Change-Detection Map Learning Using Matching Pursuit

Abstract

Learning can be of great use when dealing with problems in various fields. Inspired by locally linear embedding from manifold, we propose a novel automatic change-detection method through an offline learning approach. The proposed method comprises three steps. First, two coupled dictionaries of the difference image (DI) patches and change-detection map patches are generated from known image pairs. Second, we approximately represent each patch of the input DI with respect to the DI dictionary by using the matching the pursuit algorithm. Third, the coefficients of this representation are applied with the change-detection map dictionary to generate the output change-detection map. This way, we exploit the relationship between the DI patches and the corresponding change-detection map patches based on two coupled dictionaries. In addition, the relationship guides us to construct the change-detection map for any given input DI. Experimental results on real synthetic aperture radar databases show that the proposed method is superior to its counterparts. Our method can obtain promising results, even though the dictionaries are prepared by simple random sampling from fixed training images.