Gaussian-MambaFuse: multimodal image fusion via Gaussian splatting and spatial-channel-aware Mamba

Effective multimodal image fusion is important for enhancing scene representation in applications such as video surveillance and medical diagnosis. However, directly fusing source images with insufficient clarity often causes detail loss and information mixing, which degrades the final fusion quality. To address this issue, we propose a multimodal image fusion framework that integrates a dual-branch 2D Gaussian splatting-based fine-grained restoration (GSFR) with spatial-channel-aware Mamba interaction. Specifically, a dual-branch architecture is first designed to extract modality-specific latent features from different source images. Then, a dual-branch 2D Gaussian splatting based restoration module is introduced to refine bottleneck features with continuous Gaussian priors, so as to enhance degraded local structures, textures, and high-frequency details before cross-modal fusion. On this basis, a multiview spatial-channel fusion framework is further constructed, where a channel-aware Mamba module is used to capture cross-modal channel dependencies, and a spatial-aware Mamba module is employed to model shared spatial saliency across modalities. In addition, a Multi-View Fusion Mamba (MVFM) module is developed to jointly exploit two source-specific branches and an intermediate fused branch for deeper complementary information integration. Experiments on infrared-visible and medical image fusion tasks demonstrate that the proposed method achieves superior quantitative and qualitative performance over several state-of-the-art methods, producing fused images with clearer structures, richer textures, and more salient target information.