Weight Uncertainty Network for Low-Light Image Enhancement

Low-light image enhancement aims to recover details and visual information from corrupted low-light images. Most previous studies learn the mapping function between low/normal-light images using fixed-weighted neural net-works. However, these methods amplify the uncertainty present while parsing the content in the dark areas of the image, which consequently results in the presence of brightness artifacts and the loss of details in the enhanced results. To deal with this problem, we propose a weight uncertainty framework to enhance low-light images in an unsupervised manner. It represents network weights as probability distributions, and each weight coherently explains variability in the training data. Further, we impose the framework with the Retinex theory. Our method is trained under various brightness conditions and can generalize well to unknown brightness conditions. Extensive quantitative and qualitative experiments demonstrate that our method can achieve competitive performance against state-of-the-art solutions on different datasets.