, q is the iterations. u1 ln 1P1  ln 3 P3  H  u2  
   u1   
 ln 2 P2  ln 4 P4   1  H  u2   
To improve the iterative stability, we use half implicit t
difference scheme instead of explicit difference scheme in our
study. The detailed procedures encountered in the literatures   u1   u1  
  u1    div  R    div 
 u  
(21)
[52, 53].  u 
In (16), i is estimated by,   1   
1 

Mi   u1  
,  i  1, 2, , L    u1  div 
ˆ i   u  
(18) .
M   1 
where M i is the number of the pixels within the ith region. M is
the pixel number of the image.
Additionally, re-estimation of the PDF within each region is
required during each iteration. Supposing that I  x, y  denotes a
SAR image, the sample jth-origin moment ˆ i I;j
for the

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7

ln 1P1  ln 2 P2   H  u1 
  Some parameters need to be set before segmentation. λ is the
u2 
   u2    weighting coefficient of the penalty term. A higher value helps
 ln 3 P3  ln 4 P4   1  H  u1   
t   to approximate a signed distance function better, whereas
  u 2   u 2  weakening the effect on image segmentation. The suggested
 
  u2    div  R    div    (22) value is  =0.04 [51]. t denotes the time step in the iteration.
 u

  2   u 2  A small time step is required for the computational stability,
  u 2   thereby we set t  0.1 . α, β, and γ are weighting coefficients.
  u2  div 
 u  
. To avoid under-segmentation or over-segmentation, the
  2  proportion of each parameter should not be too large. Based on
The sample jth-origin moment ˆ Ii ;3 for the ith-region some tested images, we found that the model performs good
 i, j  1,2,3,4  can be estimated by, with   2.0 ,   3.0 , and   2.0 for most of SAR images. In
the following experiments, we changed these three parameters
 I ( x, y ) H  u  H  u  j
1 2 slightly with respect to different images to attain better
ˆ  1 

 H u  H u  performances.
I; j
1 2
 Furthermore, in order to speed up the convergence rate of the
 I ( x, y) H  u  1  H u  
j
1 2
algorithm, initial level sets in all experiments were set as
multiple circles [55]. The red and yellow curves denote u1  0
ˆ  2 

 H  u  1  H  u  
I; j
1 2 and u2  0 , respectively. The segmentation accuracies were

(23)
 I ( x, y) 1  H  u   H u j evaluated by the confusion matrices. The evaluation parameters
1 2
include User Accuracy (UA), Producer Accuracy (PA), Overall
ˆ  3 

 1  H  u   H  u  Accuracy (OA), and kappa coefficient. The higher kappa


I; j
1 2
 coefficient implies the better segmentation.
 I ( x, y) 1  H  u   1  H u  
j
1 2
All the experimental images acquired by Radarsat-2 satellite
(C band) are intensity and single-look images without speckle
ˆ 
4 

 1  H  u   1  H  u  
I; j
1 2
pre-processing. The land covers in the images are common and
 typical, denoting either homogenous regions or heterogeneous /
extremely heterogeneous regions. The corresponding optical
IV. EXPERIMENTAL RESULTS AND DISCUSSION images acquired from google earth are used as the reference
The following experiments include three parts: the first part images. It is noted that the time lags of the image acquisition
is to test the segmentation performance of the proposed method between the optical images and the SAR images are very long,
against three representative methods. The second part is to many changes occurred during these periods. Therefore, by
explore the effect of each term in the proposed level set comparing the optical images with the SAR images, we select
framework on segmentation performance. The third part is to some areas from the unchanged areas as the validation data
investigate the segmentation performance of the proposed based on visual examination.
method with respect to various number of segmented regions. The segmentations are implemented with MATLAB R2014a
on Windows 8.1. The computer is with Intel Core i3-2120 CPU
A. The performance of the proposed method against other and 8GB of memory.
representative methods The first image mainly includes a river and the river bank.
Three real SAR images are used in this experiment. Based on The SAR image is of 280×280 pixels and was acquired on
visual interpretation, a four-region model with two level sets is August 22, 2012 with HH polarization at resolution of 5.2
considered. Some works have proven that SAR image m×7.7 m. The corresponding optical image was acquired on
segmentation using the methods based on level set can achieve March 9, 2010 from the google earth. We can see from Fig. 4
better results than that by other methods [54, 55]. Therefore, we (a)(b) that it is difficult to distinguish between the farmlands
chose three representative level set methods as the comparisons. and the urban areas from the SAR image by visual observation.
The first compared method is the classical C-V method [5], There are two probably reasons accounting for this fact. On one
which utilizes regional average information to implement hand, the interaction of the backscattering from the multiple
segmentation. The second and the third compared methods are types of land cover over one pixel results in difficulties in
the Weibull-model-based level set method and the differentiating small ground objects. On the other hand, we are
G0-model-based level set method [17, 19]. These two methods not sure what changes happened over this area during the
are based on the assumption that SAR image is completely period from March 9, 2010 to August 22, 2012. For example,
described by the Weibull distribution or the G0 distribution, and the farmland may be destroyed for expanding the urban area.
such distributions are then combined to the level set To provide a valid comparison, therefore, we consider that the
frameworks. The Weibull distribution is well suited to model observed image contains only two statistically homogenous
homogenous regions [20], and the G0 distribution is better able regions in terms of visual interpretation, which is water and
to fit various roughness regions compared with other models non-water areas. In Fig. 4 (a)(b), the validation data represented
[56]. by blue color denotes water areas, and the green areas denote

