Total: 1
Registering airborne SAR with spaceborne optical images is crucial for SAR image interpretation and geo-localization. It is challenging for this cross-platform heterogeneous image registration due to significant geometric and radiation differences, which current methods fail to handle. To tackle these challenges, we propose a novel grid-based multimodal registration framework (Grid-Reg) across airborne and space-born platforms, including a new domain-robust descriptor extraction network, Hybrid Siamese Correlation Metric Learning Network (HSCMLNet) and a grid-based solver (Grid-solver) for transformation parameters estimation. Our Grid-Reg is based on detector-free and global matching loss rather than accurate keypoint correspondences. These accurate correspondences are inherently difficult in heterogeneous images with large geometric deformation. By Grid-Solver, our Grid-Reg estimates transformation parameters by optimizing robust global matching loss-based patch correspondences of whole images in a coarse-to-fine strategy. To robustly calculate the similarity between patches, specifically that have noise and change objects, we propose HSCMLNet, including a hybrid Siamese module to extract high-level features of multimodal images and a correlation learning module (CMLModule) based equiangular unit basis vectors (EUBVs). Moreover, we propose a manifold loss EUBVsLoss to constrain the normalized correlation between local embeddings of patches and EUBVs. Furthermore, we curate a new challenging benchmark dataset of SAR-to-optical registration using real-world UAV MiniSAR data and optical images from Google Earth. We extensively analyze factors affecting registration accuracy and compare our method with state-of-the-art techniques on this dataset, showing superior performance.