[Paper Review] Low-Light Image and Video Enhancement Using Deep Learning: A Survey

Chongyi Li, Chunle Guo+5 more|arXiv (Cornell University)|Apr 21, 2021

Image Enhancement TechniquesComputer Science17 references30 citations

TL;DR

This survey reviews deep learning–based low-light image and video enhancement methods, datasets, losses, and platforms, and introduces a new dataset and an online evaluation platform.

ABSTRACT

Low-light image enhancement (LLIE) aims at improving the perception or interpretability of an image captured in an environment with poor illumination. Recent advances in this area are dominated by deep learning-based solutions, where many learning strategies, network structures, loss functions, training data, etc. have been employed. In this paper, we provide a comprehensive survey to cover various aspects ranging from algorithm taxonomy to open issues. To examine the generalization of existing methods, we propose a low-light image and video dataset, in which the images and videos are taken by different mobile phones' cameras under diverse illumination conditions. Besides, for the first time, we provide a unified online platform that covers many popular LLIE methods, of which the results can be produced through a user-friendly web interface. In addition to qualitative and quantitative evaluation of existing methods on publicly available and our proposed datasets, we also validate their performance in face detection in the dark.This survey together with the proposed dataset and online platform could serve as a reference source for future study and promote the development of this research field. The proposed platform and dataset as well as the collected methods, datasets, and evaluation metrics are publicly available and will be regularly updated.

Motivation & Objective

• Survey the learning strategies, network architectures, loss functions, and datasets used for deep learning–based LLIE methods.
• Analyze generalization to real-world low-light conditions and identify open issues.
• Introduce a new low-light image/video dataset with cross-device illumination and an online platform for evaluation.
• Provide guidelines and insights to advance LLIE research and facilitate benchmarking.

Proposed method

• Categorize LLIE methods by learning strategy (supervised, reinforcement, unsupervised, zero-shot, semi-supervised).
• Discuss representative end-to-end, deep Retinex-based, and realistic data–driven supervised approaches.
• Describe Retinex-inspired and other network designs that estimate illumination and reflectance components.
• Present a real-world, cross-device LLIE dataset and a dataset for video LLIE, plus a semi-supervised and zero-shot approaches where applicable.
• Introduce an online platform that enables user-friendly, GPU-free evaluation of multiple LLIE methods on input images.
• Summarize commonly used loss functions (reconstruction, perceptual, smoothness, adversarial, exposure) and data formats (RGB and raw).

Experimental results

Research questions

• RQ1 How well do deep learning LLIE methods generalize to real-world, cross-device low-light data?
• RQ2 What learning strategies, network designs, and loss functions yield robust LLIE performance across diverse conditions?
• RQ3 Do Retinex-based models offer advantages in practice, and what are their limitations when combined with deep networks?
• RQ4 Can a unified online platform and diverse dataset accelerate LLIE research and benchmarking?

Key findings

• Supervised learning remains the mainstream approach, accounting for 73% of methods analyzed.
• A mix of network structures is used, with U-Net–like and multi-branch architectures being common; Retinex-inspired designs are widely explored.
• RGB is the dominant data format, but raw data are valuable for high dynamic range and color fidelity improvements.
• Common losses include L1/L2, SSIM, perceptual, and smoothness; non-reference losses like exposure loss are important for generalization.
• Several real-world datasets (e.g., SID, DRV, MIT-Adobe FiveK) and synthetic data strategies are discussed to improve generalization; an online platform and a new cross-device LLIE dataset are introduced as benchmarks.
• The survey validates methods on face detection in the dark, highlighting the impact of LLIE on high-level vision tasks.