Learning to see in the dark

Abstract

Low-light image enhancement aims to improve the visibility of images taken in low-light or nighttime conditions. Currently, most deep models are trained using synthetic low-light datasets or manually collected datasets with small sizes, which limits their generalization capability when encountering the low-light images captured in the wild. In this study, a domain adaptation framework is proposed to translate images between synthetic low-light images and real low-light images. Meanwhile, we embed a method into the proposed domain adaptation framework to generate low-light images of different brightness levels, which helps with the training process of low-light enhancement networks via data augmentation. Finally, an attention-guided U-Net is trained on the augmented dataset. Qualitative and quantitative evaluations show that our method is comparable to other state-of-the-art methods.