Pith. sign in

REVIEW 3 major objections 5 minor 90 references

DIME-Net claims that a single network, trained once on a blend of low-light and backlit pairs, can adaptively enhance both types of images without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A single Retinex-based network with sparse mixture-of-experts tone curves, trained on a mixed low-light/backlit dataset, improves PSNR/SSIM/LPIPS on LOLv1 and BAID without dataset-specific retraining.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The architecture is coherent and the MixBL dataset is a useful artifact, but the unstated BAID/MixBL split overlap is a real leak risk that must be fixed before the cross-dataset generalization claim holds. the 3 major comments →

arxiv 2508.13921 v1 pith:GUEAMPI4 submitted 2025-08-19 cs.CV

DIME-Net: A Dual-Illumination Adaptive Enhancement Network Based on Retinex and Mixture-of-Experts

Ziang Wang , Xiaoqin Wang , Dingyi Wang , Qiang Li , Shushan Qiao This is my paper
classification cs.CV
keywords low-light image enhancementbacklit image enhancementRetinex theorymixture of expertssparse gatingS-curve tone mappingstate-space attentionimage restoration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper sets out to show that low-light and backlit enhancement do not need separate models. It argues the two degradations have distinct brightness distributions—low-light images peak in the dark range, backlit images have both dark and bright peaks—so a fixed tone-mapping or a single Retinex illumination correction cannot serve both. DIME-Net therefore couples a Retinex-based illumination estimator to a mixture of sixteen S-curve experts chosen by sparse gating, and adds a U-Net restorer with illumination-aware cross-attention and a state-space attention module to clean the artifacts those corrections leave behind. To train this joint model, the authors assemble MixBL from existing low-light and backlit datasets, and report that the single MixBL-trained model reaches 26.05 dB PSNR on LOLv1 and 25.02 dB PSNR on BAID without retraining. If that generalization holds, it means one deployed model could handle the full range of real-world lighting degradation rather than a model per condition.

Core claim

Central claim: one network handles both low-light and backlit enhancement by choosing among a family of S-curve tone mappings rather than one fixed curve. The image is modeled as reflectance times illumination plus disturbances; a compensation map satisfying $\bar{L} \odot L = 1$ recasts enhancement as estimating illumination compensation and a residual degradation C. A top-k sparse gate over 16 S-curve experts lets single-peak dark images and dual-peak backlit images receive different curves. A U-Net damage restorer with illumination-aware cross-attention and sequential-state global attention removes C. Trained once on MixBL and tested without retraining, DIME-Net reports 26.05 dB PSNR on LOLv1 and 25

What carries the argument

The load-bearing component is the Mixture-of-Experts illumination estimator. It contains 16 parameterized S-curve experts, each a nonlinear intensity mapping $f(\sigma,n)=\frac{I^n}{I^n+\sigma^n}$ with a different n, and a top-k sparse gating network that reads the input's mean RGB and produces a softmax weight vector over experts. Around this, the Retinex identity $I \odot \bar{L} = R + C$ (with $\bar{L} \odot L = 1$) provides the justification: estimating the compensation map $\bar{L}$ and the degradation C separates brightness correction from artifact removal. The damage restorer—a U-Net with Illumination-Aware Cross Attention and Sequential-State Global Attention—carries out the artifact removal. The whole design exists so that

Load-bearing premise

The central claim assumes the BAID images randomly chosen for MixBL training do not overlap the BAID test images used for evaluation; if the sets overlap, the reported backlit generalization numbers are inflated.

What would settle it

Compute the intersection between the 600 BAID training images sampled for MixBL and BAID's official 368 test images; any overlap would invalidate the cross-dataset BAID result. Separately, test DIME-Net on a single image with an overexposed background and underexposed foreground: true dual-illumination adaptivity requires both regions to be restored at once, not just a global brightness adjustment.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single checkpoint can serve both low-light and backlit inputs, so an enhancement pipeline does not need to first classify the degradation type.
  • Joint training on mixed data is a viable route: the MixBL-trained model reaches 26.05 dB PSNR on LOLv1 and 25.02 dB on BAID without retraining, competitive with single-dataset specialists in the paper's comparisons.
  • The diversity of the expert curves is what preserves adaptivity: the ablation shows that using only n<1 or only n>1 S-curves degrades LOLv1 performance, while the full 16-expert set keeps it high.
  • The damage restorer is essential to the result: without it, LOLv1 PSNR drops from 26.05 to 9.325 dB, so illumination estimation alone does not produce clean images.
  • MixBL can serve as a common training/evaluation mix for future methods claiming joint low-light and backlit enhancement.

