DSG-Net: A Dual Semantic Guidance Network for Optical-SAR Image Registration

Abstract

Optical and synthetic aperture radar (SAR) image registration under complex geometric transformations remains a challenging task due to the pronounced discrepancy in imaging mechanisms and noise patterns, as well as the large spatial deformations. Despite recent advancements in sparse keypoint matching and dense correspondence estimation, existing approaches are still dominated by appearance-driven cross-modal similarity, whose stability and discriminability are inherently limited by their intrinsic domain gap. As an attempt to mitigate this problem, we propose DSG-Net, a dual semantic guidance network for optical-SAR image registration which exploits semantic structural information obtained from optical and SAR semantic segmentation, demonstrating enhanced cross-modal robustness against raw intensity representations. Specifically, our method decomposes segmentation maps into complementary semantic edges and semantic interiors, where semantic edges capture modality-invariant geometric structures and furnish robust cross-modal constraints along contours, while semantic interiors offer reliable localization cues under substantial spatial deformations via semantic consistency. Extensive experiments on the WHU-OPT-SAR and DFC2020 datasets show that DSG-Net consistently outperforms sparse and dense registration baselines under various geometric deformation scales. Notably, on the WHU-OPT-SAR dataset our DSG-Net achieves a CMR@1px of 79.40%, yielding a 12.7% relative gain over leading baseline. We further validate the cross-paradigm generalization of our design on both sparse and dense registration framework and achieves noticeable performance gains.