Deep Structural Information-Based Feature Point Matching for Optical and SAR Images

The reliable matching of optical and synthetic aperture radar (SAR) images is crucial for the efficient integration of multimodal remote sensing data. However, the matching accuracy is often hampered by the nonlinear radiometric distortions (NRDs), geometric distortions, and noise difference between optical and SAR images, which result in incorrect correspondences. To address these challenges, we propose a feature matching method that leverages deep structural information for optical and SAR images. Our method offers two key contributions: First, we introduce a keypoint detection network that extracts structural features through data augmentation and employs self-supervised learning, thereby reducing the reliance on manually labeled samples. Second, we design a dense descriptor network that maintains consistency under radiometric and geometric distortions by incorporating random geometric transformations during training, improving the robustness of the descriptors. Experiments conducted on two optical-to-SAR datasets demonstrate that the proposed method achieves superior matching performance across both urban images with distinct structural features and suburban images with weak textures, as evidenced by a significant improvement of at least 9% in the correct match rate (CMR) and the lowest root mean square error (RMSE) compared to other methods.