Deep Learning-Based Low-Light Image Enhancement: A Review

Abstract

With the rapid development of deep learning in the field of computer vision, low-light image enhancement (LLIE) has shifted from traditional handcrafted prior-based methods to data-driven learning paradigms. Deep models can automatically learn the complex non-linear mapping between low-light and normal-exposure images through end-to-end training, showing remarkable advantages in brightness improvement, detail restoration, and noise suppression. According to the dependence on labeled data and the form of supervision signals during training, existing deep learning-based LLIE methods are mainly divided into five categories: supervised, reinforcement, unsupervised, zero-shot, and semi-supervised learning. Among them, supervised learning is the most mature and mainstream technical route, which has evolved from convolutional neural networks and Retinex decomposition to Transformer and state-space models (e.g., Mamba), achieving accurate recovery of illumination, texture, and color with stable performance. Reinforcement learning formulates LLIE as a sequential decision-making problem but is still in the exploratory stage due to low learning efficiency and high sensitivity to reward function design. Unsupervised and zero-shot learning methods solve the problem of difficult data annotation in real scenarios by designing unpaired, no-reference, or self-supervised mechanisms, respectively. Semi-supervised learning balances data acquisition cost and enhancement performance by combining a small amount of labeled data with a large number of unlabeled data. This paper summarizes the core principles, representative works, advantages, and existing challenges of various methods, pointing out that supervised learning dominates LLIE due to its stability and precision, while other methods provide effective supplements in data-scarce or complex scenarios. Future research should focus on improving the adaptability of models to different signal-to-noise ratio regions and balancing global illumination consistency and high-frequency detail preservation.