A Novel and Robust Multi-Modal Remote Sensing Image Matching Method

The complicated nonlinear radiation distortion and geometric deformations between multi-modal remote sensing images pose a huge challenge. Usually, we use geographic information in remote sensing images to eliminate the influence of large geometric deformations. But geographic information itself may have errors and nonlinear radiation distortion makes precise matching difficult. In addition, the lack of large multi-modal remote sensing datasets makes it difficult to use deep learning to further improve matching results. In this paper, we find that edge features are less affected by non-linear radiometric distortions and are able to preserve the common structural properties of multi-modal images well. Therefore, we propose a hybrid method by integrating a pre-trained multi-scale edge detection network and a traditional template matching framework. We also use a coarse-to-fine strategy to reduce computational effort and improve matching accuracy. Extensive experiments show that our method achieves higher matchingaccuracy and better robustness as compared to previous methods