ZERRIN-Net: Adaptive low-light image enhancement using Retinex decomposition and noise extraction

Low-light image enhancement aims at correcting the exposure of images taken under underexposed conditions while removing image noise and restoring image details. Most of the previous low-light image enhancement algorithms used hand-made a priori denoising in the corrected component; however, due to the large amount of detail information of the image contained in the corrected component and the presence of some pseudo-noise, the final enhancement results obtained by these solutions do not have the original noise removed, and the image details appear blurred. To solve the above problems, we propose ZERRIN-Net, a zero-shot low-light enhancement method based on Retinex decomposition. First of all, we first design the original noise extraction network N-Net, which can adaptively extract the original noise of low-light images without losing the detailed information of the images. In addition, we propose the decomposition network RI-Net, which is based on the Retinex principle and utilizes a simple self-supervised mechanism to help decompose a low-light image into a reflection component and a light component. In this paper, we conduct extensive experiments on numerous datasets as well as advanced vision tasks such as face detection, target recognition, and instance segmentation. The experimental results show that the performance of our method is competitive with current state-of-the-art methods. The code is available at: https://github.com/liwenchao0615/ZERRINNet