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8

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

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

(i) (j) (k) (l)


Fig. 4. (a) The optical image with the validation data. (b) The SAR image with the validation data. (c) The initial iteration. (d) The iterative process of the
proposed method. (e)(f) The final iteration and the segmentation results by C-V method, respectively. (g)(h) The final iteration and the segmentation results by
Weibull-model-based method, respectively. (i)(j) The final iteration and the segmentation results by the G0-model-based method, respectively. (k)(l) The final
iteration and the segmentation results by the proposed method, respectively.

Table 2. (a) The segmentation accuracy of the C-V method corresponded (b) The segmentation accuracy of the Weibull-model-based method
to Fig. 4(f). corresponded to Fig. 4(h).
Water Non-water Water Non-water
UA(%) 62.7 97.6 UA(%) 68.8 85.1
PA(%) 99.0 41.2 PA(%) 84.0 91.2
OA(%) 70.0 OA(%) 88.6
Kappa 0.623 Kappa 0.739

(c) The segmentation accuracy of the G0-model-based method (d) The segmentation accuracy of the proposed method corresponded to
corresponded to Fig. 4(j). Fig. 4(l).
Water Non-water Water Non-water
UA(%) 76.8 96.4 UA(%) 80.5 95.2
PA(%) 86.3 93.4 PA(%) 98.7 94.4
OA(%) 92.0 OA(%) 94.4
Kappa 0.762 Kappa 0.837

non-water areas. In (16), α, β, and γ are set to 2.0, 3.0, and 2.2, the Weibull model and the G0 model [13]. The segmentation
respectively. The running time for the proposed method, the results are shown in Fig. 4.
C-V method, the Weibull-based-model method, and the We can observe from Fig. 4 (k)(l) that the image is
G0-based-model method is 16.8s, 23.7s, 22.3s and 28.6s, segmented into two regions rather than four regions by the
respectively. The proposed method takes longer time than the proposed method. It indicates that the result is still feasible
three methods because four statistical parameters need to be when applying multiple level sets to the image where the
re-estimated in each iteration for the IESE yet only one number of actual statistically homogenous regions is less than
parameter for the C-V method and two statistical parameters for that of specified regions. Thus, the accurate estimation of the

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9

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

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

(i) (j) (k) (l)


Fig.5. (a) The optical image with the validation data. (b) The SAR image with the validation data. (c) The initial iteration. (d) The iterative process of the
proposed method. (e)(f) The final iteration and the segmentation results by C-V method, respectively. (g)(h) The final iteration and the segmentation results by
Weibull-model-based method, respectively. (i)(j) The final iteration and the segmentation results by the G0-model-based method, respectively. (k)(l) The final
iteration and the segmentation results by the proposed method, respectively.

