Multimodal Image Registration via Contrastive Learning and Multi-Scale Progressive Deformation Estimation

Multimodal images can provide richer scene information. However, due to differences in imaging mechanisms, infrared and visible images often exhibit significant modality differences and spatial misalignments, which pose challenges for registration and subsequent fusion tasks. To address this issue, this paper proposes a multimodal image registration method based on contrastive learning and multiscale progressive deformation estimation-CMPE. The method first introduces a Contrastive Learning Module (CLM) to extract cross-modal shared semantic features, significantly reducing the modality gap between infrared and visible images. Subsequently, a multiscale progressive registration framework based on an encoder-decoder structure is designed, and a Global-Local Attention Module (GLAM) is employed at each scale to adaptively select the features, which are then used to predict the deformation field at each scale. The multiscale predicted deformation fields are adaptively weighted and smoothed through a Dynamic Field Fusion Module (DFFM) and vector field integration, ensuring the continuity of the output deformation. Extensive experiments demonstrate that CMPE outperforms existing methods in both qualitative and quantitative evaluations.