Where Pith is reading between the lines

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

  • The gating signal is a single global average RGB value, so the model's notion of 'illumination characteristics' is coarse; extending the gate to spatial or histogram features would be a natural way to handle images that are half dark and half bright.
  • The same recipe—Mixture-of-Experts over parameterized tone curves inside a Retinex estimation loop—could transfer to other degradations with distinct intensity distributions, such as haze, shadows, or underwater scenes.
  • A stiffer test of the no-retraining claim would be a mixed scene containing both low-light and backlit regions simultaneously; current benchmarks categorize whole images, so they do not fully exercise the sparse gate's per-image choice.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes DIME-Net, a Retinex-based enhancement network with a Mixture-of-Experts illumination estimator that uses 16 parameterized S-curve experts and top-k sparse gating, together with a U-Net damage restorer containing Illumination-Aware Cross Attention and Sequential-State Global Attention. To train one model for both degradations, the authors construct a hybrid dataset, MixBL, from low-light (LOLv2) and backlit (BAID) images. The paper reports quantitative comparisons on MixBL, external evaluation on LOLv1, held-out evaluation on BAID, and ablations of each module, claiming that a single DIME-Net model generalizes to both low-light and backlit images without retraining.

Significance. If the claims hold, a single model that handles both low-light and backlit enhancement is practically significant, and the S-curve MoE with sparse gating is a plausible mechanism for adapting to different illumination modes. The paper has real strengths: it proposes a new hybrid training set, retrains the baselines on that set, includes a cross-dataset test on LOLv1, and provides ablations isolating the MoE, illumination estimator, and damage restorer. However, the central generalization claim currently rests on an under-specified dataset split, the backlit part of the claim is not cross-dataset, and all metrics are single-run numbers without released code or data. These issues are load-bearing for the headline claim and need to be resolved.

major comments (3)
  1. [§4.1, Tables 1 and 2(b)] The MixBL construction is not specified precisely enough to rule out train/test leakage. The text says the training set includes '600 backlit images randomly selected from BAID' and the test set includes '200 BAID images, all randomly sampled', and then says the original BAID split is 2500:368. It is never stated that the 600 training images were drawn exclusively from the 2500-image BAID training split, nor that the 200 MixBL test images are disjoint from the 600 training images and from the 368-image BAID test split. Table 2(b) evaluates the MixBL-trained model on the BAID 368 test set; if any of the 600 training images overlap that test set, the reported PSNR of 25.02 dB is not a valid generalization result. The same ambiguity applies to the LOLv2 sampling used for Table 1. The authors must state the exact provenance of every MixBL train/test image and confirm disjointness, ideally by
  2. [Abstract, §4.5, Table 2(b)] The 'without retraining' claim is asymmetric across the two degradations. For low-light, the model is trained on LOLv2 and evaluated on the external LOLv1 test set, which is a genuine cross-dataset test. For backlit, both training and test come from BAID: 600 BAID training images versus the 368-image BAID test split. The BAID result therefore demonstrates held-out performance within one dataset, not generalization to an unseen backlit dataset. The paper should either evaluate on a separate backlit dataset or substantially temper the wording that it generalizes to backlit datasets without retraining.
  3. [§4.2–4.6, Tables 1–3] All quantitative claims are based on single training runs, with no error bars, no seeds, and no release of code or data. Several headline improvements are small relative to expected run-to-run variation: DIME-Net vs. RetinexFormer is +0.86 dB on MixBL, +0.70 dB on BAID, and +0.89 dB on LOLv1, and on BAID the reported BacklitNet PSNR of 25.06 actually exceeds DIME-Net's 25.02. Reporting mean±std over multiple runs, or at least a significance test, is needed to support the claim that these margins are meaningful. Without split lists and code/data, the central generalization claim cannot be independently checked.
minor comments (5)
  1. [§3.2, Eqs. (7)–(8)] The symbol n denotes both the S-curve exponent in Eq. (7) and the number of experts in the summation in Eq. (8). Using the same letter in two roles is confusing; please use distinct symbols.
  2. [§3.1, Eqs. (5)–(6)] Eq. (6) uses F_L^(1) and F_L^(2), but Eq. (5) only defines F_L. Please clarify the relationship between F_L, its intermediate features L_l, and the superscripted outputs.
  3. [§2.4, reference [11]] The text says 'Sun et al.' for the multi-modal fusion method, but the cited paper's first author is Bing Cao. Please align the citation.
  4. [§4.1] The 'randomly sampled' selections for MixBL have no seed. Provide a fixed split for exact reproducibility.
  5. [§1, Figure 2] The caption says '1,000 randomly sampled images from LOL and BAID' without specifying per-dataset counts or whether the samples come from train or test partitions. Please state these details.

