Degradation-Aware Unfolding Network With Sparsity-Gated Experts

Deep Unfolding-based Networks (DUNs) have drawn substantial attention in Compressive Imaging (CI) tasks for their high performance and partial interpretability. Yet, existing DUNs typically cascade identical iterative modules in each iteration, which restricts their flexibility to capture and recover complex details and features in different images during reconstruction. To address this issue, we propose DUMoE, a novel sparsely-activated Deep Unfolding Mixture-of-Experts architecture template applicable to diverse CI tasks. DUMoE jointly leverages degradation-aware self-attention (DAM), multi-scale expert gating (MSGate), and SpaRSA-inspired experts (DUSE), with top-1 switch routing to sparsely activate one expert per stage and reduce FLOPs compared to dense multi-expert activation. Additionally, we introduce a Degradation-Aware Mask into the self-attention mechanism to prioritize degraded regions caused by dimensionality reduction, improving reconstruction fidelity and generalization. The iterative stages are further enhanced by Deep Unfolding SpaRSA Experts, which integrate domain-specific priors with data-driven learning. Extensive experiments on three CI tasks—natural image compressive sensing, magnetic resonance imaging, and snapshot compressive imaging—demonstrate that DUMoE consistently outperforms advanced methods in terms of reconstruction quality while maintaining competitive parameter and FLOP efficiency relative to strong unfolding baselines. These results highlight the potential of DUMoE as a shared and effective architecture template for CI recovery tasks. Our code is available at DUMoE.