Table 3. (a) The segmentation accuracy of the C-V method corresponded (b) The segmentation accuracy of the Weibull-based-model method
to Fig. 5(f). corresponded to Fig. 5(h).
Water Buildings Grassland Water Buildings Grassland
UA(%) 60.4 72.9 73.5 UA(%) 71.3 38.1 70.8
PA(%) 91.0 79.0 40.8 PA(%) 78.0 58.7 68.1
OA(%) 67.4 OA(%) 55.3
Kappa 0.483 Kappa 0.317

(c) The segmentation accuracy of the G0-model-based method (d) The segmentation accuracy of the proposed method corresponded to
corresponded to Fig. 5(j). Fig. 5(l).
Water Buildings Grassland Water Buildings Grassland
UA(%) 74.0 15.1 79.3 UA(%) 89.1 42.7 90.1
PA(%) 75.7 74.4 44.8 PA(%) 93.1 47.2 88.0
OA(%) 55.8 OA(%) 86.6
Kappa 0.329 Kappa 0.738

segmented region number is not necessary in SAR image The confusion matrices with respect to the four methods are
segmentation, thereby enhancing the usability of the method. given in Table 2. Table 2 demonstrates that the proposed
However, we still recommend to set a rough segment number method performs a higher accuracy than the other three
based on the prior knowledge for an efficient and precise methods. The proposed method exhibits the overall accuracy of
evolution. For the purposes of comparison, we regard the black 94.4% and the kappa coefficient of 0.837, higher than that of
areas as water areas and the rest of the areas as non-water areas the C-V method, Weibull-model-based method, and
in all segmentation results. G0-model-based method, which are 70.0% and 0.623, 88.6%

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10

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

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

(i) (j) (k) (l)


Fig.6. (a) The optical image with the validation data. (b) The SAR image with the validation data. (c) The initial iteration. (d) The iterative process of the
proposed method. (e)(f) The final iteration and the segmentation results by C-V method, respectively. (g)(h) The final iteration and the segmentation results by
Weibull-model-based method, respectively. (i)(j) The final iteration and the segmentation results by the G0-model-based method, respectively. (k)(l) The final
iteration and the segmentation results by the proposed method, respectively.

Table 4. (a) The segmentation accuracy of the C-V method corresponded (b) The segmentation accuracy of the Weibull-model-based method
to Fig. 6(f). corresponded to Fig. 6(f).
Runways Buildings Grassland Runways Buildings Grassland
UA(%) 54.2 50.0 89.8 UA(%) 85.8 32.1 93.9
PA(%) 93.4 71.0 35.6 PA(%) 90.7 99.0 12.4
OA(%) 60.1 OA(%) 53.2
Kappa 0.492 Kappa 0.448

(c) The segmentation accuracy of the G0-model-based method (d) The segmentation accuracy of the proposed method corresponded to
corresponded to Fig. 6(j). Fig. 6(l).
Runways Buildings Grassland Runways Buildings Grassland
UA(%) 94.5 32.9 93.9 UA(%) 95.1 81.6 93.7
PA(%) 93.3 79.3 73.4 PA(%) 93.6 76.9 94.9
OA(%) 82.2 OA(%) 93.1
Kappa 0.683 Kappa 0.876

and 0.739, 92.0% and 0.762, respectively. It is observed from segmentation results when applying the C-V method to SAR
Fig. 4(e)(f) that the image is segmented into many small pieces images. The similar results occur in the following experiments.
rather than multiply statistically homogenous areas. Because In Fig. 4(g)(h) and Fig. 4(i)(j), some areas are segmented
the C-V method using statistical average information is incorrectly by the Weibull-model-based method and the
designed for optical images with addictive noise. SAR images G0-model-based method, respectively. For instance, the area in
are polluted by multiplicative noise, which cannot be reduced the middle of the river is classified into two regions by both two
by regional mean processing. This fact leads to unsatisfactory methods, whereas from Fig. 4 (a)(b) we can see that the area has