Circularity Check

0 steps flagged

No significant circularity: DIME-Net is an empirical network evaluation grounded in external benchmarks; the dataset-overlap ambiguity is a leakage risk, not circularity.

full rationale

DIME-Net is an empirical network paper. The Retinex-based formulation (Eqs. 1–6) is an algebraic decomposition that introduces the compensation map L̄ and degradation term C as definitions; the network is then trained to estimate these quantities, so there is no claim that the model derives a result from its own output. The MoE S-curve experts (Eqs. 7–10) are parameterized nonlinear mappings with a learned sparse gate; the paper does not fit a parameter to a benchmark and then call the benchmark a prediction. Cross-dataset evaluations on LOLv1 and BAID are against held-out or external splits, and the baselines are retrained under the same MixBL condition. The only concern is a reproducibility ambiguity: Section 4.1 does not explicitly state that the 600 BAID training images and 200 BAID test images in MixBL are disjoint from BAID's original 2500:368 split, which could inflate generalization numbers if overlap exists. That is a data-integrity/leakage risk, not circularity. There are no self-citations that carry a load-bearing argument, and no uniqueness theorem is imported from the authors. Hence no circular step.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper relies on standard supervised deep learning assumptions plus a few architecture-specific postulates: the additive disturbance Retinex model, the learnability of the residual C, the sufficiency of 16 S-curve experts, and the sufficiency of a global average brightness gate. No new physical entities are introduced.

free parameters (5)
  • S-curve exponent n per expert (16 values) = not reported; described only as n<1 and n>1
    Chosen by hand to span dark-lifting and highlight-compressing curves; central to the MoE's adaptive behavior, but exact values are absent.
  • S-curve sigma per expert = not reported
    Sigma is part of Eq. (7) and shifts the curve horizontally; the paper de-emphasizes it but never states how it is set for each expert.
  • Top-k gate sparsity = not reported
    The number of experts activated per image controls sparse gating behavior and is never specified.
  • Loss weights (L1, SSIM, VGG) = 1.0, 0.5, 0.1
    Chosen without sensitivity analysis; these weights directly define the optimized trade-off among fidelity, structure, and perception.
  • MixBL training composition = 800 low-light (600 real, 200 synthetic) + 600 backlit; 400 test images
    The balance between low-light and backlit training data is hand-selected and may influence cross-domain generalization.
axioms (6)
  • domain assumption Image I decomposes as R⊙L under Retinex theory (Eq. 1).
    The entire framework assumes this decomposition holds for both low-light and backlit images.
  • domain assumption Degradations enter as additive disturbances R_hat and L_hat in the Retinex decomposition (Eq. 2).
    This disturbance model is introduced to justify the residual C, but no evidence is given that real sensor noise and color casts follow this form.
  • ad hoc to paper The residual term C in I⊙L_bar = R + C can be estimated by a U-Net from I⊙L_bar and illumination features L_l (Eqs. 4-6).
    The learnability and well-conditioning of this specific residual are assumed; the paper does not analyze why this parametrization is easier to learn than direct enhancement.
  • ad hoc to paper The 16 S-curve tone mappings with varying n span the illumination transformations needed for both low-light and backlit images (Eqs. 7-8).
    The paper asserts that a diverse set of n values is essential, but does not prove or analyze coverage of the space of illumination corrections.
  • ad hoc to paper Global average RGB I* is a sufficient gating signal to distinguish low-light from backlit images (Eq. 9).
    The gating depends only on a scalar per channel; backlit images with dual histogram peaks may share average brightness with normal images, yet this is not discussed.
  • domain assumption LOLv2 and BAID paired data provide reliable ground truth for supervised training.
    The method is trained end-to-end against reference images; errors or misalignments in these datasets would propagate into the reported metrics.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DIME-Net: A Dual-Illumination Adaptive Enhancement Network Based on Retinex and Mixture-of-Experts." pith.science (2026). https://pith.science/paper/GUEAMPI4

