Retinex-based Lightweight Network for Low Light Image Enhancement

Images captured under low-light conditions not only appear overall dark but also often suffer from various noise and degradation issues. Poor-quality images are not suitable for training deep learning networks and are difficult for users to accept. Therefore, enhancing low-light images should not only increase the brightness but also eliminate noise hidden in dark areas. To address these challenges, this paper proposes a lightweight and efficient deep network composed of three components: Decomposition Network, Illumination Network, and Restoration Network. Following the Retinex theory, Decomposition Network decomposes the input image into reflectance and illumination component, which are then processed separately. A lightweight illumination-guided Restoration Network is constructed to eliminate noise hidden in dark areas. Finally, Illumination Network brightens the illumination under the guidance of the recovered reflectance. We trained the entire network with pairs of images taken under different lighting/exposure conditions and achieves high visual quality with fewer parameters. Numerous experiments were conducted to illustrate that, in terms of both parameter quantity and model performance, RLN has significant advantages compared to other methods.