GAFF: Global Attention Feature Flow Network for Optical and SAR Image Registration Under Geometric Transformations

The registration of optical and synthetic aperture radar (SAR) images under geometric distortions is critical for remote sensing applications such as image fusion and visual navigation. However, this task faces significant challenges due to inherent radiometric and geometric differences, diverse terrain conditions, and distinct noise patterns between the two modalities, which compromises the accuracy, robustness, and generalization of existing registration methods. To address these limitations, this paper introduces a novel global attention feature flow (GAFF) network, which synergistically combines model-driven and data-driven methodologies. Specifically, the framework leverages a hybrid CNN-Transformer architecture, integrating modality independent region descriptors (MIRD) for feature extraction, to achieve highly generalizable and consistent feature representations across modalities. GAFF employs a global attention mechanism to establish robust feature correspondences and generate the initial feature flow. Further, we introduce a hierarchical feature flow refinement module (HFRM) and a combined weight loss function to optimize feature flow generation. Experimental results demonstrate that GAFF delivers highly accurate, robust, and generalizable optical-SAR registration under varying geometric conditions, maintaining strong generalization capabilities across OS, WHU-OPT-SAR, and OSTerrain dataset.