Deep Semi-Supervised Learning for Low-Light Image Enhancement

Deep recursive band network(DRBN) is the first semi-supervised learning method applied in low-light image enhancement and achieved state-of-the-art results right now. However, lack of extra same domain unsupervised images and the separated supervised and unsupervised modules hinder the further improvement of the performance. To overcome these two problems, in this paper, we propose the first joint training semi-supervised low-light image enhancement algorithm. Our algorithm consists of two parts: the unsupervised image selection part and the semi-supervised low-light image enhancement part. The unsupervised image selection part overcomes the first problem. Specifically, a scoring mechanism based on the QTP theory is used to score unsupervised low-light images, images with lower score are selected as the extra same domain unsupervised images for low-light image enhancement tasks. In semi-supervised low-light image enhancement part, we extend the MixMatch based semi-supervised classification algorithm into its semi-supervised regression version, and utilize recursive band learning(RBL) which is the first stage of DRBN as the supervised learning part of our model to solve the second problem. As our method can solve the two problems of DRBN simultaneously, ours can achieve better performance. Comprehensive experimental results on real datasets demonstrate the effectiveness of our method.