pith. sign in

arxiv: 2507.04277 · v3 · submitted 2025-07-06 · 💻 cs.CV

Towards Lightest Low-Light Image Enhancement Architecture for Mobile Devices

Guangrui Bai , Hailong Yan , Wenhai Liu , Yahui Deng , Erbao Dong This is my paper

Pith reviewed 2026-05-19 05:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords low-light image enhancementlightweight neural networkunsupervised learningmobile devicesreal-time processingiterative restorationparameter efficiency
0
0 comments X

The pith

LiteIE enhances low-light images to 19.04 dB PSNR using 58 parameters and no labeled data on mobile devices.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces LiteIE as an ultra-lightweight unsupervised framework for low-light image enhancement tailored for mobile and embedded devices. It constructs the model with a minimal feature extractor of two convolutional layers and a parameter-free Iterative Restoration Module that reuses features to recover details progressively. Training uses a combination of exposure control, edge-aware smoothness, and multi-scale color consistency losses without needing paired supervision. This design allows the model to surpass existing methods in quality while drastically cutting parameter count and enabling real-time performance on smartphones. A reader would care because it addresses the practical barriers of computation and data requirements that have hindered deployment of enhancement algorithms on consumer hardware.

Core claim

LiteIE is an ultra-lightweight unsupervised enhancement framework consisting of a backbone-agnostic feature extractor with only two convolutional layers and a parameter-free Iterative Restoration Module. It is trained with an unsupervised objective integrating exposure control, edge-aware smoothness, and multi-scale color consistency losses. On the LOL dataset it reaches 19.04 dB PSNR, 1.4 dB above state-of-the-art while using 0.07% of the parameters, and runs at 30 FPS for 4K images on Snapdragon 8 Gen 3 with 58 parameters.

What carries the argument

The parameter-free Iterative Restoration Module, which reuses compact image features from two convolutional layers to progressively recover fine details without adding learnable parameters.

If this is right

  • Real-time low-light enhancement at 4K resolution becomes practical on edge devices with limited computational resources.
  • The need for large-scale paired datasets is eliminated for training effective enhancers.
  • Generalization across diverse low-light conditions improves through the multi-scale consistency losses.
  • Deployment on resource-constrained platforms is facilitated by the extreme parameter efficiency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar parameter-free iterative modules could be adapted for other mobile image processing tasks such as denoising or deblurring.
  • Exploring different iteration counts in the restoration module could reveal speed-quality trade-offs on specific hardware.
  • Integration into on-device camera pipelines might enable video enhancement without added hardware cost.

Load-bearing premise

The unsupervised training objective that combines exposure control, edge-aware smoothness, and multi-scale color consistency losses is sufficient to train a high-quality generalizable enhancer without any labeled paired data.

What would settle it

Applying LiteIE to a new low-light test set with ground-truth pairs and checking whether its PSNR falls below that of supervised competitors on the same data.

Figures

Figures reproduced from arXiv: 2507.04277 by Erbao Dong, Guangrui Bai, Hailong Yan, Wenhai Liu, Yahui Deng.

