MHDBN: Mamba-based hybrid dual-branch network for multi-focus image fusion.

Abstract

This paper proposes a Mamba-based Hybrid Dual-Branch Network (MHDBN) for high-quality multi-focus image fusion (MFIF). The network employs a parallel Mamba-ConvNeXt architecture. The Mamba branch efficiently captures long-range dependencies, while the ConvNeXt branch reinforces local texture representation. These branches are progressively aligned in the Hierarchical Feature Interaction Module (HFIM) and leverage the Multi-Scale Feature Aggregation Module (MSFAM) to adaptively emphasize in-focus regions. Finally, aided by an upsampling module, the model generates precise decision maps and fused outputs. On the three public datasets of Lytro, MFFW and SAVIC, a systematic comparison with 13 of the latest MFIF methods on 12 objective metrics demonstrate that MHDBN achieves the best performance on 6, 10, and 9 metrics, respectively. Notably, on SAVIC MHDBN improves mutual information (MI) by 17.4 % over the second-best method. Extensive quantitative and qualitative results collectively validate the overall superiority of MHDBN.