Robust Self-supervised Correspondence Learning for SAR-optical Image Matching

SAR-optical image matching is a prerequisite for the joint use of these two modalities in many remote sensing applications. Due to significant nonlinear radiometric difference, it is a challenging task to perform robust matching between SAR and optical images. Recently, supervised deep learning methods have become the mainstream in the task of image feature matching, demonstrating the great power in feature description. One drawback of supervised deep learning methods is the huge demand for a vast amount of labelled training data, causing a big obstacle to its application in remote sensing. For this problem, we propose a self-supervised deep learning method to find correspondences using local image structure feature for SAR-optical image matching. In order to learn to exploit image structure cues, our network operates on the gradient of the input image instead of raw image intensity to produce dense descriptors, which paves the way for the self-supervised training. Next, a 4-D correlation map of SAR-optical images is computed using these descriptors. Lastly, reliable point-to-point correspondences can be extracted from the correlation map by means of the mutual nearest neighbor (MNN) and the RANSAC-based outlier removal. Extensive experimental results demonstrate that our method can establish robust correspondences between SAR and optical images with no demand for labelled data.