CRetinex: A Progressive Color-Shift Aware Retinex Model for Low-Light Image Enhancement

Low-light environments introduce various complex degradations into captured images. Retinex-based methods have demonstrated effective enhancement performance by decomposing an image into illumination and reflectance, allowing for selective adjustment and removal of degradations. However, different types of pollutions in reflectance are often treated together. The absence of explicit distinction and definition of various pollution types results in residual pollutions in the results. Typically, the color shift, which is generally spatially invariant, differs from other spatially variant pollution and proves challenging to eliminate with denoising methods. The remaining color shift compromises color constancy both theoretically and in practice. In this paper, we consider different manifestations of degradations and further decompose them. We propose a color-shift aware Retinex model, termed as CRetinex, which decomposes an image into reflectance, color shift, and illumination. Specific networks are designed to remove spatially variant pollution, correct color shift, and adjust illumination separately. Comparative experiments with the state-of-the-art demonstrate the qualitative and quantitative superiority of our approach. Furthermore, extensive experiments on multiple datasets, including real and synthetic images, along with extended validation, confirm the effectiveness of color-shift aware decomposition and the generalization of CRetinex over a wide range of low-light levels.