Traditional statistically-based SAR image change detection is usually limited by imaging conditions. We propose an unsupervised change detection algorithm based on the image inherent geometrical structures. Firstly, an approximate complete data set is constructed to describe the image local features via fuzzy derivatives,and then a feature vector space is made to represent the image inherent structures using image local neighborhood information and spatial correlation.Finally the change map is produced by simulated annealing-based k-means (SAKM) clustering arithmetic. Experimental results show great detection performance both in contour and region. Besides,it is robust to noises and registration.%传统的基于统计特性的SAR图像变化检测方法,易受成像条件的影响而误差很大.本文针对图像内在几何结构,提出了一种无监督SAR图像变化检测算法.通过模糊微分构造描述图像局部特征的近似完整数据集,并充分利用图像局部邻域信息和空间相关性,构造深入描述图像结构化信息的特征矢量空间,最后通过基于模拟退火的K均值(SAKM)聚类算法,实现图像变化区域与非变化区域的分类.实验证明,本文方法不仅能够很好的检测出图像的轮廓变化和图像的区域变化,而且对噪声及配准精度均具有很好的鲁棒性.
展开▼