Liang Jingyun ; Cao Jiezhang ; Sun Guolei ; Zhang Kai ; Van Gool Luc ; Timofte Radu

Abstract

Image restoration is a long-standing low-level vision problem that aims torestore high-quality images from low-quality images (e.g., downscaled, noisyand compressed images). While state-of-the-art image restoration methods arebased on convolutional neural networks, few attempts have been made withTransformers which show impressive performance on high-level vision tasks. Inthis paper, we propose a strong baseline model SwinIR for image restorationbased on the Swin Transformer. SwinIR consists of three parts: shallow featureextraction, deep feature extraction and high-quality image reconstruction. Inparticular, the deep feature extraction module is composed of several residualSwin Transformer blocks (RSTB), each of which has several Swin Transformerlayers together with a residual connection. We conduct experiments on threerepresentative tasks: image super-resolution (including classical, lightweightand real-world image super-resolution), image denoising (including grayscaleand color image denoising) and JPEG compression artifact reduction.Experimental results demonstrate that SwinIR outperforms state-of-the-artmethods on different tasks by $\textbf{up to 0.14$\sim$0.45dB}$, while thetotal number of parameters can be reduced by $\textbf{up to 67%}$.