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 11

the similarly statistical properties. Fig. 4 (k)(l) demonstrate that


the segmented result is achieved with correct boundary
localization and homogeneous areas.
The second image mainly includes a river, buildings, and
grassland. The SAR image is of 320×320 pixels and was
acquired on June 11, 2012 with VV polarization and 1.6 m ×
2.8m resolution. The corresponding optical image was acquired
on February 16, 2015 from google earth. In Fig. 5(a)(b), the
validation areas with blue color denote water areas, the green
areas denote grassland areas, and the red areas denote urban (a) (b)
areas. α, β, and γ are set to 2.0, 3.2, and 2.0, respectively. The
running time for the proposed method, the C-V method, the
Weibull-based-model method, and the G0-based-model method
is 21.8s, 37.4s, 31.2s and 46.4s, respectively. The segmentation
accuracies are described in Table 3.
Table 3 shows that both the Weibull-model-based method
and the G0-model-based method performance present much
lower performance in the tested example than the proposed
method performance. Compared with Table 3(b) and Table 3(c), (c) (d)
the overall accuracy from Table 3(d) increases by 31.3% and
30.8%, and the kappa coefficient increases by 0.421 and 0.409,
respectively. Since the result is determined only by the region
information by the Weibull-model-based method and the
G0-model-based method, the different areas presenting the
similarly statistical information will result in incorrect
segmentation. On the contrary, the proposed method combining
edge feature with region information attains a good
performance. For instance, in Fig. 5(k)(l), since the scattering
(e) (f)
mechanism of the pond is similar with that of the river, the pond Fig. 7. The final iteration and segmentation results by the proposed method
is segmented into the same regions of the river, whereas much without using (a)(b) region term, respectively, (c)(d) edge term,
grassland in Fig. 5(g)(h) and Fig. 5(i)(j) is segmented respectively, and (e)(f) boundary regularization term, respectively.
incorrectly into the region of the buildings.
proposed method, the C-V method, the Weibull-based-model
It should be noted that the proposed method utilizes texture
method, and the G0-based-model method is 33.8s, 63.1s, 55.7s
information to achieve segmentation, hence the same regions
and 73.5s, respectively. The segmentation accuracies are
that are of the similar texture properties in a SAR image are
described in Table 4.
identified as one of statistically homogenous regions. In other
We observe from Table 4 that the proposed method achieves
words, the different types of regions with the similar texture
a better segmentation result. The overall accuracy and the
information will be segmented into one same region. For
kappa coefficient by the proposed method are greater than
example, still water and road are two entirely different types of
39.9% and 0.428,10.9% and 0.193, respectively, compared
land covers, whereas they are of the similar texture properties
with that by the Weibull-model-based method and the
(e.g. low grey value and roughness) in radar images, thereby
G0-model-based method. Fig. 6 (g)(h) indicate that
resulting in the difficulties in distinguishing them without
Weibull-model-based cannot attain a good segmentation
knowing other information such as scattering mechanisms.
performance for this SAR image. Most of grassland is
It is also important to note that some changes took place
partitioned into building class. The probably reason is that the
during the period. Particularly, for Fig. 5(a)(b), many new
Weibull distribution has poor fitting capability for these two
buildings were built and some old buildings were dismantled in
land covers for this SAR image. In Fig. 6(i)(j), some areas are
the middle-upper images after the acquisition time of the SAR
segmented erroneously by the G0-model-based method. A part
image. Therefore, the accuracies for the four methods are not
of grassland between the runways is not segmented into the
absolute because we cannot guarantee that visual interpretation
grassland class. Furthermore, some segmented areas at the
is completely correct. However, it still provides a relative
top-left corner of the image do not coincide with the actuality.
analysis.
In Fig. 6(k)(l), different types of regions are segmented well.
The third image mainly includes grassland, buildings, and
However, due to the low image resolution, some small targets,
runways (roads). The SAR image is of 512×512 pixels and was
such as the small aircrafts at the left-bottom corner of the image,
acquired on July 29, 2012 with HH polarization and
cannot be segmented into a new region even if their texture
5.2m×7.7m resolution. The corresponding optical image was
features are quite different from others.
acquired on January 23, 2013 from the google earth. α, β, and γ
The above three experiments show that the method
are 2.3, 3.3, and 1.8, respectively. The running time for the

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 12

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

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

(m) (n)
Fig. 8. (a)(c)(e)(g)(i)(k)(m) are the simulated SAR images with 2-8 regions, respectively. (b)(d)(f)(h)(j)(l)(n) are their corresponding segmentation results,
respectively.

