Abstract

Low-light images suffer from low contrast and unclear details, which not only reduces the available information for humans but limits the application of computer vision algorithms. Among the existing enhancement techniques, Retinex-based and learning-based methods are under the spotlight today. The former methods focus on Retinex decomposition to obtain illumination and reflectance, which is a highly ill-posed problem that requires additional constraints. Nevertheless, more sophisticated constraints can make the two components more coupled. The latter methods train neural networks and learn the priors from datasets. However, the difficulty in collecting real datasets limits the performance. In this paper, we bridge the gap between these two methods. First, we propose a novel "two-one" strategy for Retinex decomposition, by which the decomposition is casts as a generation problem. Second, based on the strategy, a unified deep framework for low-light image enhancement is proposed. The networks take the randomly sampled noise as inputs to generate the latent illumination and reflectance. Third, the proposed method does not require any external dataset for training. Last but not least, our method can weaken the coupling relationship between the two components while performing Retinex decomposition. The proposed method is compared with ten competitive algorithms on seven public datasets, and the experimental results demonstrate the superiority of our method. Code is available at:https://github.com/zhaozunjin/RetinexDIP.

Retinex Decomposition of RetinexDIP

Comparisons with other methods

To visually compare our enhanced images against other methods, you can (1) click the picture regions with your mouse for obtaining the focus and then press the left or right keys on the keyboard to switch the pictures, or (2) click the arrow on the pictures with your mouse.

Cite

Z. Zhao, B. Xiong, L. Wang, Q. Ou, L. Yu and F. Kuang, "RetinexDIP: A Unified Deep Framework for Low-light Image Enhancement," in IEEE Transactions on Circuits and Systems for Video Technology, doi: 10.1109/TCSVT.2021.3073371.