Grid-Reg: Grid-Based SAR and Optical Image Registration Across Platforms
Journal:
arXiv
Published Date:
Jul 6, 2025
Abstract
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.