@misc{pith2026250813921,
  author       = {Pith},
  title        = {Pith review of: DIME-Net: A Dual-Illumination Adaptive Enhancement Network Based on Retinex and Mixture-of-Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUEAMPI4}},
  note         = {Machine review of arXiv:2508.13921}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Image degradation caused by complex lighting conditions such as low-light and backlit scenarios is commonly encountered in real-world environments, significantly affecting image quality and downstream vision tasks. Most existing methods focus on a single type of illumination degradation and lack the ability to handle diverse lighting conditions in a unified manner. To address this issue, we propose a dual-illumination enhancement framework called DIME-Net. The core of our method is a Mixture-of-Experts illumination estimator module, where a sparse gating mechanism adaptively selects suitable S-curve expert networks based on the illumination characteristics of the input image. By integrating Retinex theory, this module effectively performs enhancement tailored to both low-light and backlit images. To further correct illumination-induced artifacts and color distortions, we design a damage restoration module equipped with Illumination-Aware Cross Attention and Sequential-State Global Attention mechanisms. In addition, we construct a hybrid illumination dataset, MixBL, by integrating existing datasets, allowing our model to achieve robust illumination adaptability through a single training process. Experimental results show that DIME-Net achieves competitive performance on both synthetic and real-world low-light and backlit datasets without any retraining. These results demonstrate its generalization ability and potential for practical multimedia applications under diverse and complex illumination conditions.

Figures

Figures reproduced from arXiv: 2508.13921 by Dingyi Wang, Qiang Li, Shushan Qiao, Xiaoqin Wang, Ziang Wang.

Figure 1
Figure 1. Figure 1: PSNR comparison on low-light and backlit datasets. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Histogram peak distribution of brightness in the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our model architecture, primarily composed of an illumination estimator module based on a Mixture-of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of S-Curve Parameters on Histogram Map [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with baseline methods on the proposed MixBL dataset. DIME-Net effectively restores brightness [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on the LOLv1 (top) and BAID (bottom). (b) is a low-light enhancement method trained on [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

90 extracted references · 48 canonical work pages · 3 internal anchors

  1. [1]

    Mohammad Abdullah-Al-Wadud, Md Hasanul Kabir, M Ali Akber Dewan, and Oksam Chae. 2007. A dynamic histogram equalization for image contrast en- hancement.IEEE transactions on consumer electronics53, 2 (2007), 593–600

    2007

  2. [2]

    Masato Akai, Yoshiaki Ueda, Takanori Koga, and Noriaki Suetake. 2021. A Single Backlit Image Enhancement Method For Improvement Of Visibility Of Dark Part. In2021 IEEE International Conference on Image Processing (ICIP). 1659–1663. doi:10.1109/ICIP42928.2021.9506526

    arXiv 2021

  3. [3]

    Masato Akai, Yoshiaki Ueda, Takanori Koga, and Noriaki Suetake. 2023. Low- Artifact and Fast Backlit Image Enhancement Method Based on Suppression of Lightness Order Error.IEEE Access11 (2023), 121231–121245. doi:10.1109/ ACCESS.2023.3328534

    work page arXiv 2023

  4. [4]

    Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. 2021. Xcit: Cross-covariance image transformers.Advances in neural information processing systems34 (2021), 20014–20027

