Abstract

Noise and structural distortion significantly affect the performance of Single Image Super-Resolution (SISR). While recent advancements leverage Generative Adversarial Networks (GANs) to produce photo-realistic images, handling noise and preserving structure remain challenging. This paper introduces a novel approach, termed SNS, which enhances SISR for noisy images by integrating a denoising preprocessing module and a structure-preserving gradient branch. The denoising module learns the noise distribution and employs residual-skip connections to effectively remove noise before super-resolution. Simultaneously, the gradient branch restores high-resolution gradient maps and incorporates gradient and spatial losses to guide optimization, thus enforcing structural fidelity. GAN-based mechanisms are retained to synthesize fine details. Experimental evaluations demonstrate that the proposed method achieves superior performance on noisy datasets, with improvements in Perceptual Index (PI) and Learned Perceptual Image Patch Similarity (LPIPS), while maintaining competitive Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) compared to existing approaches, including those combined with DNCNN. On the Urban100 dataset (noise level 25), SNS achieved 3.6976 (PI), 0.1124 (LPIPS), 24.652 (PSNR), and 0.9481 (SSIM), confirming its effectiveness.