Low-light image enhancement with joint illumination and noise data distribution transformation

Sheng Guo, Wei Wang, Xiao Wang, Xin Xu

Abstract: Images captured in low-light environments often face two problems: low contrast and high noise, which are caused by the low number of photons in the environment. Existing low-light image enhancement methods mainly focus on the previous problem to increase visibility while the latter one is usually addressed with a post-processing module. However, there is a coupling relationship between illumination and noise, and ignoring it will result in under-/over-smoothing of the enhanced images. To solve this problem, we propose a novel low-light image enhancement method based on simultaneous adjustment on illumination and noise using unpaired data. In other words, we consider illumination and noise as a joint data distribution. The proposed method consists of two main branches: a Distribution Extraction branch which is used to extract the joint distribution of illumination and noise in normal-light images, and a Distribution Transformation branch which transforms the low-light images in a spatial domain through the joint distribution. Extensive experiment results show that the proposed model can reach the network capability that trained with rich paired data and achieved satisfactory results.