DBMLFusion: dual feature supplementary branch mutual learning for the comprehensiveness and saliency of infrared-visible image fusion

Image fusion using visible and infrared images preserves the advantages of different modalities in fused images, e.g., detailed textures and salient features. Current methods usually cannot guarantee the comprehensiveness and saliency of multi-source image features in fused images due to focusing on extracting features from different modalities separately or jointly and using a single branch to obtain fusion results. To address this problem, a Dual Feature Supplementary Branch Mutual Learning (DBMLFusion) framework is proposed. Specifically, a proposed Cross-Modal Feature Extractor (CMFE) and a novel interactive learning loss are employed to extract features between two modalities and retain more sufficient salient features. Then, an Infrared Feature Supplementary Fusion Branch (IFSFB) with an Infrared Feature Fusion Module (IFFM) and a Visual Feature Supplementary Fusion Branch (VFSFB) with a Visual Feature Fusion Module (VFFM) that mutually learn two kinds of features to approach each other for systematically preserving comprehensive features. Such schemes can drive the last fusion module to achieve more sufficient fusion using comprehensive features. Extensive experiments demonstrate that the proposed DBMLFusion achieves a 4% improvement over other methods in the Infrared-Visible image Fusion (IVF) task and a 1.2% improvement in downstream visual tasks compared to state-of-the-art methods.