combining the modified ROEWA and the IESE performs best Fig. 7(a)(b) that without using region term the image is
compared with other three methods. It proves that the proposed segmented only into two regions since the divide curves are
method is able to cope with images containing various texture only decided by edge intensities. It demonstrates that using
regions. The experiments also indicate that the classical C-V only edge information cannot be applied to multi-region
method cannot be applied to SAR image segmentation due to segmentation. Fig. 7 (c)(d) indicates that the method without
multiplicative noise, and the G0-model-based method generally using edge information performs worse compared with the
can achieve better segmentation performances compared with method using all information. It is observed that the divide
the Weibull-model-based method, which verifies that the G0 curves cannot locate the correct boundaries of construction.
model is better suitable for the various heterogenous regions in This possibly can be attributed to the strong noise as well as the
a SAR image compared with other empirical distributions. imperfect fitting capability of the IESE. It demonstrates that the
It is noted that this method only uses texture information to use of edge information is able to enhance the segmentation
implement segmentation. It is practicable to discriminate the performance. Fig. 7 (e)(f) shows many incorrectly tiny
different targets presenting different texture information, such segmented areas in the image. The boundary regularization
as water, vegetation, and buildings, whereas it cannot be term is to guarantee the smoothness of the curve and avoid such
applied to those targets with similar texture information. For tiny segmented areas. If not using this term, divide curve easily
example, it is extremely difficult to distinguish between road moves to and gets in small areas which are completely caused
and water, or between short vegetation and farmland. In other by strong noise. The fourth term in equation (15) is the penalty
words, the method is suited to coarse segmentation, but not to term, which is used to keep level set functions always to be
fine segmentation. The method is suited for single polarization close to signed distance functions. This term simplifies the
SAR images segmentation. whole calculating procedures and significantly improves the
efficiencies. Without using this term, re-initialization of level
B. The effect of each term in the level set framework
set functions in each iteration is needed. We refer readers to [50]
We use the third experimental SAR image (Fig. 6(b)) to test for more information. Consequently, each term in the proposed
the effect of each term in the proposed level set framework level set framework has a positive influence on the
(equation (16)). The initial condition and the weighting segmentation performance.
parameter values are exactly the same as that in the above third
experiment. The effect of each term on segmentation C. The segmentation performance in terms of various number
performance can be tested by setting  ,  ,  related to each of segmented regions
term to 0 respectively. The segmented results without using The segmentation performance reduces sharply with the
region information, edge information, or boundary increasing number of segmented regions. We use simulated
regularization are shown in Fig. 7. The results are qualitatively SAR images with various regions to qualitatively test the
analyzed by visual examination. It is interesting to note from performance of the proposed method in terms of different
segmented number. Seven simulated SAR images of 500×500

