Reproduce “M2Restore: Mixture-of-Experts-based Mamba-CNN Fusion Framework for All-in-One Image Restoration” on OpenResearch

About this paper

Yongzhen Wang, Yongjun Li, Zhuoran Zheng, Xiao-Ping Zhang, Mingqiang Wei

M2Restore presents a hybrid Mixture-of-Experts (MoE) and Mamba-CNN framework for all-in-one image restoration, adaptively processing diverse degradations while balancing global context and local detail. The model achieved the highest average PSNR (31.65 dB) and SSIM (0.936) across several benchmark datasets, demonstrating robust performance against various degradation types.

Key insights

• Dynamically routing inputs to specialized experts based on degradation-aware priors, guided by multimodal information, significantly enhances a model's adaptability and generalization to diverse degradation types.
• The synergistic combination of Mamba's long-range dependency modeling with CNNs' local feature extraction, mediated by an adaptive gating mechanism, effectively preserves both global coherence and fine-grained structural details.
• Integrating a frozen vision-language model (DA-CLIP) can provide valuable degradation priors and content-aware features, enabling more precise expert selection and feature fusion in complex restoration tasks.

Summary by alphaXiv.