Illuminating the night: A light source-aware and exposure-balanced low-light enhancement approach for real nighttime scenes

Abstract

Low-light image enhancement aims to improve the visibility and quality of images taken under dim lighting conditions. Most existing approaches in this field often encounter challenges such as low or excessive enhancement, and exaggerated light sources in real nighttime scenes. To address these issues, we present a comprehensive and modular approach that combines several key elements. First, we employ a novel variation of self-supervised retinex-based network to achieve effective enlightening. Second, we utilize an adaptive light source-aware enlightenment module to consider the presence of light sources. Then, our illumination-aware exposure-balanced fusion module integrates the outputs of the two stages. This method significantly improves the quality of low-light and nighttime images by balancing exposure, contrast, and saturation, producing well-exposed results. Comprehensive experiments on two referenced datasets (LOL and EnlightenGAN) and two non-referenced datasets (LIME and ExDark) validate the effectiveness of our approach. Our approach consistently achieves balanced exposure and preserves natural color tones, as reflected in key metrics. Specifically, it demonstrates an average improvement of 8.96% in FID score and 4.43% in LPIPS score for referenced datasets, along with a 0.1186 enhancement in NIQE score for non-referenced datasets. The code and implementation instructions are available at https://github.com/PaulMahdizadeh123/LowLightEnh.