A comprehensive experiment-based review of low-light image enhancement methods and benchmarking low-light image quality assessment

Low-light image enhancement is a notoriously challenging problem. Enhancement of low-light images is intended to increase contrast, adjust the tone, suppress noise, and produce better aesthetic quality images. Over the years, a large number of state-of-the-art classical and deep learning-based methods have been designed. The purpose of this experimental review is to examine the generalization ability of these methods. In order to examine the generalization ability of these methods (classical and deep learning-based) empirical analysis and relevant comparisons are performed on a wide range of commonly used test datasets using a variety of evaluation techniques. In addition, we have studied the inconsistency of low-light evaluation methods and highlighted their drawbacks. This inconsistency raises the question of whether the existing methods are able to fairly evaluate enhancement methods. In order to address this question, we propose a large no-reference perceptual image quality assessment (PIQA) dataset. Additionally, different deep learning-based methods have been trained on this PIQA dataset in order to provide benchmarking for developing learning-based low-light assessment methods. Finally, this review paper is concluded with current challenges and suggestions for future work.