A Remote sensing images registration method based on compound outlier removal strategy

摘要

Registration is an essential preprocessing procedure for extensive application of remote sensing images. As a robust feature-based descriptor used for image registration, SIFT shows great superiority in describing graphical structure with its invariant characteristic to illumination, rotation and scale changes. However, it is not well performed under low contrast and weak edge conditions, which has a high appearance with remote sensing images. In this paper, a novel remote sensing image registration is proposed to cope with issues that SIFT hardly performs an acceptable accurate feature point matching. Frequency based and spatial based algorithms, which namely are Log Gabor wavelet transformation and discrete probability relaxation respectively, are introduced as a compound strategy to figure out mismatching candidates. Experiment results show that the proposed method could get lower root mean square error (RMSE) and higher correct matching ratio (CMR) than other typical methods utilized to compare with. As proposed compound outlier removal strategy is out of relying on transformation modal, it is considerably feasible and well performed without RANSAC algorithm. Furthermore, this method is invariant to shift, rotation, illumination and scale change dues to SIFT participation.