Figure 1
Figure 1. Figure 1: Performance and efficiency comparisons with state-of-the-art methods. (a) Image quality metrics (PSNR, SSIM) and runtime efficiency (CPU, GPU, and mobile phone SoC FPS). Our method consistently outperforms other approaches. (b) Visual results show LiteIE produces natural and perceptually pleasing enhancements. constraints of embedded and mobile platforms. Our goal is to explore the extreme lightweight boun… view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed LiteIE framework, consisting of a Lightweight Feature Extraction Network and an Iterative Restoration Module. The feature extractor uses two shared convolutional layers, repeated three times, with the final feature map as the enhancement matrix. Each iteration applies the Iterative Restoration Module to ensure detail recovery, improving overall image quality. to computational e… view at source ↗
Figure 3
Figure 3. Figure 3: Two weight-sharing convolutional layers (Conv3→1 3×3 and Conv1→3 3×3 ) constitute the feature extraction operator (𝑥), progressively extracting image features 𝜙1 , 𝜙2 , and 𝜙3 as references for restoration. The final feature map, 𝜙3 , serves as the output enhancement matrix. models and suggest further potential for LLIE, albeit with attention to computational cost and training complexity. 3. Proposed meth… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of models with and without the Iterative Restoration Module (IRM) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study on loss functions. Removing the color consistency loss (w/o lcol,gcol) leads to strong color shifts , while excluding the exposure loss (w/o exp), images appear dim, and discarding the smoothing term (w/o EA-TV) introduces visible artifacts [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison of representative methods. From left to right: Zero-DCE++, Zero-DCE, Zero-DCE + new loss, LiteIE + old loss, LiteIE + new loss, and ground truth. Zero-DCE++ shows noticeable color shifts. Incorporating the proposed global color loss improves color fidelity for both Zero￾DCE and LiteIE. Our full model (LiteIE + new loss) produces the most natural and visually balanced results. By utilizing… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of the LiteIE feature extraction network across channel configurations in [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study of the loss weighting parameters 𝛼∕𝛽 on PSNR performance across LOL-V1 [34], LOL-V2, and LSRW [7] datasets. 4. Evaluating Method Performance This section examines the lightweight LiteIE feature ex￾traction and evaluates the IRM’s impact on low-light image enhancement. 4.1. Ablation Study on Loss Functions We conduct ablation experiments to evaluate the contri￾bution of each loss component. A… view at source ↗
Figure 9
Figure 9. Figure 9: Visual Comparison of Low-Light Image Enhancement Methods [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Detection performance visualization on the Dark￾Face dataset [42] with DSFD [19] [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visual comparison of low-light image enhancement methods. Results are compared across LOL [34], MEF [23], NPE [33], and VV [31] datasets [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visual comparison of low-light image enhancement methods across LOL [34], MEF [23], NPE [33], and VV [31] datasets [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Limitation in extremely low-light conditions: our method enhances structural visibility but also amplifies noise in dark areas (red boxes), with zoomed-in patches showing noise and slight color deviations. RTX 4090 GPU and an Intel Xeon Silver 4310 CPU, as well as on smartphones powered by Kirin 990 5G and Snapdragon 8 Gen 3 SoCs [PITH_FULL_IMAGE:figures/full_fig_p011_14.png] view at source ↗
read the original abstract

Real-time low-light image enhancement on mobile and embedded devices requires models that balance visual quality and computational efficiency. Existing deep learning methods often rely on large networks and labeled datasets, limiting their deployment on resource-constrained platforms. In this paper, we propose LiteIE, an ultra-lightweight unsupervised enhancement framework that eliminates dependence on large-scale supervision and generalizes well across diverse conditions. We design a backbone-agnostic feature extractor with only two convolutional layers to produce compact image features enhancement tensors. In addition, we develop a parameter-free Iterative Restoration Module, which reuses the extracted features to progressively recover fine details lost in earlier enhancement steps, without introducing any additional learnable parameters. We further propose an unsupervised training objective that integrates exposure control, edge-aware smoothness, and multi-scale color consistency losses. Experiments on the LOL dataset, LiteIE achieves 19.04 dB PSNR, surpassing SOTA by 1.4 dB while using only 0.07\% of its parameters. On a Snapdragon 8 Gen 3 mobile processor, LiteIE runs at 30 FPS for 4K images with just 58 parameters, enabling real-time deployment on edge devices. These results establish LiteIE as an efficient and practical solution for low-light enhancement on resource-limited platforms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes LiteIE, an ultra-lightweight unsupervised low-light image enhancement framework for mobile devices. It consists of a backbone-agnostic feature extractor using only two convolutional layers to generate compact features and enhancement tensors, paired with a parameter-free Iterative Restoration Module that reuses features for progressive detail recovery. Training relies on an unsupervised objective combining exposure control, edge-aware smoothness, and multi-scale color consistency losses. The central empirical claims are that LiteIE achieves 19.04 dB PSNR on the LOL dataset (surpassing SOTA by 1.4 dB while using 0.07% of the parameters) and runs at 30 FPS for 4K images on Snapdragon 8 Gen 3 with only 58 parameters.

