Employing Multiple Priors in Retinex-Based Low-Light Image Enhancement

Abstract

In the field of low-light image enhancement, images captured under low illumination suffer from severe noise and artifacts, which are often exacerbated during the enhancement process. Our method, grounded in the Retinex theory, tackles this challenge by recognizing that the illuminance component predominantly contains low-frequency image information, whereas the reflectance component encompasses high-frequency details, including noise. To effectively suppress noise in the reflectance without compromising detail, our method uniquely amalgamates global, local, and non-local priors. It utilizes the tensor train rank for capturing global features along with two plug-and-play denoisers: a convolutional neural network and a Color Block-Matching 3D filter (CBM3D), to preserve local details and non-local self-similarity. Furthermore, we employ Proximal AlternatingMinimization (PAM) and the Alternating DirectionMthod ofMultipliers (ADMM) algorithms to effectively separate the reflectance and illuminance components in the optimization process. Extensive experiments show that our model achieves superior or competitive results in both visual quality and quantitative metrics when compared with state-of-the-art methods. Our code is available at https://github.com/YangWeipengscut/GLON-Retinex.