Semantic Segmentation-Guided Generative Network for Blind Image Super-Resolution

Early super-resolution (SR) methods assumed predefined degradation processes, but real-world degradation is complex and unknown. Blind SR methodologies are necessary for improving SR precision when blur kernels are unknown. Researchers have found that images exhibit repetitiveness and similarity across different-sized image patches. Generative Adversarial Networks (GANs) can extract degradation features from low-resolution (LR) images. With deep learning, one can obtain blur kernels with maximum similarity and use a trained Generator network (G) to generate LR-HR pairs for SR. However, CNNs in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$G$</tex> have difficulty capturing larger spatial structures and global information. To address this problem, we propose SDiffGAN, a Semantic Segmentation-Guided Generative Network for Blind Image Super-Resolution, which generates fine-grained structural information and learns internal image distributions. SDiffGAN proposes CLIP and SAM models to incorporate text prompts into segmentation masks, enabling independent noise mean adjustment across different semantic distributions. SDiffGAN also propose the High-Similarity-Pass Attention Network (HSPAN) to capture global information. Experimental results show SDiffGAN's effectiveness, with a maximum PSNR value on the Set5 dataset exceeding existing methods by up to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1. 5 ~ d B}$</tex>.