A Single-Stage Unsupervised Denoising Low-Illumination Enhancement Network Based on Swin-Transformer

Abstract

Traditional low-light enhancement methods are often based on paired datasets for training. The training data is difficult to obtain and the resulting model has poor generalization. In unsupervised low-light enhancement networks, because paired data is not used for training, there is a lack of effective denoising methods, which often leads to noise problems. In this paper, we propose a low-light enhancement network based on Swin-Transformer and CNN, guided by brightness information. It is worth noting that it also has an unsupervised denoising module, which allows our network to achieve low-light restoration and noise removal goals with a single training session without any additional or fake samples. Our network is based on the idea of generative adversarial networks. We conducted extensive experiments on datasets such as LOL and VE-LOL-L, and compared our results with mainstream unsupervised learning, zero-shot learning, and even partially supervised learning low-light enhancement methods using a series of image indicators and visual effects such as PSNR, SSIM, NIQE, and LOE. The results show that our network is superior to mainstream unsupervised learning and partially supervised learning methods.