Low-Light Image and Video Enhancement: A Comprehensive Survey and Beyond

TL;DR

This survey addresses the core challenges of low-light image and video enhancement by introducing two mixed-exposure image datasets (SICE_Grad and SICE_Mix) and a large-scale nighttime video dataset (Night Wenzhou) to better reflect real-world conditions. It provides a structured critique of learning strategies, network architectures, loss functions, and evaluation metrics, and benchmarks representative LLIE methods across diverse datasets. The work highlights that mixed exposure and real-world video data remain underexplored, and it proposes future directions including semantic-aware enhancement, real-time LLVE, standardized benchmarks, and improved metrics. Together, these contributions advance understanding of LLIE and offer practical resources for developing and evaluating robust, real-world capable methods.

Abstract

This paper presents a comprehensive survey of low-light image and video enhancement, addressing two primary challenges in the field. The first challenge is the prevalence of mixed over-/under-exposed images, which are not adequately addressed by existing methods. In response, this work introduces two enhanced variants of the SICE dataset: SICE_Grad and SICE_Mix, designed to better represent these complexities. The second challenge is the scarcity of suitable low-light video datasets for training and testing. To address this, the paper introduces the Night Wenzhou dataset, a large-scale, high-resolution video collection that features challenging fast-moving aerial scenes and streetscapes with varied illuminations and degradation. This study also conducts an extensive analysis of key techniques and performs comparative experiments using the proposed and current benchmark datasets. The survey concludes by highlighting emerging applications, discussing unresolved challenges, and suggesting future research directions within the LLIE community. The datasets are available at https://github.com/ShenZheng2000/LLIE_Survey.

Figures (17)

• Figure 1: Real-world Images from the SICE cai2018learning Dataset. These images exhibit diverse exposure and lighting, making visual aesthetics and scene understanding challenging.
• Figure 2: A Milestone for Recent Representative Low-Light Image and Video Enhancement Methods. (1) Traditional Learning methods: PIE fu2015probabilistic and LIME guo2016lime. (2) Unsupervised Learning methods: EnlightenGAN jiang2021enlightengan and SCI ma2022toward. (3) Semi-Supervised Learning method: DRBN yang2020fidelity. (4): Zero-Shot Learning methods: ExCNet zhang2019zero, Zero-DCE guo2020zero, Zero-DCE++ li2021learning, RUAS liu2021retinex, RetinexDIP zhao2021retinexdip, and SGZ zheng2022semantic. (5) Supervised Learning methods: LLNet lore2017llnet, MBLLEN lv2018mbllen, LightenNet li2018lightennet, Retinex-Net pham2020low, SID cheng2016learning, DeepUPE wang2019underexposed, EEMEFN zhu2020eemefn, KinD zhang2019kindling, Xu et al. xu2020learning, DLN wang2020lightening, DeepLPF moran2020deeplpf, KinD++ zhang2021beyond, Zhang et al. zhang2021learning, UTVNett zheng2021adaptive, SDSD wang2021seeing, LLFlow wang2022low, SNR-Aware xu2022snr, URetinex-Net wu2022uretinex, Dong et al. dong2022abandoning, MAXIM tu2022maxim, BIPNet dudhane2022burst, LCDPNet wang2022local, and IAT cui2022illumination.
• Figure 3: A Hierarchical Taxonomy of Learning Strategies for Low-Light Image and Video Enhancement Methods.
• Figure 4: Learning Strategies for Traditional Learning-based Low-Light Image and Video Enhancement. See Section \ref{['learning_strategies']} for details.
• Figure 5: Learning Strategies for Deep Learning-based Low-Light Image and Video Enhancement. See Section \ref{['learning_strategies']} for details.