Remote sensing image registration is very important for multi-temporal remote sensing data analysis. However, multi-temporal images have different radiation and geometric features, which leads to inadequate extraction of keypoints and unrobust descriptor construction, thereby affecting subsequent applications.
In this study, we propose a self-supervised neural network based on frequency domain information guidance(FIGNet). FIGNet introduces the prior of frequency domain by constructing a multi-way detector, making the predicted keypoints more representative. In addition, in the data generation phase, we employ a nonlinear image enhancement strategy to improve the robustness of the network to radiometric variations. Besides, we improve the accuracy of the network by adopting the matching strategy based on patches.
Finally, experimental results on the VIS-SAR dataset demonstrate the effectiveness of our proposed method.