Infrared and visible image fusion based on multi-scale cross-attention

To address the shortcomings of existing infrared and visible image fusion methods in balancing modality-specific and shared structures, this paper proposes a fusion method based on multi-scale cross-attention. First, multi-scale deep features are extracted and enhanced through dual-branch residual dense blocks to highlight modality-specific information; then, a multi-scale cross-attention fusion layer is designed to integrate shallow and deep features at multiple scales, achieving effective fusion of complementary salient features. Comparative experiments on public datasets show that this method better preserves texture details and improves visual quality. Specifically, on the TNO dataset, the MI, SF, SD, and VIF metrics increased by 11.4%, 8.6%, 5.1%, and 3.6%, respectively; on the MSRS dataset, the MI, SD, VIF, and PSNR metrics increased by 26.9%, 4%, 12.8%, and 9.4%, respectively.