Code

Swinir: Image restoration using swin transformer

Abstract ： Image restoration is a low-level visual problem . At present, most image restoration methods are based on cnn, As for those who shine in the field of advanced visual tasks Transformer But rarely use . So this paper is mainly based on Swin transformer A benchmark model is established . This model consists of three parts ： Shallow feature extraction 、 Deep feature extraction 、 Image reconstruction . especially , The depth feature extraction module consists of several residuals Swin transformer modular （RSTB） form , Each block has several Swin transformer Layer and a residual connection . The author conducted experiments on three representative tasks ： Image super-resolution （ Including classics 、 Lightweight and real-world image super-resolution ）、 Image denoising （ Including gray and color image denoising ） and JPEG Compression artifacts are reduced . Experimental results show that ,SwinIR Performance on different tasks is better than the most advanced methods , Up to 0.14∼0.45dB, The total number of parameters can be reduced at most 67%.

introduction

Most are based on CNN Our approach focuses on fine architectural design , For example, residual learning [43,51] And dense connections [97,81], Despite the traditional