Gradient-Guided Swin Transformer with Edge-Aware Fusion for Enhanced Image Super-Resolution

Abstract Existing methods still suffer from undesirable structural distortions.To address this issue, we propose incorporating gradients as high-frequency information to enhance structural preservation. Building on this idea, we introduce the Gradient-Guided Swin Transformer-based (SwinGSR), a novel dual-branch transformer for image SR.Specifically, successive Residual Swin Transformer Blocks (RSTBs) are used to extract deep features from the SR branch (SRB) and gradient branche (GB).SwinGSR leverages high-frequency information to prevent edge structure loss and improve image reconstruction with gradient guidance.Furthermore, we propose a Edge-Aware Fusion Module (EAFM), which incorporates several Channel Shuffle Residual Blocks (CSRBs) as its basic blocks and a Multi-Dconv Head Transposed Attention Block (MDTAB) to integrate features from both branches. This approach addresses the Swin Transformer's limited receptive field and improves feature fusion.The CSRBs enhance generalization by applying channel shuffling.Meanwhile, the MDTAB fully utilizes edge feature information.Extensive experiments demonstrate that SwinGSR outperforms previous methods.