Progressive Keypoint Detection with Dense Siamese Network for SAR Image Registration

Abstract

SAR image registration plays a key role in the applications such as change detection, image mosaic, which is one of the most challenging tasks in recent years due to speckle noise, geometric distortion and nonlinear radiation differences between SAR images. In the feature-based SAR image registration methods, the repeatability of keypoints and the effectiveness of feature descriptors could directly affect the registration accuracy. In this paper, we propose a stable feature intersection-based (FI) keypoint detector, which contains three progressive detectors, i.e., phase coherency (PC) detector, horizontal/vertical and major/minor diagonal oriented gradient detectors, and local coefficient of variation (LCoV) detector. Our proposed keypoint detector can not only effectively extract keypoints with high repeatability, but also greatly reduce the number of false keypoints, thus reducing the computational cost of feature description and matching. Then we propose the Cross Stage Partial Siamese Network (CSPSNet) to rapidly extract feature descriptors containing both deep and shallow features, which can be used to obtain more correct matching point pairs than traditional man-made shallow descriptors. In addition, we design a Hard Example Mining Loss (HEMLoss) to minimize the matching distance between the matching descriptor and adaptive selected non-matching descriptors. Experimental results on different pairs of SAR images demonstrate that our proposed method achieves better performance than other state-of-the-art methods.