MDFN: A Multi-Scale Dual-Attention Fusion Network For Infrared And Visible Images
Abstract
Infrared-visible image fusion seeks a single representation that preserves thermal saliency while retaining visible-spectrum structure. Existing networks often perform fusion with a single interaction mechanism, leaving the modality gap, long-range scene consistency, and fine-grained cross-modal complementarity insufficiently modeled. This paper proposes the Multi-scale Dual-attention Fusion Network (MDFN). Two symmetric encoders are first pretrained with cross-modal contrastive learning. Mid-level features are enhanced by global-context self-attention, whereas deep features are exchanged through patch-adaptive cross-attention. A hierarchical gated decoder then selects scale-specific skip information for artifact-suppressed reconstruction. On TNO and M3FD, the model achieves EN/PSNR scores of 7.121/14.8314 and 7.3105/16.1962, respectively. Ablations show that contrastive pretraining and the complementary attention branches jointly improve information content, spatial detail, and visual fidelity.