A Novel Multi-level Feature Collaborative Matching Network for Optical and Synthetic Aperture Radar Image Registration

Due to the complementary characteristics of synthetic aperture radar (SAR) and optical images, image registration as a prerequisite for their information fusion has received increasing attention. Currently, learning-based methods can better handle the significant radiometric and geometric differences between optical and SAR images compared to traditional registration approaches, but they still have limitations in distinguishing difficult samples, making high-precision registration a remaining challenge. To address these challenges, this paper proposes a multi-level feature collaborative matching network (MFC-Net) that effectively integrates high-level abstract features and low-level spatial features for precise registration. Furthermore, a novel dual-dimension joint attention module (DDJA) is designed to dynamically capture feature dependencies across both channel and spatial dimensions, enhancing cross-modal feature consistency and improving matching performance. Additionally, to address the problem of similarity between hard positive and negative samples caused by high-precision registration requirements, a dynamic differentiation factor is introduced at the loss function level, enabling the model to better distinguish between these similar samples in training. Extensive experiments conducted on the WHU-OPT-SAR data set and WHU-SENCity data set demonstrate that the proposed MFC-Net outperforms state-of-the-art methods in both matching accuracy and precision, validating its superiority in cross-modal image registration tasks.