A histogram-based Chan-Vese model driven by local
contrast pattern for texture image segmentation
Haiying Tian
School of Information Science
and Technology
Sun Yat-sen University
Guangzhou, China
Email: tian882006@126.com
Yanfei Liu
School of Information Science
and Technology
Sun Yat-sen University
Guangzhou, China
Email: farelyf@163.com
Jian-huang Lai
∗
School of Information Science
and Technology
Sun Yat-sen University
Corresponding Author
Email: stsljh@mail.sysu.edu.cn
Abstract—This paper proposes a novel local contrast pattern
(LCP) to drive the histogram-based Chan-Vese (CV) model for
texture image segmentation. The local contrast pattern has two
maps, differential contrast map and orientation map, which are well
suited to describe texture structure, especially the texture orientation
information. In order to enable the extraction of accurate local texture
features, a truncated Gaussian kernel function is also incorporated
into the improved model. Then, a novel histogram-based CV model
is guided by the LCP feature maps and a truncated Gaussian kernel to
obtain the texture segmentation. Moreover, we verify the robustness
for illumination, noise and initialization of the proposed model in
level set framework. Experiments and comparisons demonstrate that
the proposed model is effective on various types of image for texture
segmentation.
I. INTRODUCTION
Active contour models (ACMs) have been extensively
applied in image segmentation. According to the type of
extracted image information, active contour models can be
roughly categorized into two classes: edge-based models [1]–
[5] and region-based models [6]–[11]. Edge-based ACMs use
the image edgemap as the stopping function. That is to say,
active contour models follow the high gradient information
to evolve the curve, which is ineffective when the contrast
between foreground and background is low. Region-based
ACMs utilize the region information to segment image based
on the different statistics of intensity or texture in regions.
Therefore, they are more robust to noise and object without
edge.
Over the last decades, many texture analysis methods have
been developed to segment texture image [12]–[16]. However,
texture image segmentation is still a thorny task, and many
researchers tend to find a good feature statistic or feature de-
scriptor to achieve the segmentation of texture image. Rousson
et al. [10] proposed to use the Gaussian distribution to repre-
sent the probability densities of different regions, and then the
segmentation results were obtained by maximizing a posteriori
probability (MAP) principle. Inspired by this work, Li Wang et
al. [11] introduced a Gaussian kernel function into the energy
function to localize the feature statistic, and the improved
method could well partition intensity inhomogeneity image and
texture image. Recently, the histogram has become a powerful
tool to describe image features for its rotation invariance
and approximately homogeneity within a local region. Some
similar models based on histogram statistic were presented
in [17]–[20]. In [20], a histogram-based statistic appearance
model was estimated, which was trained to segment the human
left kidney in CT image. In [16], the authors incorporated
histogram statistic into the segmentation model by using the
Wasserstein distance to quantify the similarity between the
object histogram and the average histogram to obtain a good
segmentation result.
In this paper, a histogram-based Chan-Vese (CV) model
driven by local contrast pattern and a truncated Gaussian kernel
function is proposed. There are two main contributions. Firstly,
the LCP descriptor is proposed to suppress undesirable clusters
and describe image feature correctly. It can not only eliminate
inhomogeneous intensity in an intra-texture but also distinguish
different texture types. Secondly, we incorporate the LCP
descriptor into a histogram-based CV model to segment texture
image. In order to measure texture information effectively, a
truncated Gaussian kernel function is used to localize texture
information in the improved method. Experiments demonstrate
that the proposed scheme can effectively achieve segmentation
and has a good robustness for illumination, noise and curve
initialization.
The remainder of the paper is organized as follows. Sec-
tion II reviews some relative works based on ACMs and texture
statistics. Section III describes our method. Firstly, a local
contrast pattern is proposed, then a local histogram-based CV
model is reformulated by the LCP descriptor and a truncated
Gaussian kernel function. The experiments are demonstrated
in Section IV, and Section V summarizes the paper.
II. R
ELATED WORKS
A. Local Histogram-based (LH) Model
The local histogram-based active contour model [18] can
be applied to segment texture image. Essentially, it is a defor-
mation process of evolving curve by minimizing the distance
between the object histogram and the average histogram in
the CV model, and it is able to well segment texture image
with noise and complex intensity variation compared with the
traditional CV model.
Given a gray-scale image I :Ω→ [0,L], which has two
regions to be partitioned, Ω is the image domain and the
evolving contour C divides I into two areas denoted by in(C)
and out(C). Let N
x,r
be a local patch centered at a pixel
2014 22nd International Conference on Pattern Recognition
1051-4651/14 $31.00 © 2014 IEEE
DOI 10.1109/ICPR.2014.174
954