Dynamic multi-scale fusion single image super-resolution: a method based on Swin Transformer

Abstract

Single image super-resolution (SISR) aims to reconstruct high-resolution images from low-resolution inputs, essential for medical imaging, surveillance, and remote sensing. Traditional methods falter on complex textures and edges due to limited feature extraction and inefficient cross-scale fusion. This study introduces a dynamic multi-scale fusion mechanism using the Swin Transformer for SISR. It begins with a pre-trained Swin Transformer backbone to extract hierarchical features from the low-resolution image, spanning local to global scales. A dynamic fusion module then adaptively weights features across resolutions via scale-attention learned from input texture complexity through a multi-layer perceptron (MLP), reducing losses from fixed upsampling. Multi-head cross-attention with residuals aggregates these features for enhanced representation. A pixel reconstruction head yields the high-resolution output. Training combines L1 and VGG-based perceptual losses for better visual and structural quality. Evaluations on Set5, Set14, BSD100, Urban100, and Manga109 show superior results, with average PSNR gains of 1.5-2.5 dB over EDSR, RCAN, and SwinIR. Ablations validate component efficacy, while sensitivity tests confirm hyperparameter robustness. This advances SISR via adaptive Transformer fusion, providing a lightweight solution for practical use.