Unsupervised low-light Image Enhancement with Quality-Task-Perception Loss

Unsupervised low-light image enhancement is a classic task in image processing area. Deep learning based methods are the main approaches to tackle this task due to their effectiveness in whole visual areas. However, blur image or fake color, insufficient enhancement and missing semantic information are three common issues when utilize these methods. In this paper, we propose a new hybrid loss named Quality-Task-Perception (QTP) to alleviate these three problems simultaneously. Specifically, this hybrid loss includes three components: quality, task and perception. The quality part overcomes the blur image or fake color problem by enforcing the similarity between image quality scores of enhancement results and those of samples in daytime domain. The task part tackles the insufficient enhancement issue by constraining enhancement results to have higher daytime probabilities, where daytime probabilities are achieved by our proposed new pipeline. The perception part copes with the missing semantic information by restricting the phase consistency between the low-light image and its enhanced version. As this new hybrid loss is special designed to deal with these three problems, a more accurate low-light image enhancement model can be learned. Comprehensive experiments on both supervised and unsupervised datasets demonstrate the effectiveness of our proposed approach.