Global Attention Super-Resolution Algorithm for Nature Image Edge Enhancement

Single-image super-resolution (SR) has long been a research hotspot in computer vision, playing a crucial role in practical applications such as medical imaging, public security and remote sensing imagery. However, all currently available methods focus on reconstructing texture details, resulting in blurred edges and incomplete structures in the reconstructed images. To address this problem, an edge-enhancement-based global attention image super-resolution network (EGAN) combining channel- and self-attention mechanisms is proposed for modeling the hierarchical features and intra-layer features in multiple dimensions. Specifically, the channel contrast-aware attention (CCA) module learns the correlations between the intra-layer feature channels and enhances the contrast in the feature maps for richer features in the edge structures. The cyclic shift window multi-head self-attention (CS-MSA) module captures the long-range dependencies between layered features and captures more valuable features in the global information network. Experiments are conducted on five benchmark datasets for × 2, × 3 and × 4 SR. The experimental results show that for × 4 SR, our network improves the average PSNR by 0.12 dB, 0.19 dB and 0.12 dB over RCAN, HAN and NLSN, respectively, and can reconstruct a clear and complete edge structure.