Unsupervised Low Light Image Enhancement Transformer Based on Dual Contrastive Learning

Low-light image enhancement aims to recover normal-light images from the images captured under very dim environments. While deep learning-based methods have achieved substantial success in this field, most of them require paired training data, which is difficult to be collected. We propose an Unsupervised Dual Contrastive Learning Transformer (UDCL-Transformer) where the unsupervised contrastive learning is for the first time introduced to the low light image enhancement task. From a different yet new perspective, we explore contrastive learning with an adversarial training effort to leverage unpaired low-light images and normal-light images. Our proposed method leveraged dual contrastive learning and generative adversarial networks to restore low light image. Patch-wise contrastive learning maximizes the mutual information between raw and restored images. Pixel-wise contrastive learning encourages the restored images to approach the positive samples and keep away from the negative samples in the embedding space. Generator based on Parallel Convolution Transformer (PC-Former) is proposed to capture the rich features of global and local context for better aggregate information. Extensive experiments with comparisons to recent approaches further demonstrate the superiority of our proposed method.