Remote Sensing Image Registration Via Cyclic Parameter Synthesis and Spatial Transformation Network

Abstract

Aiming at the problems of insufficient feature extraction ability, many mismatching points and low registration accuracy of some remote sensing image registration algorithms, this study proposes a remote sensing image registration algorithm via cyclic parameter synthesis spatial transformation network. (1) We propose a feature extraction network framework combined with the improved spatial transformation network and improved Densely Connected Networks (DenseNet), which can focus on important areas of images for feature extraction.This framework can effectively improve the feature extraction ability of the model, so as to improve the model accuracy. (2) In the matching stage, we design the coarse filter and fine filter double filter architecture. Thus, the false matching points are effectively filtered out, which not only improves the robustness of the model but also improves the registration accuracy. Compared with the two traditional methods and two deep learning methods, the experimental results of this model are better in many indexes.