Multi-Modal Image Fusion with State Space Mamba Networks

To tackle problems of structural redundancy, detail loss, and limited cross-modal feature fusion in infrared and visible image fusion, we introduce a two-stage fusion network based on state-space modeling with Mamba. Specially, the proposed network uses a Transformer-CNN decoupled encoder to extract global semantic and local texture features, while the Mamba module dynamically fuses base and detail information. With its selective state-space modeling mechanism, Mamba captures cross-modal dependencies efficiently under linear computational complexity, boosting feature expressiveness and reducing redundancy. The final fused output is refined using channel mixing, residual convolution, and channel compression modules. Training follows a two-phase scheme to separately optimize feature decoupling reconstruction and the quality of the fused image. Experiments on the MSRS, TNO, and RoadScene datasets demonstrate superior performance across 8 metrics, with notable advantages in VIF, SSIM, and MI. The proposed approach holds significant promise for Multi-modal applications, including remote sensing, smart grid systems, and digital twin modeling.