IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 9, NO. 6, NOVEMBER 2012 1041
MPM SAR Image Segmentation Using Feature
Extraction and Context Model
Biao Hou, Member, IEEE, Xiangrong Zhang, Member, IEEE,andNanLi
Abstract—A new synthetic aperture radar (SAR) image seg-
mentation method based on a maximization of posterior marginals
(MPM) algorithm with feature extraction and context model is
proposed in this letter. First, Gabor wavelet and texture descriptor
are used to extract features, which enhance intraclass similarities
and interclass differences. Second, the number of regions within
the same class is reduced in order to improve the reliability of
the regional statistical characteristics. Finally, the MPM of each
region combined with the context model is calculated by consider-
ing both the intralayer correlation and interlayer correlation. The
experimental results show that the proposed method is efficient
and effective for SAR image segmentation.
Index Terms—Hierarchical Markov random field model, maxi-
mization of posterior marginals (MPM), SAR image segmentation,
synthetic aperture radar (SAR), watershed segmentation.
I. INTRODUCTION
S
YNTHETIC aperture radar (SAR) image segmentation is
a classic problem of SAR image understanding and inter-
pretation. Due to the characteristic of image formation process
and content, a SAR image is extremely speckled, which makes
SAR image segmentation a challenging problem [1]. Markov
random field (MRF) models have been extensively used in
SAR image segmentation [2]–[5]. These segmentation methods
can be mainly divided into two categories: transform-domain
methods and spatial-domain methods.
Owing to the advantage of transform-domain tools, hidden
Markov tree (HMT) models in the transform domain have
been widely applied to image segmentation. Choi and Baraniuk
[6] proposed a wavelet-domain HMT segmentation method.
However, wavelets are limited to capture direction informa-
tion and represent images with smooth contours, which easily
lead to directional edge vagueness and singularity diffusion in
segmentation results. To overcome the directional vagueness,
Wu et al. [7] proposed a contourlet-domain HMT method which
can efficiently capture the smooth contours using directional
filter banks. Hence, the segmentation results contain more
accurate edge than wavelet. A segmentation method using
complex wavelet transform and HMT was proposed in [8],
which made full use of the orientation selection of complex
wavelet transform and improved the segmentation result.
Manuscript received June 10, 2011; revised January 29, 2012; accepted
February 21, 2012. This work was supported in part by the National Natural
Science Foundation of China under Grants 60971128 and 60803097 and in part
by the Program for New Century Excellent Talents in University under Grant
NCET-10-0666.
The authors are with the Key Laboratory of Intelligent Perception and Image
Understanding of the Ministry of Education of China, Xidian University, Xi’an
710071, China (e-mail: avcodec@hotmail.com).
Digital Object Identifier 10.1109/LGRS.2012.2189352
In the spatial-domain methods, the traditional noncausal
MRF models [9]–[11] are computationally demanding, and
segmentation results are often unsatisfactory in homogeneous
regions and exact edges. Therefore, the causal hierarchical
models [12] have been proposed for segmentation of SAR
images. In order to reduce the computational complexity, a
discrete Markov model was proposed in [13]. However, block
artifacts and discontinuous edges still existed i n segmentation
results because intralayer correlations were not considered.
Moreover, a region-based hierarchical MRF model was pro-
posed in [14], which can efficiently reduce the block artifacts
but still contained some drawbacks.
1) The capability of features in characterizing images was
poor because only gray information was taken into
account.
2) The regions in this model were produced by the tra-
ditional watershed method [15]. As is well known, the
watershed algorithm usually oversegments an image into
large numbers of small regions, and these regions are easy
to be misclassified.
3) Only the interlayer interactions were considered in the
maximization of posterior marginals (MPM) algorithm
[16], and the interactions of regions in a layer were
not taken into account; consequently, the accuracy of
segmentation results decreased.
To solve the aforementioned problems, a new region-based
MPM algorithm using feature extraction and context model is
proposed in this letter. Gabor wavelet and texture descriptor
are used to extract features. An improved watershed method
based on gradient modification and region merging is used to
reduce the number of small regions, which also ensures that
terrain edges are complete and clear. In the proposed algorithm,
both the intralayer correlation and the interlayer correlation
are considered, and then, the MPM of each region combined
with the context model is calculated, improving the accuracy of
segmentation.
This letter i s organized as follows. In Section II, the region-
based hierarchical MRF model used for observation field and
label field is described. Section III describes the proposed seg-
mentation method based on the region-based MPM estimation
using feature extraction and context model. The experimental
results are illustrated in Section IV, and Section V is the
conclusion.
II. R
EGION-BASED HIERARCHICAL MRF MODEL
The noncausal MRF model [9] demonstrated global behavior
using local interactions at a scale of a single pixel and its nearest
neighbors. The fundamental limitation of the noncausal MRF
1545-598X/$31.00 © 2012 IEEE