Structure-guided hyperspectral and multispectral image registration via cross-modal region-to-point alignment

Multimodal image registration aims to spatially align corresponding structures from different modalities and is a critical prerequisite for fusion-based downstream tasks. However, accurate registration remains challenging because differing sensor viewpoints and acquisition conditions induce geometric distortions and radiometric discrepancies, especially in hyperspectral images (HSIs) and multispectral images (MSIs). Most learning-based methods adopt a bottom-up strategy that aggregates pixel-level similarities, making them vulnerable to feature ambiguity in repetitive or homogeneous patterns. In this paper, we propose SCOPE, a structure-guided registration framework for hyperspectral and multispectral images that integrates structural prior learning with cross-modal region-to-point alignment. SCOPE adopts a hierarchical pipeline that models global shape priors, localizes modality-invariant regions, and estimates confidence-filtered point correspondences to refine alignment from the scene level to the point level. It uses synthetic geometric layouts with rule-based region of interest (ROI) annotations to train a detector that encodes structural priors and domain-specific spatial layouts. A self-supervised region localization task is designed to learn modality-invariant features by integrating the Mutual Feature Alignment Module (MFAM) into the basic detector, using region-level pseudo-labels constructed by aggregating predictions under diverse radiometric and geometric variations. To establish reliable correspondences across modalities, SCOPE performs confidence-filtered region-to-point refinement, focusing on top-ranked ROIs and high-confidence keypoint matches to improve homography estimation accuracy. SCOPE outperforms state-of-the-art methods in aligning HSIs and MSIs across multimodal datasets from both medical and remote sensing domains, demonstrating strong generalization and practical effectiveness, as further evidenced by high-quality fused images in a downstream fusion task.