Blind SR: Semantic-assisted Spatially-Varying Blur Estimation

Low-Resolution (LR) images resulted from object motion or camera defocus generally have space-variant blur kernels. Traditional blind Super-Resolution (SR) reconstruction methods fail to consider this variability, resulting in subpar performance in practical applications. To address this issue, this paper introduces a multi-task learning framework and proposes a Semantic-Enhanced Blur Prediction Network (SEBPN). With blur prediction as the primary task, we incorporate a Cross-Spatial Interaction (CSI) network and an Improved Grouping Interactive Attention (Improved-GIA) module as dual-feature fusion network to effectively utilize both blur and semantic features. The CSI network enables feature interactions across different scales and stages, extending receptive field by introducing positional information and convolutional kernels of varying sizes. This extension enhances the network's sensitivity to spatial details, allowing it to more accurately capture regions with spatially varying blur. Additionally, we inrtoduce coordinate attention to improve the fusion of dual features, significantly boosting blur prediction accuracy. This enhancement, which combined with existing non-blind Super-Resolution reconstruction method, leads to higher-quality reconstructed images. This paper utilizes the NYUv2-BSR and Cityscapes-BSR datasets, which proposed by CMOS in space-variant blur domain, and conducts qualitative and quantitative experiments with the most advanced methods on real images. Results demonstrate the superiority of our approach, leading to state-of-the-art blind SR performance.