A robust patch-statistical active contour model for image segmentation
Qi Ge
a
, Liang Xiao
a,1
, Jun Zhang
b
, Zhi Hui Wei
a,
⇑
a
School of Computer Science, Nanjing University of Science & Technology, Nanjing, Jiangsu 210094, China
b
School of Science, Nanjing University of Science & Technology, Nanjing, Jiangsu 210094, China
article info
Article history:
Received 4 October 2011
Available online 28 April 2012
Communicated by A. Fernandez-Caballero
Keywords:
Active contour model
Image segmentation
Weighting-estimated statistical information
Intensity variation penalization term
Gradient information
abstract
This paper proposes a novel region-based active contour model (ACM) for image segmentation, which is
robust to noise and intensity non-uniformity. The energy functional of the proposed model consists of
three terms, i.e., the patch-statistical region fitting term, the improved regularization term, and the inten-
sity variation penalization term. The patch-statistical region fitting term computes the local statistical
information in each patch as the basis for driving the curve accurately with resist to intensity non-
uniformity and weak boundaries. And the regularization term coupling with the gradient information
improves the ability of capturing the boundaries with cusps and narrow topology structures. Further-
more, an intensity variation penalization term is proposed to make sure that the segmentation result
is robust to the irregular intensity variation. Experiments on medical and natural images show that
the proposed model is more robust than the popular active contour models for image segmentation with
noise and intensity non-uniformity.
Ó 2012 Elsevier B.V. All rights reserved.
1. Introduction
Active contour model (ACM) is one of the most successful meth-
ods for image segmentation, which plays an important role in the
computer vision. The basic idea is to drive a contour under some
constraints to the desired object. According to the nature of con-
straints, the existing ACMs can be categorized into two groups:
the edge-based models (Kass et al., 1991; Caselles et al., 1997;
Kichenesamy et al., 1996; Mishra et al., 2011) and the region-based
models (Chan and Vese, 2001; Chan et al., 2006; Bresson et al.,
2007; Qin and Clausi, 2010; Ni et al., 2009; Aubert et al., 2003;
Jehan-Besson et al., 2003; Herbulot et al., 2006; Lecellier et al.,
2006; Meziou et al., 2011). The latter ones are less sensitive to
noise, as well as the location of the initial contour, and have better
performance with weak boundaries. Hence these models can effi-
ciently detect the exterior and interior boundaries simultaneously.
There are several successful region-based models such as the
Mumford–Shah functional (Mumford and Shah, 1989) and the
Chan–Vese (CV) model (Chan and Vese, 2001). The CV model,
implemented by the level set method (Osher and Sethian, 1988),
has been successfully applied in binary phase segmentation under
the assumption that each image region is statistically homoge-
neous. Vese and Chan extended their work (Vese and Chan,
2002) to represent multiple regions utilizing multiphase level set
functions. These models are referred to as piecewise constant
(PC) models. However, both the CV and the PC models often lead
to poor segmentation results for images with intensity non-unifor-
mity due to their underlying assumption that the intensities in
each region always keep constant.
In order to segment images with smooth intensity non-unifor-
mity, the local statistical information of intensity is widely em-
ployed in the region-based active contour models to approximate
the images (Yezzi et al., 1999; Li et al., 2008a,b; Zhang et al.,
2010; Brox and Cremers, 2010; Aubert et al., 2003; Ni et al.,
2009). Li et al. (2008a,b) proposed a local binary fitting (LBF) mod-
el, which embedded the local statistical information of intensity
into a region-based active contour model by a Gaussian kernel
with a scale parameter. The LBF model superiorly segments the
images with spatially smooth intensity non-uniformity. However,
with a too small scale parameter, the model would output undesir-
able contours due to noise and unsmooth intensity non-unifor-
mity. Moreover a large scale parameter induces very few of
undesired contours, it may be computationally expensive and lead
to inaccurate segmentation results. The reason is that these models
extract the local statistical information by assuming that the distri-
bution of intensity is uniform. Nevertheless, the intensities of
images are not necessarily described by one specific distribution
because the intensity non-uniformity also varies spatially. Hence
it will be inaccurate to partition each region by the statistical infor-
mation of intensity using a fixed-scale estimation method. In order
to solve this problem, the region-based contour models were pro-
posed in recent literature (Aubert et al., 2003; Jehan-Besson et al.,
2003; Ni et al., 2009), by using the histogram of the intensity in a
0167-8655/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.patrec.2012.03.029
⇑
Corresponding author. Tel.: +86 13512510326.
E-mail addresses: carolgestone@yahoo.com.cn (Q. Ge), xtxiaoliang@163.com
(L. Xiao), phil_zj@163.com (J. Zhang), gswei@mail.njust.edu.cn (Z.H. Wei).
1
Tel.: +86 02584318108.
Pattern Recognition Letters 33 (2012) 1549–1557
Contents lists available at SciVerse ScienceDirect
Pattern Recognition Letters
journal homepage: www.elsevier.com/locate/patrec