An Algorithm Based on a Feature Interaction-based Keypoint Detector and Sim-CSPNet for SAR Image Registration

Abstract

Synthetic Aperture Radar (SAR) image registration has recently been one of the most challenging tasks because of speckle noise, geometric distortion and nonlinear radiation differences between SAR images. The repeatability of keypoints and the effectiveness of feature descriptors directly affect the registration accuracy of feature-based methods. In this paper, we propose a novel Feature Intersection-based (FI) keypoint detector, which contains three parallel detectors, i.e., a Phase Congruency (PC) detector, horizontal/vertical oriented gradient detectors, and a Local Coefficient of Variation (LCoV) detector. The proposed FI detector can effectively extract keypoints with high repeatabilityand greatly reduce the number of false keypoints, thus greatly reducing the computational cost of feature description and matching. We further propose the Siamese Cross Stage Partial Network (Sim-CSPNet) to rapidly extract feature descriptors containing deep and shallow features, which can obtain more correct matching point pairs than traditional synthetic shallow descriptors. Through the registration experiments on multiple sets of SAR images, the proposed method is verified to have better registration results than the three existing methods.

Keywords

• sar image registration
• local coefficient of variation (lcov)
• phase congruency (pc)
• structure tensor
• dense siamese network