摘要
提出一种结合KL(Kullback-Leibler)距离和图像域分块的SAR图像分割算法.首先,利用规则划分技术将图像域划分成若干规则子块,以子块为处理单元,假设子块内像素服从高斯分布,并构建特征场概率模型;其次,采用广义Potts模型定义刻画邻域子块相关性的标号场概率模型,根据贝叶斯定理,得到后验概率模型;再次,采用KL距离定义刻画同质区域间统计分布差异的异质性系数,并通过非约束吉布斯表达式构建概率分布函数,结合后验概率和吉布斯概率分布函数建立图像分割模型;然后,设计M-H(Metropolis-Hastings)采样方法,包括改变子块标号操作和分裂子块操作,模拟上述分割模型,从而获得最优分割结果;最后,通过对所提出算法和对比算法的SAR图像分割结果进行分析,充分验证了所提出算法的有效性和优越性.
In this paper, a segmentation method for synthetic aperture radar（SAR） images based on Kullback-Leibler（KL）distance and regular tessellation is proposed. Firstly, the image domain is divided into several sub-blocks by a regular tessellation, and the divided blocks are considered as basic processing units during segmentation. It is assumed that all pixels in a sub-block follow Gaussian distribution, while to modeling feature field of a given image. Then a general Potts model is utilized to model relationship between neighbor sub-blocks in label field. According to Bayes theorem, the posterior probability model is obtained by combining pixels＇ feature and sub-blocks＇ labels. Thereafter, the heterogeneity coefficient between classes is characterized with KL distance, and the corresponding probability distribution function is constructed by a non-constrained Gibbs distribution. Combining the posterior probability model and the non-constrained Gibbs distribution, the image segmentation model is established. In order to simulate the segmentation model, a Metropolis-Hastings（M-H） sampling method is designed, including the operations of changing label and splitting subblocks. By analyzing the segmentation results of the proposed algorithm and the comparing algorithms, the validity and superiority of the proposed algorithm are fully verified.
作者
赵泉华
高郡
赵雪梅
李玉
ZHAO Quan-huai;GAO Jun;ZHAO Xue-mei;LI Yu(School of Geomatics,Liaoning Technical University,Fuxin 123000 China;China Sciences Group Remote Sensing Group Technology Co Ltd,Tianjin 300380,China)
出处
《控制与决策》
EI
CSCD
北大核心
2018年第10期1767-1774,共8页
Control and Decision
基金
国家自然科学基金项目(41271435
41301479)
辽宁省自然科学基金项目(2015020090)
