IHMambaSR: An importance-guided hierarchical mamba with dynamic prompt for single image super-resolution
[2026/01/06] News: 🎉 IHMambaSR has been accepted by Pattern Recognition (PR)!
Characterized by the global receptive fields and linear computational complexity, Mamba-based methods have demonstrated remarkable potential in single image super-resolution (SISR). However, due to Mamba's inherent causal modeling, the fixed scanning methods employed by most existing approaches lead to suboptimal feature utilization, failing to prioritize content-rich regions crucial for reconstruction. Additionally, Mamba's long-range decay weakens feature interactions over long distances, compromising the model's ability to model global context for coherent texture reconstruction. Furthermore, directly flattening the image into a 1D sequence further degrades local structural correlations. These challenges highlight a fundamental mismatch between Mamba's 1D sequential design and the non-causal and spatial nature of image data, hindering its full potential in SISR. To overcome these limitations, we propose IHMambaSR, a framework underpinned by a novel Importance-Guided Hierarchical State-Space Module (IH-SSM) that integrates three core contributions. First, we introduce a Patch-wise Scanning Strategy (PSS) that globally evaluates patch importance scores to adaptively reorder the scanning sequence, breaking the rigid causality of conventional methods to ensure the scanning prioritizes highly informative regions. Second, a Hierarchical Patch Partition Strategy (HPPS) is employed to capture multi-scale local cues while concurrently preserving spatial continuity. Finally, we formulate a Zero-Centered Dynamic Prompt (ZCDP) mechanism, which repurposes the importance scores to bolster distant feature representations and effectively counteract long-range decay. Extensive experiments on multiple standard benchmarks show that our IHMambaSR achieves state-of-the-art performance on lightweight, classical, and real-world SR tasks, surpassing existing methods in both quantitative metrics and visual fidelity.
