CSIFT: A Structural Feature Encoding Framework for Cross-modal Image Registration

Multimodal sensor collaborative detection serves as an indispensable complementary technology in Earth observation. However, significant cross-modal discrepancies in multi-modal images severely limit the operational efficiency and alignment accuracy of traditional registration methods. Feature-based multimodal image registration techniques struggle to maintain stable performance across different modalities due to inherent radiometric differences. To address this limitation, this paper proposes a cross-modal structural information feature transform (CSIFT) framework by optimizing intermediate modality construction. The method employs the oriented FAST and rotated BRIEF (ORB) detector to extract keypoints and construct circular Gaussian-weighted oriented gradient features (CGOGF) through a Gaussian-weighted directional gradient allocation mechanism applied to neighborhood information. The gradient magnitude distribution based on annular distance eliminates truncation effects caused by linear interpolation, while principal directionaligned blocks ensure rotational consistency. By transforming cross-modal registration into a homogeneous feature matching problem through structural feature encoding, experimental results demonstrate that CGOGF significantly enhances crossmodal performance while preserving information integrity compared to traditional methods. Extensive evaluations on two public datasets show that CSIFT achieves the highest success rates (97.58% and 97.19%) across various multi-modal datasets, outperforming multiple traditional registration approaches.