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 13

pixels with two to eight regions are generated based on region information. Consequently, the segmentation is
different greyscale levels, and then the images are polluted by achieved by solving the level set functions obtained by
the multiplicative noise modelled by the uniformly distributed minimizing the energy functional in terms of variation principle
random noise with zero-mean and 0.03-variance. For a better and gradient descent flow theorem.
visual examination, each region in the segmentation results is The experimental results show that the proposed method is
represented by a random color. The simulated SAR images and able to cope with a SAR image composed of different types of
the corresponding segmentation results are shown in Fig. 8. texture regions. Compared with the C-V method, the
Fig. 8 shows that the proposed method for two-region Weibull-model-based method, and the G0-model-based method,
segmentation using one level set performs excellent. The the proposed method attains the higher segmentation accuracies
three-region and four-region segmentation results using two with respect to the three tested examples. The experimental
level sets are also satisfactory. There are slight noises around images can be segmented into multiple regions with correct
edges between regions in these two segmentation results. When boundary localization and statistically homogeneous areas by
using three level sets to implement segmentation, the the proposed method. Besides, it demonstrates that each term in
performance of the proposed method decreases as the number the proposed level set framework can improve either the
of regions increases due to the interaction of level set functions. segmentation performance or the effectiveness of the method.
For five-region segmentation in Fig. 8(h), boundary location is Furthermore, the experiment indicates that this method is
not precise in the intersections of multiply regions. Edge noise considered can attain acceptable segmentation performances
occurs more in six-region and seven-region segmentation when applying it to a SAR image composed of less than eight
results. Besides, for seven-region segmentation, it can be seen regions. Consequently, the method is suited to coarse
from Fig. 8(l) that some pixels in the homogenous regions are segmentation, but not to fine segmentation. The proposed
incorrectly segmented into other regions. The worst result is method provides a feasible coarse segmentation method for
eight-region segmentation. In addition to abovementioned single polarization SAR images in the absence of
problems, the proposed method segments Fig. 8(m) into seven multi-polarimetric SAR images.
regions rather than eight regions as shown in Fig. 8(n). It is Further work will focus on the adaptivity of the weighted
observed that the two regions in Fig. 8(m) corresponded to two parameters, which has an effect on the segmentation results.
“red” regions in Fig. 8(n) are similar, although they are of
totally different grey levels before being polluted by noise. REFERENCES
These two regions have approximately statistical information [1] M. Kass, A. Witkin, and D. Terzopoulos, "Snakes: Active contour
and are thus difficult to be distinguished by the statistical term models," International journal of computer vision, vol. 1, pp.
321-331, 1988.
in the proposed level set framework due to the imperfect fitting [2] R. Malladi, J. A. Sethian, and B. C. Vemuri, "Shape modeling with
ability of the IESE. The segmentation result from Fig. 8(n) front propagation: A level set approach," Pattern Analysis and
indicates that with the increasing number of segmented regions, Machine Intelligence, IEEE Transactions on, vol. 17, pp. 158-175,
a model with much better fitting ability is required in order to 1995.
[3] A. Yezzi Jr, S. Kichenassamy, A. Kumar, P. Olver, and A.
achieve a satisfactory segmentation. Based on the analysis, we Tannenbaum, "A geometric snake model for segmentation of
can infer that the proposed method can perform good using one medical imagery," Medical Imaging, IEEE Transactions on, vol. 16,
or two level sets. It also can be used to a SAR image with less pp. 199-209, 1997.
[4] V. Caselles, R. Kimmel, and G. Sapiro, "Geodesic active contours,"
than eight regions, however, when using three level sets the International journal of computer vision, vol. 22, pp. 61-79, 1997.
segmentation performance is dependent on the image [5] T. F. Chan and L. A. Vese, "Active contours without edges," Image
characteristics. Although the proposed method can be Processing, IEEE Transactions on, vol. 10, pp. 266-277, 2001.
[6] L. A. Vese and T. F. Chan, "A multiphase level set framework for
generalized for arbitrary number of regions segmentation, the image segmentation using the Mumford and Shah model,"
segmentation result probably is not satisfactory when using International journal of computer vision, vol. 50, pp. 271-293,
four level sets or more. 2002.
[7] T. Brox and J. Weickert, "Level set segmentation with multiple
regions," Image Processing, IEEE Transactions on, vol. 15, pp.
V. CONCLUSION 3213-3218, 2006.
[8] B. Sandberg, S. H. Kang, and T. F. Chan, "Unsupervised feature
In this paper, we present a multi-region segmentation method balancing multiphase segmentation," UCLA CAM report, pp. 08-02,
for SAR images based on the multi-texture model with level 2008.
sets. For the purpose of multi-region segmentation, we derive [9] G. Gao, "Statistical modeling of SAR images: A survey," Sensors,
vol. 10, pp. 775-795, 2010.
the energy functional that arises from the multi-texture model
[10] C. J. Oliver, A. Blake, and R. G. White, OPTIMUM TEXTURE
which encompasses edge feature and region information. This ANALYSIS OF SYNTHETIC-APERTURE RADAR IMAGES vol.
energy functional has the general form and can be generalized 2230. Bellingham: Spie - Int Soc Optical Engineering, 1994.
for arbitrary number of regions segmentation. The ROEWA [11] F. Ulaby and M. Dobson, "Handbook of radar statistics for terrain,"
Artech House, Norwood, MA, 1989.
operator based on the multi-edge model is effective and [12] M. Sekine and Y. Mao, Weibull radar clutter: IET, 1990.
practical to extract edge feature for SAR images. For the need [13] C. C. Freitas, A. C. Frery, and A. H. Correia, "The polarimetric G
of level set evolution, we modify the ROEWA operator to distribution for SAR data analysis," Environmetrics, vol. 16, pp.
13-31, Feb 2005.
obtain edge feature. The IESE, which is able to give any SAR [14] M. Horritt, "A statistical active contour model for SAR image
image distribution an appropriately statistical expression with segmentation," Image and Vision Computing, vol. 17, pp. 213-224,
respect to various types of regions, is developed to obtain 1999.

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806201, IEEE
Transactions on Image Processing
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 14