    2021

  5. [5]

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. Vivit: A video vision transformer. InProceedings of the IEEE/CVF international conference on computer vision. 6836–6846

    2021

  6. [6]

    Jiesong Bai, Yuhao Yin, Qiyuan He, Yuanxian Li, and Xiaofeng Zhang. 2024. Retinexmamba: Retinex-based mamba for low-light image enhancement.arXiv preprint arXiv:2405.03349(2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Samprit Bose, Sahil Nawale, Dhruv Khut, and Maheshkumar H. Kolekar. 2023. LumiNet: Multispatial Attention Generative Adversarial Network for Backlit Image Enhancement.IEEE Transactions on Instrumentation and Measurement72 (2023), 1–14. doi:10.1109/TIM.2023.3317384

    work page arXiv 2023

  8. [8]

    Antoni Buades, Jose-Luis Lisani, Ana Belen Petro, and Catalina Sbert. 2020. Backlit Images Enhancement Using Global Tone Mappings and Image Fusion.IET Image Processing14, 2 (2020), 211–219. doi:10.1049/iet-ipr.2019.0814

    work page arXiv 2020

  9. [9]

    Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang

  10. [10]

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang

  11. [11]

    Bing Cao, Yiming Sun, Pengfei Zhu, and Qinghua Hu. 2023. Multi-modal gated mixture of local-to-global experts for dynamic image fusion. InProceedings of the IEEE/CVF international conference on computer vision. 23555–23564

    2023

  12. [12]

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. 2022. Swin-unet: Unet-like pure transformer for medical image segmentation. InEuropean conference on computer vision. Springer, 205–218

    2022

  13. [13]

    Turgay Celik and Tardi Tjahjadi. 2011. Contextual and variational contrast enhancement.IEEE Transactions on Image Processing20, 12 (2011), 3431–3441

    2011

  14. [14]

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. 2018. Learning to see in the dark. InProceedings of the IEEE conference on computer vision and pattern recognition. 3291–3300

    2018

  15. [15]

    Qili Chen, Junfang Fan, and Wenbai Chen. 2021. An improved image enhancement framework based on multiple attention mechanism.Displays70 (2021), 102091

    work page 2021

  16. [16]

    Xingbo Dong, Wanyan Xu, Zhihui Miao, Lan Ma, Chao Zhang, Jiewen Yang, Zhe Jin, Andrew Beng Jin Teoh, and Jiajun Shen. 2022. Abandoning the bayer-filter to see in the dark. InProceedings of the ieee/cvf conference on computer vision and pattern recognition. 17431–17440

    work page 2022

  17. [17]

    Minhao Fan, Wenjing Wang, Wenhan Yang, and Jiaying Liu. 2020. Integrating semantic segmentation and retinex model for low-light image enhancement. In Proceedings of the 28th ACM international conference on multimedia. 2317–2325

    work page 2020

  18. [18]

    Tatiana Gaintseva, Martin Benning, and Gregory Slabaugh. 2024. RAVE: Residual Vector Embedding for CLIP-Guided Backlit Image Enhancement. doi:10.48550/ arXiv.2404.01889 arXiv:2404.01889 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.01889 2024

  19. [19]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752(2023)

    Pith/arXiv arXiv 2023

  20. [20]

    Albert Gu, Karan Goel, and Christopher Re. 2022. Efficiently Modeling Long Sequences with Structured State Spaces. InInternational Conference on Learning Representations. https://openreview.net/forum?id=uYLFoz1vlAC

    work page 2022

  21. [21]

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. 2020. Zero-reference deep curve estimation for low- light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1780–1789

    2020

  22. [22]

    Jiahao Huang, Liutao Yang, Fanwen Wang, Yinzhe Wu, Yang Nan, Weiwen Wu, Chengyan Wang, Kuangyu Shi, Angelica I Aviles-Rivero, Carola-Bibiane Schön- lieb, et al. 2025. Enhancing global sensitivity and uncertainty quantification in medical image reconstruction with Monte Carlo arbitrary-masked mamba. Medical Image Analysis99 (2025), 103334

    work page 2025

  23. [23]

    Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jian- chao Yang, Pan Zhou, and Zhangyang Wang. 2021. Enlightengan: Deep light enhancement without paired supervision.IEEE transactions on image processing 30 (2021), 2340–2349

    2021

  24. [24]

    Yeying Jin, Wenhan Yang, and Robby T Tan. 2022. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision. Springer, 404–421

    2022

  25. [25]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

    Pith/arXiv arXiv 2014

  26. [26]

    Edwin H Land and John J McCann. 1971. Lightness and retinex theory.Journal of the Optical society of America61, 1 (1971), 1–11

    work page 1971

  27. [27]

    Chongyi Li, Chunle Guo, and Chen Change Loy. 2021. Learning to enhance low-light image via zero-reference deep curve estimation.IEEE transactions on pattern analysis and machine intelligence44, 8 (2021), 4225–4238

    2021

  28. [28]

    Chongyi Li, Jichang Guo, Fatih Porikli, and Yanwei Pang. 2018. LightenNet: A convolutional neural network for weakly illuminated image enhancement. Pattern recognition letters104 (2018), 15–22

    work page 2018

  29. [29]

    Jiaqian Li, Juncheng Li, Faming Fang, Fang Li, and Guixu Zhang. 2020. Luminance- aware pyramid network for low-light image enhancement.IEEE Transactions on Multimedia23 (2020), 3153–3165

    work page 2020

  30. [30]

    Zhenhao Li, Kai Cheng, and Xiaolin Wu. 2015. Soft binary segmentation-based backlit image enhancement. In2015 IEEE 17th International Workshop on Multi- media Signal Processing (MMSP). 1–5. doi:10.1109/MMSP.2015.7340808

    work page arXiv 2015

  31. [31]

    Zhenhao Li and Xiaolin Wu. 2017. Learning-based restoration of backlit images. IEEE Transactions on Image Processing27, 2 (2017), 976–986

    work page 2017

  32. [32]

    Jinxiu Liang, Yong Xu, Yuhui Quan, Boxin Shi, and Hui Ji. 2022. Self-supervised low-light image enhancement using discrepant untrained network priors.IEEE Transactions on Circuits and Systems for Video Technology32, 11 (2022), 7332– 7345

    work page 2022

  33. [33]

    Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. 2023. Iterative prompt learning for unsupervised backlit image enhancement. InProceedings of the IEEE/CVF International Conference on Computer Vision. 8094– 8103

    work page 2023

  34. [34]

    Seokjae Lim and Wonjun Kim. 2020. DSLR: Deep stacked Laplacian restorer for low-light image enhancement.IEEE Transactions on Multimedia23 (2020), 4272–4284

    work page 2020

  35. [35]

    Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. 2021. Retinex- inspired unrolling with cooperative prior architecture search for low-light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10561–10570

    2021

  36. [36]

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. 2024. Vmamba: Visual state space model. Advances in neural information processing systems37 (2024), 103031–103063

    2024

  37. [37]

    Kin Gwn Lore, Adedotun Akintayo, and Soumik Sarkar. 2017. LLNet: A deep au- toencoder approach to natural low-light image enhancement.Pattern Recognition 61 (2017), 650–662

    2017

  38. [38]

    Kun Lu and Lihong Zhang. 2020. TBEFN: A two-branch exposure-fusion network for low-light image enhancement.IEEE Transactions on Multimedia23 (2020), 4093–4105

    work page 2020

  39. [39]

    Feifan Lv, Bo Liu, and Feng Lu. 2020. Fast enhancement for non-uniform il- lumination images using light-weight CNNs. InProceedings of the 28th ACM International Conference on Multimedia. 1450–1458

    work page 2020

  40. [40]

    Feifan Lv, Feng Lu, Jianhua Wu, and Chongsoon Lim. 2018. MBLLEN: Low-light image/video enhancement using cnns.. InBmvc, Vol. 220. Northumbria University, 4

    work page 2018

  41. [41]

    Xiaoqian Lv, Shengping Zhang, Qinglin Liu, Haozhe Xie, Bineng Zhong, and Huiyu Zhou. 2022. BacklitNet: A Dataset and Network for Backlit Image En- hancement.Computer Vision and Image Understanding218 (April 2022), 103403. doi:10.1016/j.cviu.2022.103403

    work page arXiv 2022

  42. [42]

    Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. 2022. Toward fast, flexible, and robust low-light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5637–5646

    2022

  43. [43]

    2019.PyTorch: an imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019.PyTorch: an imperative style, high-per...