Significance. If the reported PSNR, parameter count, and runtime results hold under fair and reproducible conditions, the work would be significant for real-time computer vision on edge devices. The extreme parameter efficiency (58 parameters) combined with a parameter-free iterative module and fully unsupervised training removes reliance on large labeled datasets, which could enable broader deployment of low-light enhancement on resource-constrained platforms. The design choices demonstrate that high performance need not require large model capacity.

major comments (2)
  1. [Abstract] Abstract: The headline claim of 19.04 dB PSNR (1.4 dB above SOTA) with only 0.07% of the parameters is load-bearing for the contribution, yet the abstract provides no details on the specific SOTA baselines compared, their parameter counts, implementation sources, data splits on LOL, or evaluation protocol. Without this information, the efficiency and performance assertions cannot be verified or reproduced from the given text.
  2. [Unsupervised Training Objective] Unsupervised training objective: The paper states that the combination of exposure control, edge-aware smoothness, and multi-scale color consistency losses is sufficient to train a high-quality enhancer without any paired ground truth. However, no ablation isolating the contribution of each loss term or direct comparison of the identical two-conv + parameter-free architecture trained with a supervised L1/L2 objective on the same LOL splits is reported. Given the deliberately minimal capacity, this assumption directly supports the 19.04 dB claim and requires explicit validation.
minor comments (1)
  1. [Abstract] Abstract: The phrasing 'Experiments on the LOL dataset, LiteIE achieves 19.04 dB PSNR...' is grammatically incomplete and should be revised for clarity (e.g., 'On the LOL dataset, LiteIE achieves...').

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive feedback on our manuscript. We address each of the major comments below and outline the revisions we plan to make to strengthen the paper.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline claim of 19.04 dB PSNR (1.4 dB above SOTA) with only 0.07% of the parameters is load-bearing for the contribution, yet the abstract provides no details on the specific SOTA baselines compared, their parameter counts, implementation sources, data splits on LOL, or evaluation protocol. Without this information, the efficiency and performance assertions cannot be verified or reproduced from the given text.

    Authors: We agree with the referee that additional details in the abstract would enhance the verifiability of our claims. In the revised manuscript, we will update the abstract to specify the SOTA method being compared (the one with the highest reported PSNR on the LOL dataset), include its parameter count for context, clarify the LOL data splits used, and briefly describe the evaluation protocol. This revision will make the performance and efficiency assertions more transparent without altering the core claims. revision: yes

  2. Referee: [Unsupervised Training Objective] Unsupervised training objective: The paper states that the combination of exposure control, edge-aware smoothness, and multi-scale color consistency losses is sufficient to train a high-quality enhancer without any paired ground truth. However, no ablation isolating the contribution of each loss term or direct comparison of the identical two-conv + parameter-free architecture trained with a supervised L1/L2 objective on the same LOL splits is reported. Given the deliberately minimal capacity, this assumption directly supports the 19.04 dB claim and requires explicit validation.

    Authors: We acknowledge the value of ablations to validate the loss terms. We will add an ablation study in the experiments section of the revised paper to isolate the contribution of each loss component (exposure control, edge-aware smoothness, and multi-scale color consistency) to the overall performance. Regarding a direct comparison with supervised L1/L2 training on the same architecture, we note that our work focuses on the unsupervised setting to eliminate the need for paired data, which is a key advantage for practical deployment. We can include a discussion on this point in the revised manuscript. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results from unsupervised training on standard benchmarks

full rationale

The paper describes an architecture (two-conv feature extractor plus parameter-free Iterative Restoration Module) and an unsupervised objective (exposure control + edge-aware smoothness + multi-scale color consistency losses) then reports measured PSNR on the LOL dataset. These are experimental outcomes after training, not quantities obtained by algebraic reduction to the paper's own equations or by renaming fitted parameters as predictions. No self-citation is invoked as a uniqueness theorem or load-bearing premise for the performance numbers, and the 19.04 dB figure is not defined in terms of the training losses themselves. The derivation chain is therefore self-contained as a standard empirical validation of a proposed model and loss combination.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are described beyond standard convolutional layers and common unsupervised loss terms.

pith-pipeline@v0.9.0 · 5764 in / 1152 out tokens · 54220 ms · 2026-05-19T05:20:58.819145+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages · 2 internal anchors

  1. [1]

    Transform coefficient histogram-based image enhancement algorithms using contrast en- tropy

    Agaian, S.S., Silver, B., Panetta, K.A., 2007. Transform coefficient histogram-based image enhancement algorithms using contrast en- tropy. IEEE transactions on image processing 16, 741–758

    work page 2007

  2. [2]