[15] Y. Shuai, H. Sun, and G. Xu, "SAR image segmentation based on with an active contour approach," Image Processing, IEEE
level set with stationary global minimum," Geoscience and Remote Transactions on, vol. 10, pp. 72-78, 2001.
Sensing Letters, IEEE, vol. 5, pp. 644-648, 2008. [35] R. Fjortoft, A. Lop, P. Marthon, and E. Cubero-Castan, "Different
[16] M. Silveira and S. Heleno, "Separation Between Water and Land in approaches to multiedge detection in SAR images," in Geoscience
SAR Images Using Region-Based Level Sets," Geoscience and and Remote Sensing, 1997. IGARSS'97. Remote Sensing-A
Remote Sensing Letters, IEEE, vol. 6, pp. 471-475, 2009. Scientific Vision for Sustainable Development., 1997 IEEE
[17] R. C. P. Marques, F. Medeiros, x00E, N. tima, and J. Santos Nobre, International, 1997, pp. 2060-2062.
"SAR Image Segmentation Based on Level Set Approach and G [36] L. D. Cohen, "On active contour models and balloons," CVGIP:
Model," Pattern Analysis and Machine Intelligence, IEEE Image understanding, vol. 53, pp. 211-218, 1991.
Transactions on, vol. 34, pp. 2046-2057, 2012. [37] C. Xu and J. L. Prince, "Snakes, shapes, and gradient vector flow,"
[18] F. Galland, J. M. Nicolas, H. Sportouche, M. Roche, F. Tupin, and P. Image Processing, IEEE Transactions on, vol. 7, pp. 359-369, 1998.
Refregier, "Unsupervised Synthetic Aperture Radar Image [38] A. A. Amini, T. E. Weymouth, and R. C. Jain, "Using dynamic
Segmentation Using Fisher Distributions," Geoscience and Remote programming for solving variational problems in vision," Pattern
Sensing, IEEE Transactions on, vol. 47, pp. 2966-2972, 2009. Analysis and Machine Intelligence, IEEE Transactions on, vol. 12,
[19] I. B. Ayed, N. Hennane, and A. Mitiche, "Unsupervised Variational pp. 855-867, 1990.
Image Segmentation/Classification Using a Weibull Observation [39] A. Papoulis and S. U. Pillai, Probability, random variables, and
Model," Image Processing, IEEE Transactions on, vol. 15, pp. stochastic processes: Tata McGraw-Hill Education, 2002.
3431-3439, 2006. [40] A. Stuard and J. Ord, "Kendall’s advanced theory of statistics," ed:
[20] C. Tison, J. M. Nicolas, F. Tupin, and H. Maitre, "A new statistical Griffin London, 1987.
model for Markovian classification of urban areas in high-resolution [41] J. W. Goodman, "Statistical properties of laser speckle patterns," in
SAR images," Geoscience and Remote Sensing, IEEE Transactions Laser speckle and related phenomena, ed: Springer, 1975, pp. 9-75.
on, vol. 42, pp. 2046-2057, 2004. [42] C. L. Nikias, "Higher-order spectral analysis," in Engineering in
[21] Y. Huang and Y.-C. Huang, "Segmenting SAR satellite images with Medicine and Biology Society, 1993. Proceedings of the 15th
the multilayer level set approach," Selected Topics in Applied Earth Annual International Conference of the IEEE, 1993, pp. 319-319.
Observations and Remote Sensing, IEEE Journal of, vol. 4, pp. [43] C. Oliver and S. Quegan, Understanding synthetic aperture radar
632-642, 2011. images: SciTech Publishing, 2004.
[22] G. Akbarizadeh, "A new statistical-based kurtosis wavelet energy [44] S. Kullback, "Letter to the editor: the Kullback-Leibler distance,"
feature for texture recognition of SAR images," Geoscience and 1987.
Remote Sensing, IEEE Transactions on, vol. 50, pp. 4358-4368, [45] D. H. Johnson and S. Sinanovic, "Symmetrizing the kullback-leibler
2012. distance," 2003.
[23] L. Hualin, W. Wei, Z. Guixu, and P. Ling, "A Variational Approach [46] D. Peng, B. Merriman, S. Osher, H. Zhao, and M. Kang, "A
for Bias Correction and Boundary Delineation of SAR Image with PDE-based fast local level set method," Journal of Computational
Intensity Inhomogeneity," in Image and Signal Processing, 2009. Physics, vol. 155, pp. 410-438, 1999.
CISP '09. 2nd International Congress on, 2009, pp. 1-5. [47] M. Sussman, A. S. Almgren, J. B. Bell, P. Colella, L. H. Howell,
[24] I. Ben Ayed, A. Mitiche, and Z. Belhadj, "Multiregion level-set and M. L. Welcome, "An adaptive level set approach for
partitioning of synthetic aperture radar images," Pattern Analysis incompressible two-phase flows," Journal of Computational
