A Level Set SAR Image Segmentation Approach Based on Fisher
Distribution
1
Xun Wang,
2
Yinghao Fan,
3
Dingke Kong
1, 2
School of Computer and Information Engineering, Zhejiang Gongshang University,
wx@mail.zjgsu.edu.cn; fanyinghao0705@126.com
*3Corresponding Author
School of Computer and Information Engineering, Zhejiang
Gongshang University, kdk@zjgsu.edu.cn
Abstract
Traditional level set approaches for SAR image segmentation are generally based on gamma
distribution assumptions. Although those gamma distribution based approaches can work well in the
case of low-resolution SAR images, they are not obviously suitable for the high-resolution SAR images,
especially with textures or strong reflectors. Aiming at overcoming this limitation, a novel level set
approach for SAR image segmentation is proposed. This approach is based on Fisher probability
distribution functions (pdfs) of the intensity fluctuations with an efficient parameter estimation strategy.
Fisher distribution is general enough to encompass most of the existing models. In order to take into
account the entire diversity of the scenes, Fisher pdfs is chosen. And the energy functional can be
efficiently minimized by alternative iteration between the Euler-Lagrange descent equations of curve
evolution and a gradient descent update of the parameters of Fisher distribution. Experimental results
show that the proposed method is much better adapted to SAR images especially with textures or
strong reflectors compared with gamma pdfs.
Keywords: Synthetic Aperture Radar (SAR), Image Segmentation, Fisher Distribution, Level Set,
Parameter Estimation
1. Introduction
Synthetic aperture radar (SAR) has been widely used for many years in remote sensing applications
[1]
. Such imaging systems have many advantages in comparison to more standard optical imaging
systems. They can operate any time of days under any weather conditions and offer a high spatial
resolution. In spite of these advantages, SAR images are affected by the coherent speckles, so the
segmentation of SAR images is generally acknowledged as a challenging task compared to optical
image processing
[2]
.
Recently, active contour model (ACM) based on curve evolution and level set method
[3, 4]
are
received widely attention to address SAR image segmentation
[5-9]
. There are several desirable
advantages of geometric ACM over classical image segmentation methods, such as histogram
thresholding
[10]
, edge detection, and region growing
[11]
. At first, geometric ACM can achieve sub-
pixel accuracy of object boundaries
[12]
. It can adaptively deal with topological changes and provide
high-precision closed partition curve, which are necessary and used for further applications, such as
shape analysis and recognition
[13]
. In addition, ACM can also manage to define energy function
without speckle noise reduction and get accurate segmentation results.
Besides, statistical approaches are also generally required. Second-kind statistics (or Log-statistics)
have been used and are particularly appropriate to deal with such images, such as K, Weibull or
Rayleigh distribution
[14, 15]
. And the most of existing level set methods for SAR image segmentation
are based on gamma distribution
[16, 17]
. The gamma distribution can be extensively applied in the case
of low-resolution SAR images. However, the gamma model is not obviously suitable for high-
resolution SAR images, since it is weak in strong reflections or textures and its segmentation results
have a certain redundancy
[18]
. Therefore, none of them provide a good solution for high-resolution
SAR images, mainly because they do not have a sufficiently large field of applications. In order to take
into account the entire diversity of the scenes, we propose to choose a class of models that is general
enough to encompass most of the existing models. Therefore, Fisher probability distribution function
(pdf) is chosen as the statistical model, whose feature is more adapted to high-resolution SAR images
especially with textures or strong reflectors as well as low-resolution SAR images.
A Level Set SAR Image Segmentation Approach Based on Fisher Distribution
Xun Wang, Yinghao Fan, Dingke Kong
Journal of Convergence Information Technology(JCIT)
Volume 7, Number 21, Nov 2012
doi : 10.4156/jcit.vol7.issue21.32