Fuzzy topology–based method for unsupervised change detection

Abstract

This letter presents a novel framework for the analysis of difference image (DI) in unsupervised change detection problems based on fuzzy topology. First, the DI is softly categorized into unchanged and changed classes. In other words, a membership function is computed for each class. Second, each class is decomposed into three parts – interior, fuzzy boundary and exterior – by analysing its membership function through fuzzy topology. Third, for each class, its interior pixels which have a high membership degree are classified as the current class; its exterior, denoting the interior and fuzzy boundary parts of the other class, is ignored; and its fuzzy boundary pixels associated with a low membership degree are reclassified using the supported connectivity in fuzzy topology. As a result, the proposed approach can solve the problem of misclassifying pixels in the fuzzy boundary of unchanged or changed class to some extent, providing improved change detection accuracy. Experiments were conducted on two different datasets and the results confirm the effectiveness of the proposed framework.