and Machine Intelligence, IEEE Transactions on, vol. 27, pp. Physics, vol. 148, pp. 81-124, 1999.
793-800, 2005. [48] H.-K. Zhao, T. Chan, B. Merriman, and S. Osher, "A variational
[25] J. L. Feng, Z. J. Cao, and Y. M. Pi, "Multiphase SAR Image level set approach to multiphase motion," Journal of computational
Segmentation With G(0)-Statistical-Model-Based Active physics, vol. 127, pp. 179-195, 1996.
Contours," Ieee Transactions on Geoscience and Remote Sensing, [49] S. C. Zhu and A. Yuille, "Region competition: Unifying snakes,
vol. 51, pp. 4190-4199, Jul 2013. region growing, and Bayes/MDL for multiband image
[26] I. Ben Ayed, A. Mitiche, and Z. Belhadj, "Polarimetric image segmentation," Pattern Analysis and Machine Intelligence, IEEE
segmentation via maximum-likelihood approximation and efficient Transactions on, vol. 18, pp. 884-900, 1996.
multiphase level-sets," Pattern Analysis and Machine Intelligence, [50] J. Gomes and O. Faugeras, "Reconciling distance functions and
IEEE Transactions on, vol. 28, pp. 1493-1500, 2006. level sets," Journal of Visual Communication and Image
[27] R. Fjortoft, A. Lopes, P. Marthon, and E. Cubero-Castan, "An Representation, vol. 11, pp. 209-223, 2000.
optimal multiedge detector for SAR image segmentation," [51] C. Li, C. Xu, C. Gui, and M. D. Fox, "Level set evolution without
Geoscience and Remote Sensing, IEEE Transactions on, vol. 36, pp. re-initialization: a new variational formulation," in Computer Vision
793-802, 1998. and Pattern Recognition, 2005. CVPR 2005. IEEE Computer
[28] J. Inglada and G. Mercier, "A new statistical similarity measure for Society Conference on, 2005, pp. 430-436.
change detection in multitemporal SAR images and its extension to [52] D. Mumford and J. Shah, "Optimal approximations by piecewise
multiscale change analysis," Geoscience and Remote Sensing, IEEE smooth functions and associated variational problems,"
Transactions on, vol. 45, pp. 1432-1445, 2007. Communications on pure and applied mathematics, vol. 42, pp.
[29] J. Zheng and H. You, "A New Model-Independent Method for 577-685, 1989.
Change Detection in Multitemporal SAR Images Based on Radon [53] C. Samson, L. Blanc-Féraud, G. Aubert, and J. Zerubia, "A level set
Transform and Jeffrey Divergence," Geoscience and Remote model for image classification," International journal of computer
Sensing Letters, IEEE, vol. 10, pp. 91-95, 2013. vision, vol. 40, pp. 187-197, 2000.
[30] F. Bovolo and L. Bruzzone, "An Adaptive Technique based on [54] K. Karantzalos and D. Argialas, "Automatic detection and tracking
Similarity Measures for Change Detection in Very High Resolution of oil spills in SAR imagery with level set segmentation,"
SAR Images," in Geoscience and Remote Sensing Symposium, 2008. International Journal of Remote Sensing, vol. 29, pp. 6281-6296,
IGARSS 2008. IEEE International, 2008, pp. III - 158-III - 161. 2008.
[31] Z. Shi and K. B. Fung, "A comparison of digital speckle filters," in [55] R. C. P. Marques, F. N. S. de Medeiros, and D. Ushizima, "Target
Geoscience and Remote Sensing Symposium, 1994. IGARSS'94. Detection in SAR Images Based on a Level Set Approach," Systems,
Surface and Atmospheric Remote Sensing: Technologies, Data Man, and Cybernetics, Part C: Applications and Reviews, IEEE
Analysis and Interpretation., International, 1994, pp. 2129-2133. Transactions on, vol. 39, pp. 214-222, 2009.
[32] S. Luo, L. Tong, Y. Chen, and M. Jia, "SAR image segmentation [56] L. Bombrun, G. Vasile, M. Gay, and F. Totir, "Hierarchical
based on the advanced level set," in IOP Conference Series: Earth segmentation of polarimetric SAR images using heterogeneous
and Environmental Science, 2014, p. 012246. clutter models," IEEE Transactions on Geoscience and Remote
[33] R. Touzi, A. Lopes, and P. Bousquet, "A statistical and geometrical Sensing, vol. 49, pp. 726-737, 2011.
edge detector for SAR images," Geoscience and Remote Sensing,
IEEE Transactions on, vol. 26, pp. 764-773, 1988.
[34] O. Germain and P. Réfrégier, "Edge location in SAR images:
Performance of the likelihood ratio filter and accuracy improvement

1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like