High-Resolution and Large-Scene Optical-SAR Registration via Star-UNet: A Local-to-Global Approach With Implicit Kernel Mapping

High-resolution, large-scene optical and SAR image registration is a fundamental task in remote sensing data processing, but it faces two main challenges: significant nonlinear radiometric differences and high computational costs. To address these challenges, this paper proposes a local-to-global registration framework with two core contributions. First, we designed the Star-UNet deep network, whose key component, the Star Block, maps features into a high-dimensional non-linear space through element-wise multiplication. This design enhances the separability between shared structures (e.g., edges) and modality-specific artifacts (e.g., SAR speckle noise), thereby generating features that are more robust to radiometric differences and noise. Second, we proposed the Localization and Regression Matching Loss (LRML). This loss function combines a classification task for robust localization with a regression task for optimizing the response map's peak to improve sub-pixel matching accuracy. We systematically evaluated the framework on public datasets with three distinct resolutions (10 m, 1 m, and 0.5 m) and applied it to multiple real-world large-scene image pairs. Experimental results show that the method achieves reliable registration accuracy across different resolutions, providing a solution for large-scene, multi-sensor image registration that balances accuracy, efficiency, and flexibility.