    Learning photographicglobaltonaladjustmentwithadatabaseofinput/output image pairs, in: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition

    Bychkovsky, V., Paris, S., Chan, E., Durand, F., 2011. Learning photographicglobaltonaladjustmentwithadatabaseofinput/output image pairs, in: The Twenty-Fourth IEEE Conference on Computer Vision and Pattern Recognition

    work page 2011

  3. [3]

    IQA-PyTorch: Pytorch toolbox for im- age quality assessment

    Chen, C., Mo, J., 2022. IQA-PyTorch: Pytorch toolbox for im- age quality assessment. [Online]. Available: https://github.com/ chaofengc/IQA-PyTorch

    work page 2022

  4. [4]

    Fast Context-Based Low-Light Image Enhancement via Neu- ral Implicit Representations

    Chobola, T., Liu, Y., Zhang, H., Schnabel, J.A., Peng, T., 2024. Fast Context-Based Low-Light Image Enhancement via Neu- ral Implicit Representations. Springer Nature Switzerland. p. 413–430. URL: http://dx.doi.org/10.1007/978-3-031-73016-0_24 , doi:10.1007/978-3-031-73016-0_24

    work page doi:10.1007/978-3-031-73016-0_24 2024

  5. [5]

    Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.,

    work page

  6. [6]

    1780–1789

    Zero-reference deep curve estimation for low-light image en- hancement,in:ProceedingsoftheIEEE/CVFconferenceoncomputer vision and pattern recognition, pp. 1780–1789

    work page

  7. [7]

    Lime:Low-lightimageenhancement via illumination map estimation

    Guo,X.,Li,Y.,Ling,H.,2016. Lime:Low-lightimageenhancement via illumination map estimation. IEEE Transactions on image pro- cessing 26, 982–993

    work page 2016

  8. [8]

    R2rnet: Low-light image enhancement via real-low to real-normal network.JournalofVisualCommunicationandImageRepresentation 90, 103712

    Hai, J., Xuan, Z., Yang, R., Hao, Y., Zou, F., Lin, F., Han, S., 2023. R2rnet: Low-light image enhancement via real-low to real-normal network.JournalofVisualCommunicationandImageRepresentation 90, 103712

    work page 2023

  9. [9]

    Low-light image enhancementwithsemi-decoupleddecomposition

    Hao, S., Han, X., Guo, Y., Xu, X., Wang, M., 2020. Low-light image enhancementwithsemi-decoupleddecomposition. IEEEtransactions on multimedia 22, 3025–3038

    work page 2020

  10. [10]

    Jte- cflow for low-light enhancement and zero-element pixels restoration with application to night traffic monitoring images

    Hu, C., Hu, Y., Xu, L., Guo, Y., Cai, Z., Jing, X., Liu, P., 2024a. Jte- cflow for low-light enhancement and zero-element pixels restoration with application to night traffic monitoring images. IEEE Transac- tions on Intelligent Transportation Systems

    work page

  11. [11]

    Zmar-snflow: Restoration for low-light images with massive zero-element pixels

    Hu, C., Zhang, B., Hu, K., Xu, L., Wu, F., Cai, Z., Ye, M., Lu, X., 2024b. Zmar-snflow: Restoration for low-light images with massive zero-element pixels. Computers and Electrical Engineering 120, 109750

    work page

  12. [12]

    A reflectance re- weighted retinex model for non-uniform and low-light image en- hancement

    Jia, F., Wong, H.S., Wang, T., Zeng, T., 2023. A reflectance re- weighted retinex model for non-uniform and low-light image en- hancement. Pattern Recognition 144, 109823

    work page 2023

  13. [13]

    Lightendiffusion: Unsupervisedlow-lightimageenhancementwithlatent-retinexdiffu- sion models

    Jiang, H., Luo, A., Liu, X., Han, S., Liu, S., 2024. Lightendiffusion: Unsupervisedlow-lightimageenhancementwithlatent-retinexdiffu- sion models. arXiv preprint arXiv:2407.08939 . Guangrui Bai et al.:Preprint submitted to Elsevier Page 11 of 12 Expert Systems with Applications

    work page arXiv 2024

  14. [14]

    Low-lightimageenhancementwithquality-orientedpseudo labelsviasemi-supervisedcontrastivelearning

