TSFuse: Triple-Stream Network for Infrared and Visible Image Fusion

In the field of infrared and visible image fusion, mainstream methods predominantly adopt a unified architecture to separately extract features from both modalities, often resulting in asymmetry in the extracted modal information. To tackle this limitation, we introduce a triple-stream fusion network, namely TSFuse. The framework, in addition to independently extracting features from infrared and visible images, incorporates an additional j oint feature extraction stream. We employ a joint transformer and Multi-receptive Field Block (MFB) to capture visible image features, while utilizing a joint transformer and DenseBlocks for extracting features from both infrared and joint modalities. This design ensures adaptive and comprehensive feature representation tailored to the distinct characteristics of each modality. We also design an Interactive Transformer Fusion (ITF) module and a Spatial Attention Fusion Module (SAFM) aiming to deeply integrate features from different modalities. Extensive comparative and ablation experiments demonstrate the great fusion performance of the proposed model, as well as the effectiveness of its individual modules. The codes will be available at https://github.com/Chlzzz/TSFuse.