Low-Light Image Enhancement via Unsupervised Learning

The models based on unsupervised learning methods have achieved prominent achievement in several low-level tasks such as image restoration and low-light enhancement. Many of them are based on generative adversarial networks such as EnlightenGAN. Although EnlightenGAN can be trained without the need for paired images, there are still existing some issues such as insufficient illumination and color distortion. Inspired by the achievement in visual tasks made by Vision Transformer(ViT), we propose a discriminator based on ViT to replace the original fully convolutional network to solve this problem. Furthermore, to improve the illumination enhancement effect, we devise a new loss function enlightened by the luminance in SSIM and multi-scale SSIM. Our method surpasses the state-of-the-art on mainstream testing datasets.