    Jiang, N., Cao, Y., Zhang, X.Y., Wang, D.H., He, Y., Wang, C., Zhu, S.,2025. Low-lightimageenhancementwithquality-orientedpseudo labelsviasemi-supervisedcontrastivelearning. ExpertSystemswith Applications 276, 127106

    work page 2025

  15. [15]

    Enlightengan: Deep light enhancement without paired supervision

    Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z., 2021. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing 30, 2340–2349

    work page 2021

  16. [16]

    Lightness and retinex theory

    Land, E.H., McCann, J.J., 1971. Lightness and retinex theory. Josa 61, 1–11

    work page 1971

  17. [17]

    Contrast enhancement based on layereddifferencerepresentationof2dhistograms

    Lee, C., Lee, C., Kim, C.S., 2013. Contrast enhancement based on layereddifferencerepresentationof2dhistograms. IEEEtransactions on image processing 22, 5372–5384

    work page 2013

  18. [18]

    Low-lightimageandvideoenhancementusingdeeplearning: A survey

    Li, C., Guo, C., Han, L., Jiang, J., Cheng, M.M., Gu, J., Loy, C.C., 2021a. Low-lightimageandvideoenhancementusingdeeplearning: A survey. IEEE transactions on pattern analysis and machine intelli- gence 44, 9396–9416

    work page

  19. [19]

    Learning to enhance low-light image via zero-reference deep curve estimation

    Li, C., Guo, C., Loy, C.C., 2021b. Learning to enhance low-light image via zero-reference deep curve estimation. IEEE transactions on pattern analysis and machine intelligence 44, 4225–4238

    work page

  20. [20]

    Dsfd:dualshotfacedetector,in:Proceedingsofthe IEEE/CVF conference on computer vision and pattern recognition, pp

    Li,J.,Wang,Y.,Wang,C.,Tai,Y.,Qian,J.,Yang,J.,Wang,C.,Li,J., Huang,F.,2019. Dsfd:dualshotfacedetector,in:Proceedingsofthe IEEE/CVF conference on computer vision and pattern recognition, pp. 5060–5069

    work page 2019

  21. [21]

    Iterative prompt learning for unsupervised backlit image enhancement, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp

    Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C., 2023. Iterative prompt learning for unsupervised backlit image enhancement, in: ProceedingsoftheIEEE/CVFInternationalConferenceonComputer Vision, pp. 8094–8103

    work page 2023

  22. [22]

    Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z., 2021. Retinex-inspired unrolling with cooperative prior architecture search for low-light imageenhancement,in:ProceedingsoftheIEEE/CVFconferenceon computer vision and pattern recognition, pp. 10561–10570

    work page 2021

  23. [23]

    Ntire 2024 challenge on low light image enhancement: Methods and results

    Liu,X.,Wu,Z.,Li,A.,Vasluianu,F.A.,Zhang,Y.,Gu,S.,Zhang,L., Zhu, C., Timofte, R., Jin, Z., et al., 2024. Ntire 2024 challenge on low light image enhancement: Methods and results. arXiv preprint arXiv:2404.14248

    work page arXiv 2024

  24. [24]

    Perceptual quality assessment for multi-exposureimagefusion.IEEETransactionsonImageProcessing 24, 3345–3356

    Ma, K., Zeng, K., Wang, Z., 2015. Perceptual quality assessment for multi-exposureimagefusion.IEEETransactionsonImageProcessing 24, 3345–3356

    work page 2015

  25. [25]

    Toward fast, flexible, and robust low-light image enhancement, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z., 2022. Toward fast, flexible, and robust low-light image enhancement, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5637–5646

    work page 2022

  26. [26]

    Cycle-interactive generative adversarial network for robust unsu- pervised low-light enhancement, in: Proceedings of the 30th ACM International Conference on Multimedia, pp

    Ni, Z., Yang, W., Wang, H., Wang, S., Ma, L., Kwong, S., 2022. Cycle-interactive generative adversarial network for robust unsu- pervised low-light enhancement, in: Proceedings of the 30th ACM International Conference on Multimedia, pp. 1484–1492

    work page 2022

  27. [27]

    Cheby- lighter: Optimal curve estimation for low-light image enhancement, in: Proceedings of the 30th ACM international conference on multi- media, pp

