A Multi-Scale Hybrid Attention Residual Network for Super-Resolution Image Reconstruction

Single Image Super-Resolution (SISR) aims to recover high-resolution details from low-resolution inputs, a task that remains challenging due to the significant loss of high-frequency information. To address limitations in cross-dimensional feature interaction and detail restoration, this paper proposes a Multi-scale Hybrid Attention Residual Network (MHARN). MHARN integrates three novel modules: a Progressive Convolution Group (PCG) that employs parallel dilated branches for dynamic multi-scale shallow feature aggregation; a Multi-head Hybrid Attention Module (MHAM) that combines channel, spatial, and positional information to capture detailed multi-dimensional features and long-range dependencies; and a Dynamic Feature Enhancement Module (DFEM) that generates parameterized kernels for adaptive feature extraction based on local image content. We evaluate the proposed model on five standard benchmark datasets: Set5, Set14, BSD100, Urban100, and Manga109. Extensive quantitative and qualitative experiments demonstrate that MHARN achieves excellent Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) performance. The results show that MHARN significantly reduces visual artifacts and recovers sharper textures compared to state-of-the-art methods, providing a robust solution for high-fidelity image reconstruction.