    2019

  44. [44]

    Xiaohuan Pei, Tao Huang, and Chang Xu. 2024. Efficientvmamba: Atrous selective scan for light weight visual mamba.arXiv preprint arXiv:2403.09977(2024)

    Pith/arXiv arXiv 2024

  45. [45]

    Etta D Pisano, Shuquan Zong, Bradley M Hemminger, Marla DeLuca, R Eugene Johnston, Keith Muller, M Patricia Braeuning, and Stephen M Pizer. 1998. Con- trast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms.Journal of Digital imaging11 (1998), 193–200

    work page 1998

  46. [46]

    Stephen M Pizer, E Philip Amburn, John D Austin, Robert Cromartie, Ari Geselowitz, Trey Greer, Bart ter Haar Romeny, John B Zimmerman, and Karel Zuiderveld. 1987. Adaptive histogram equalization and its variations.Computer vision, graphics, and image processing39, 3 (1987), 355–368

    work page 1987

  47. [47]

    Wenqi Ren, Sifei Liu, Lin Ma, Qianqian Xu, Xiangyu Xu, Xiaochun Cao, Junping Du, and Ming-Hsuan Yang. 2019. Low-light image enhancement via a deep hybrid network.IEEE Transactions on Image Processing28, 9 (2019), 4364–4375

    work page 2019

  48. [48]

    Ali M Reza. 2004. Realization of the contrast limited adaptive histogram equaliza- tion (CLAHE) for real-time image enhancement.Journal of VLSI signal processing MM ’25, October 27–31, 2025, Dublin, Ireland Ziang Wang, Xiaoqin Wang, Dingyi Wang, Qiang Li, & Shushan Qiao systems for signal, image and video technology38 (2004), 35–44

    work page 2004

  49. [49]

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. 2017. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538(2017)

    Pith/arXiv arXiv 2017

  50. [50]

    Alvy Ray Smith. 1978. Color gamut transform pairs.ACM Siggraph Computer Graphics12, 3 (1978), 12–19

    work page 1978

  51. [51]

    Danai Triantafyllidou, Sean Moran, Steven McDonagh, Sarah Parisot, and Gre- gory Slabaugh. 2020. Low light video enhancement using synthetic data produced with an intermediate domain mapping. InComputer vision–ECCV 2020: 16th Eu- ropean conference, glasgow, UK, August 23–28, 2020, proceedings, part XIII 16. Springer, 103–119

    work page 2020

  52. [52]

    Thaweesak Trongtirakul, Werapon Chiracharit, and Sos S. Agaian. 2020. Single Backlit Image Enhancement.IEEE Access8 (2020), 71940–71950. doi:10.1109/ ACCESS.2020.2987256

    work page arXiv 2020

  53. [53]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  54. [54]

    Javier Vazquez-Corral, Praveen Cyriac, and Marcelo Bertalmío. 2018. Perceptually-Based Restoration of Backlit Images.Color and Imaging Conference26, 1 (Nov. 2018), 32–37. doi:10.2352/ISSN.2169-2629.2018.26.32

    work page arXiv 2018

  55. [55]

    Qiuhong Wang, Xueyang Fu, Xiao-Ping Zhang, and Xinghao Ding. 2016. A Fusion- Based Method for Single Backlit Image Enhancement. In2016 IEEE International Conference on Image Processing (ICIP). 4077–4081. doi:10.1109/ICIP.2016.7533126

    work page arXiv 2016

  56. [56]

    Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. 2019. Underexposed photo enhancement using deep illumination estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6849–6857

    work page 2019

  57. [57]

    Shuang Wang, Qianwen Lu, Boxing Peng, Yihe Nie, and Qingchuan Tao. 2024. DPEC: Dual-Path Error Compensation Method for Enhanced Low-Light Image Clarity.arXiv preprint arXiv:2407.09553(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  58. [58]

    Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. 2023. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 2654–2662

    2023

  59. [59]

    Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Lap-Pui Chau, and Alex Kot. 2022. Low-light image enhancement with normalizing flow. InProceedings of the AAAI conference on artificial intelligence, Vol. 36. 2604–2612

    work page 2022

  60. [60]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 4 (2004), 600–612

    2004

  61. [61]

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. 2018. Deep retinex decomposition for low-light enhancement.arXiv preprint arXiv:1808.04560(2018)

    Pith/arXiv arXiv 2018

  62. [62]

    Jiangwei Weng, Zhiqiang Yan, Ying Tai, Jianjun Qian, Jian Yang, and Jun Li. 2024. Mamballie: Implicit retinex-aware low light enhancement with global-then-local state space.arXiv preprint arXiv:2405.16105(2024)

    Pith/arXiv arXiv 2024

  63. [63]

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph E Gonzalez, Kurt Keutzer, and Peter Vajda. 2021. Visual transformers: Where do transformers really belong in vision models?. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 599–609

    work page 2021

  64. [64]

    Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. 2022. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5901–5910

    2022

  65. [65]

    Bin Xiao, Yunqiu Xu, Han Tang, Xiuli Bi, and Weisheng Li. 2019. Histogram learn- ing in image contrast enhancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 0–0

    work page 2019

  66. [66]

    Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. 2020. Learning to restore low- light images via decomposition-and-enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2281–2290

    work page 2020

  67. [67]

    Rui Xu, Shu Yang, Yihui Wang, Yu Cai, Bo Du, and Hao Chen. 2024. Visual mamba: A survey and new outlooks.arXiv preprint arXiv:2404.18861(2024)

    Pith/arXiv arXiv 2024

  68. [68]

    Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. 2022. Snr-aware low- light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 17714–17724

    2022

  69. [69]

    Kai-Fu Yang, Cheng Cheng, Shi-Xuan Zhao, Hong-Mei Yan, Xian-Shi Zhang, and Yong-Jie Li. 2023. Learning to adapt to light.International Journal of Computer Vision131, 4 (2023), 1022–1041

    work page 2023

  70. [70]

    Wenhan Yang, Shiqi Wang, Yuming Fang, Yue Wang, and Jiaying Liu. 2020. From fidelity to perceptual quality: A semi-supervised approach for low-light image enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3063–3072

    2020

  71. [71]

    Jing Ye, Changzhen Qiu, and Zhiyong Zhang. 2023. Spatio-temporal propagation and reconstruction for low-light video enhancement.Digital Signal Processing 139 (2023), 104071

    work page 2023

  72. [72]

    Jing Ye, Changzhen Qiu, and Zhiyong Zhang. 2024. A survey on learning-based low-light image and video enhancement.Displays81 (2024), 102614

    work page 2024

  73. [73]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shah- baz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5728–5739

    2022

  74. [74]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. 2020. Learning Enriched Features for Real Image Restoration and Enhancement. InComputer Vision – ECCV 2020, Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm (Eds.). Springer International Publishing, Cham, 492–511. doi:10.1007/9...

    doi:10.1007/978-3-030-58595- 2020

  75. [75]

    Lin Zhang, Lijun Zhang, Xiao Liu, Ying Shen, Shaoming Zhang, and Shengjie Zhao. 2019. Zero-shot restoration of back-lit images using deep internal learning. InProceedings of the 27th ACM international conference on multimedia. 1623–1631

    work page 2019

  76. [76]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  77. [77]

    Xuanqi Zhang, Haijin Zeng, Jinwang Pan, Qiangqiang Shen, and Yongyong Chen

    work page

  78. [78]

    Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. 2021. Beyond brightening low-light images.International Journal of Computer Vision129 (2021), 1013–1037

    work page 2021

  79. [79]

    Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. 2019. Kindling the darkness: A practical low-light image enhancer. InProceedings of the 27th ACM international conference on multimedia. 1632–1640

    2019

  80. [80]

    Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. 2022. Deep color consistent network for low-light image enhance- ment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1899–1908

    work page 2022

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.