Structure-Aware Mamba Sequence Modeling for Image Super-Resolution

This paper introduces a structure-aware Mamba sequence modeling approach for image super-resolution (SR). While Mamba has shown success in NLP with linear complexity and long-sequence modeling, its application to visual tasks faces challenges due to the two-dimensional structural constraints of images. To address this, we propose a lightweight structure-aware prior module that quantifies regional importance using gradient and local variance cues. Based on this prior, we design a dualbranch Mamba block (DCMB) that jointly models local structural details and global contextual dependencies, fused via an adaptive gating mechanism. Furthermore, an adaptive scanning path strategy is introduced to improve structural continuity and reduce computational redundancy. Experiments on Urban100 demonstrate that our method achieves a 0.34 dB PSNR gain with only 0.84 M parameters and improves inference speed by approximately 32%, validating the effectiveness of structureaware sequence modeling for efficient visual reconstruction.