Answers are grounded in this paper and cite their sources.
Infrared-visible image registration is a fundamental preprocessing step for multimodal image fusion, object detection, and complex scene understanding. However, due to the heterogeneous imaging mechanisms of infrared and visible sensors, the two modalities exhibit pronounced non-linear radiometric differences, weakly textured regions, and local occlusions, which make cross-modal geometric correspondence estimation challenging. To address the susceptibility of existing methods to mismatches in unreliable regions and their limited ability to characterize fine-grained geometry in low-texture areas, this paper proposes a confidence-aware infrared-visible image registration network with Multiple Local Detail Convolution, termed MCCAR. Specifically, a CNN-Transformer encoder is first employed to extract modality-shared structural representations and modality-specific detail features. Then, a multiple local detail convolution module is designed to enhance cross-modal geometric cues from multi-scale structures, omnidirectional edges, and Laplacian high-frequency residuals. Furthermore, a pixel-wise confidence estimation branch is introduced to adaptively assess regional matching reliability under occlusion, thermal contrast variation, and texture degradation, with the resulting confidence map serving as a soft mask in the loss function to suppress unreliable correspondences during optimization. Finally, hard negative contrastive learning is adopted to improve the discriminative capability of the network for subtle geometric deviations. Experiments on the RoadScene dataset demonstrate that MCCAR achieves superior registration performance in terms of mutual information, correlation coefficient, and Mean Square Error, while further improving the quality of registration-guided multimodal image fusion.
No takes yet. Share an insight, caveat, or question.
Fang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: