SAR图像双Markov-EAR模型的纹理无监督分割  

An Unsupervised and Higher-Precision Segmentation Method for Textured SAR Images Based on Pairwise Markov-EAR Model

作　　者：丁明涛[1] 田铮[1] 句彦伟[1] 

机构地区：[1]西北工业大学理学院应用数学系,陕西西安710072

出　　处：《西北工业大学学报》2006年第6期736-740,共5页Journal of Northwestern Polytechnical University

基　　金：国家自然科学基金(60375003);航空基础科学基金(03I53059)资助

摘　　要：单视SAR图像保留了最大的分辨率和场景可观测的全部纹理信息,根据单视SAR图像的统计性质,在双M arkov模型的框架下,对低层M arkov随机场提出了指数自回归EAR纹理模型,并对纹理含噪情形下的高层M arkov随机场模型给出了一种参数估计方法及相应的无监督分割算法。实验结果表明,与以往的有监督GAR模型和不考虑纹理的模型相比,无监督的双M arkov-EAR模型能大量降低分割时的错分率。Aim. The existing pairwise Markov-GAR（Gauss autoregressive） model for the segmentation of textured SAR（synthetic aperture radar） images suffers from two shortcomings .it requires supervision and its precision is not as good as can be. We now present what we believe to be a better segmentation method based on Markov-EAR（exponential autoregressive） model. In the full paper, we explain our Markov-EAR model in detail; in the abstract, we just add some pertinent remarks to listing the three topics of explanation. （1） pairwise Markov-EAR model; （2） parameter estimation based on pairwise Markov-EAR model; （3） the unsupervised segmentation method based on pairwise Markov-EAR model for textured SAR images; the two subtopics of topic 1 are Gibbs distribution based on labeling data （subtopic 1.1） and EAR model for SAR image data （subtopic 1.2）; the two subtopics of topic 2 are parameter estimation based on LLMRF（low level Markov random field）（subtopic 2. 1） and parameter estimation based on HLMRF（high level Markov random field）（subtopic 2.2）;under subtopic 1.2, we derive eqs. （4）, （5）, and （7）; under subtopic 2.1, we derive eqs. （9）, （10）, and （11）; under subtopic 2.2, we derive eqs. （13）, （14）, and （15）; the unsupervised segmentation algorithm given in topic 3 is directly based on the explanation given in topic 2. Finally we summarize our experimental results in Figs. 2 and 3 and Table 1. The experimental results show preliminarily that ： （1） the pixels incorrectly segmented are respectively 1665 and 2809 for the Markov-EAR model that does not require supervision and the Markov-GAR model that does; （2） the percentages of incorrect segmentation are respectively 2.5 and 4.4 for Markov-EAR model and Markov-GAR model.

关 键 词：SAR图像 纹理双Markov模型 无监督分割 EAR ECM算法 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术]