    Pan, J., Zhai, D., Bai, Y., Jiang, J., Zhao, D., Liu, X., 2022. Cheby- lighter: Optimal curve estimation for low-light image enhancement, in: Proceedings of the 30th ACM international conference on multi- media, pp. 1358–1366

    work page 2022

  28. [28]

    Pizer, S.M., Amburn, E.P., Austin, J.D., Cromartie, R., Geselowitz, A.,Greer,T.,terHaarRomeny,B.,Zimmerman,J.B.,Zuiderveld,K.,

    work page

  29. [29]

    Computer vision, graphics, and image processing 39, 355–368

    Adaptive histogram equalization and its variations. Computer vision, graphics, and image processing 39, 355–368

    work page

  30. [30]

    Lr3m: Robust low- light enhancement via low-rank regularized retinex model

    Ren, X., Yang, W., Cheng, W.H., Liu, J., 2020. Lr3m: Robust low- light enhancement via low-rank regularized retinex model. IEEE Transactions on Image Processing 29, 5862–5876

    work page 2020

  31. [31]

    Shi, Y., Liu, D., Zhang, L., Tian, Y., Xia, X., Fu, X., 2024. Zero- ig: zero-shot illumination-guided joint denoising and adaptive en- hancement for low-light images, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 3015– 3024

    work page 2024

  32. [32]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    Simonyan,K.,2014.Verydeepconvolutionalnetworksforlarge-scale image recognition. arXiv preprint arXiv:1409.1556

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  33. [33]

    Vvdataset

    Vonikakis,V.,2017. Vvdataset. Availableat: https://sites.google. com/site/vonikakis/datasets

    work page 2017

  34. [34]

    Exploring clip for assessing the look and feel of images, in: AAAI

    Wang, J., Chan, K.C., Loy, C.C., 2023. Exploring clip for assessing the look and feel of images, in: AAAI

    work page 2023

  35. [35]

    Naturalness preserved enhancement algorithm for non-uniform illumination images

    Wang, S., Zheng, J., Hu, H.M., Li, B., 2013. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing 22, 3538–3548

    work page 2013

  36. [36]

    Deep Retinex Decomposition for Low-Light Enhancement

    Wei, C., Wang, W., Yang, W., Liu, J., 2018. Deep retinex decompo- sition for low-light enhancement. arXiv preprint arXiv:1808.04560

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [37]

    Self- referencedeepadaptivecurveestimationforlow-lightimageenhance- ment

    Wen, J., Wu, C., Zhang, T., Yu, Y., Swierczynski, P., 2023. Self- referencedeepadaptivecurveestimationforlow-lightimageenhance- ment. arXiv preprint arXiv:2308.08197

    work page arXiv 2023

  38. [38]

    Uretinex-net: Retinex-based deep unfolding network for low-light imageenhancement,in:ProceedingsoftheIEEE/CVFconferenceon computer vision and pattern recognition, pp

    Wu, W., Weng, J., Zhang, P., Wang, X., Yang, W., Jiang, J., 2022. Uretinex-net: Retinex-based deep unfolding network for low-light imageenhancement,in:ProceedingsoftheIEEE/CVFconferenceon computer vision and pattern recognition, pp. 5901–5910

    work page 2022

  39. [39]

    Zero-shot adaptive low light en- hancement with retinex decomposition and hybrid curve estimation, in:2023InternationalJointConferenceonNeuralNetworks(IJCNN), IEEE

    Xia, Y., Xu, F., Zheng, Q., 2023. Zero-shot adaptive low light en- hancement with retinex decomposition and hybrid curve estimation, in:2023InternationalJointConferenceonNeuralNetworks(IJCNN), IEEE. pp. 1–8

    work page 2023

  40. [40]

    Upt-flow: Multi-scale transformer-guided normalizing flow for low-light image enhancement

    Xu, L., Hu, C., Hu, Y., Jing, X., Cai, Z., Lu, X., 2025. Upt-flow: Multi-scale transformer-guided normalizing flow for low-light image enhancement. Pattern Recognition 158, 111076

    work page 2025

  41. [41]

    Igdnet:Zero-shotrobustunder- exposed image enhancement via illumination-guided and denoising

    Yan,H.,Huang,J.,Huang,T.,2025a. Igdnet:Zero-shotrobustunder- exposed image enhancement via illumination-guided and denoising. arXiv preprint arXiv:2507.02445

    work page arXiv

  42. [42]

