DSSRM级联分割的SAR图像变化检测

SRM(Statistical Region Merging)分割算法具有快速、稳定和抗噪强的优点,基于此,本文提出一种基于DSSRM(Dynamic Sorting Statistical Region Merging)级联分割的SAR图像变化检测方法。首先,针对SRM算法基于单特征静态排序导致的过分割问题,提出一种动态排序模式的DSSRM算法以减少差异图像分割错误,该算法建立基于合并区域的多特征马氏距离排序准则,在每次合并之后更新区域邻接矩阵并重新排序;然后,基于互信息最小化准则构造多通道差异数据集以提高算法对区域合并的约束能力;最后,提出一种级联分割变化检测框架,第1级利用SRM算法将差异图像映射到超像素空间,第2级采用DSSRM算法对超像素进行动态合并获得收敛的分割结果,第3级采用简化SRM方法进行三次合并获得最终的变化检测图。实验结果表明,该方法可以获得比SRM方法和目前流行方法更好的检测性能。

SAR image change detection method of DSSRM based on cascade segmentation

Synthetic Aperture Radar (SAR) images have region homogeneity with gray and texture.Considering that SRM (Statistical Region Merging) algorithms of image segmentation are efficient,stable and robust against noise,we propose a novel change detection method based on cascade segmentation with Dynamic Sorting Statistical Region Merging (DSSRM) algorithm.Firstly a DSSRM algorithm based on dynamic sorting is proposed to overcome conventional SRM''s over-segmentation problem caused by single feature and static sorting.This algorithm takes the Manhattan distance of multi-feature of regions to be merged as the sorting criterion,and updates the adjacency matrix after each merging.Secondly based on the rule of minimizing mutual information we design a multi-channel complementary appearance model to improve the capability of constraint for region merging.Finally we present a cascade change detection framework with multiple levels.The first level projects difference image to super-pixel space via SRM,the second level utilizes DSSRM to dynamically merge regions;and the third level leverages a simplified SRM to realize region merging again to obtain final change detection map.Experimental results of the proposed method and proposed methods based on PCA and MRF are presented.By analysis and quantitative comparisons,the false alarm number and total of error number by DSSRMare decreased thereby the performance of KAPPA can get higher than methods based on PCA and MRF.DSSRM method is based on dynamic sorting algorithm with Manhattan distance of multi-feature of regions,it makes similar regions to be merged firstly.Experiments on construction of multi-channel illustrates that the more is the difference between channels the better is the performance of change detection.Our method improved the performance of SRM algorithm to avoid the over-segmentation phenomenon.Comparison experiments show that this method can obtain better performance of change detection than conventional SRM and state of art algorithms.