    Zero-shot image segmentationforsceneobjectsbasedonthel0gradientminimization andadaptivesuperpixelmethod

    Yan, H., Huang, J., Zheng, M., Tang, Y., 2025b. Zero-shot image segmentationforsceneobjectsbasedonthel0gradientminimization andadaptivesuperpixelmethod. NeuralComputingandApplications , 1–21

    work page

  43. [43]

    Yan,H.,Li,A.,Zhang,X.,Liu,Z.,Shi,Z.,Zhu,C.,Zhang,L.,2025c. Mobileie: An extremely lightweight and effective convnet for real- time image enhancement on mobile devices, in: Proceedings of the IEEE/CVF International Conference on Computer Vision

    work page

  44. [44]

    Advancing image understanding in poor visibility environments: A collective benchmarkstudy

    Yang, W., Yuan, Y., Ren, W., Liu, J., Scheirer, W.J., Wang, Z., Zhang, T., Zhong, Q., Xie, D., Pu, S., et al., 2020. Advancing image understanding in poor visibility environments: A collective benchmarkstudy. IEEETransactionsonImageProcessing29,5737– 5752

    work page 2020

  45. [45]

    Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model,in:ProceedingsoftheIEEE/CVFInternationalConferenceon Computer Vision, pp

    Yi, X., Xu, H., Zhang, H., Tang, L., Ma, J., 2023. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model,in:ProceedingsoftheIEEE/CVFInternationalConferenceon Computer Vision, pp. 12302–12311

    work page 2023

  46. [46]

    Beyond brightening low-light images

    Zhang, Y., Guo, X., Ma, J., Liu, W., Zhang, J., 2021. Beyond brightening low-light images. International Journal of Computer Vision 129, 1013–1037

    work page 2021

  47. [47]

    Zhang, Y., Teng, B., Yang, D., Chen, Z., Ma, H., Li, G., Ding, W.,

    work page

  48. [48]

    IEEE Transactions on Circuits and Systems for Video Technology 34, 5995–6008

    Learningasingleconvolutionallayermodelforlowlightimage enhancement. IEEE Transactions on Circuits and Systems for Video Technology 34, 5995–6008

    work page

  49. [49]

    Kindling the darkness: A practicallow-lightimageenhancer,in:Proceedingsofthe27thACM international conference on multimedia, pp

    Zhang, Y., Zhang, J., Guo, X., 2019. Kindling the darkness: A practicallow-lightimageenhancer,in:Proceedingsofthe27thACM international conference on multimedia, pp. 1632–1640

    work page 2019

  50. [50]

    Zhang, Z., Zhao, S., Jin, X., Xu, M., Yang, Y., Yan, S., Wang, M.,

    work page

  51. [51]

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    Noise self-regression: A new learning paradigm to enhance low-light images without task-related data. IEEE Transactions on Pattern Analysis and Machine Intelligence

    work page

  52. [52]

    Semantic-guided zero-shot learn- ing for low-light image/video enhancement, in: Proceedings of the IEEE/CVFWinterconferenceonapplicationsofcomputervision,pp

    Zheng, S., Gupta, G., 2022. Semantic-guided zero-shot learn- ing for low-light image/video enhancement, in: Proceedings of the IEEE/CVFWinterconferenceonapplicationsofcomputervision,pp. 581–590

    work page 2022

  53. [53]

    Cwfas-net: Low- light image enhancement using curvelet wavelet attention and fourier transform

    Zhou, H., Cao, C., Tong, J., Zheng, K., 2025. Cwfas-net: Low- light image enhancement using curvelet wavelet attention and fourier transform. Expert Systems with Applications 278, 127263

    work page 2025

  54. [54]

    Zero- shot restoration of underexposed images via robust retinex decom- position, in: 2020 IEEE International Conference on Multimedia and Expo (ICME), IEEE

    Zhu,A.,Zhang,L.,Shen,Y.,Ma,Y.,Zhao,S.,Zhou,Y.,2020. Zero- shot restoration of underexposed images via robust retinex decom- position, in: 2020 IEEE International Conference on Multimedia and Expo (ICME), IEEE. pp. 1–6. Guangrui Bai et al.:Preprint submitted to Elsevier Page 12 of 12

    work page 2020