Pith. sign in

REVIEW 3 major objections 69 references

Depth estimated from Retinex reflectance, fused by multi-scale attention, restores low-light images while keeping scene structure intact.

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 ↗

T0 review · grok-4.5

2026-07-11 10:13 UTC pith:6JY35CPJ

load-bearing objection Solid engineering extension of depth-guided LLIE with a useful dataset; depth-prior accuracy is the soft spot, not a collapse of the claim. the 3 major comments →

arxiv 2607.05005 v1 pith:6JY35CPJ submitted 2026-07-06 cs.CV

Geometry-aware Depth-guided Representation Learning for Structure-preserving Low-light Image Enhancement

Fang Gao , Jiongkai Qin , Jiabao Wang , Jingfeng Tang , Ming Cheng , Hanbo Zheng , Qingbao Huang , Cheng Wu This is my paper
classification cs.CV
keywords low-light image enhancementgeometry-aware representation learningdepth-guided feature learningcross-modal attentionstructure-preserving restorationRetinex decompositionLOL-D dataset
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.

Low-light images lose not only brightness but also the geometric cues that tell where objects sit in space. Most enhancers only brighten and denoise; they treat the image as a flat map of colors and therefore blur edges and collapse depth layers. This paper claims that a reliable depth map, obtained by first stripping illumination via Retinex and then running a monocular depth model on the resulting reflectance, can serve as a structural prior that steers restoration. That prior is injected only inside the encoder through a multi-scale depth-aware attention block, so geometry guides feature learning without contaminating the final reconstruction. The resulting network, DMSA-Net, produces brighter, cleaner images that keep object boundaries and spatial layout more consistent than appearance-only baselines, and the authors release the depth-augmented LOL-D dataset so others can build on the same idea.

Core claim

Reliable monocular depth estimated from illumination-invariant reflectance, when fused with multi-scale appearance features by encoder-only depth-aware attention, yields low-light images that are both photometrically restored and geometrically consistent.

What carries the argument

The Multi-scale Depth Fusion (MDF) block, whose Depth-Aware Attention Fusion (DAAF) module uses cross-modal attention so depth queries reweight RGB features (and vice versa) only inside the encoder, thereby injecting geometric constraints without decoder-stage noise.

Load-bearing premise

The depth map produced by a frozen monocular model on Retinex reflectance is accurate enough, even in dark or textureless regions, to serve as a trustworthy geometric guide.

What would settle it

On a set of real low-light scenes with measured ground-truth depth, replace the estimated depth prior with random or inverted maps and check whether PSNR/SSIM/LPIPS and boundary fidelity still improve; if they do not, the claimed geometric benefit collapses.

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

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 / 0 minor

Summary. The paper proposes DMSA-Net, a supervised low-light image enhancement network that injects monocular depth as a geometric prior. A Retinex decomposition first yields an illumination-invariant reflectance map; a frozen DepthAnythingV2 model then estimates depth from that reflectance (Eq. 9). Depth is fused only in the encoder via Multi-scale Depth Fusion blocks that combine a Multi-branch Attention Module (MbAM/MFMA) with a Depth-Aware Attentional Feature Fusion (DAAF) cross-attention module (Eqs. 15–18) and residual dense blocks. The authors report competitive PSNR/SSIM/LPIPS on LOL-v1, LOL-v2 (real/synthetic), and SICEv2, competitive no-reference scores on DICM/LIME/MEF, progressive ablations (Tables 3–5), plug-and-play DAAF tests (Table 6), and release a depth-augmented LOL-D dataset.

Significance. If the depth prior is reliable and the fusion design is the main driver of the gains, the work is a useful incremental contribution to geometry-aware low-light enhancement: it couples Retinex reflectance with a modern monocular depth foundation model, provides a carefully designed encoder-only cross-modal attention fusion path, and ships LOL-D as a reusable resource. The progressive ablations and plug-and-play DAAF experiments are concrete strengths. Significance is tempered by the existence of prior depth-guided LLIE (notably Lin et al. [24]) and by the lack of quantitative depth-quality evidence, so the paper’s distinctive claim is currently more architectural and empirical than foundational.

major comments (3)
  1. Section 3.2.3, Eq. (9): The central geometry-aware claim rests on D = P(R(I_L)) with a frozen DepthAnythingV2 applied to Retinex reflectance. The manuscript asserts “robust, illumination-invariant” priors but only points to qualitative supplement comparisons with normal-light depth. No quantitative depth metrics (AbsRel, RMSE, δ<1.25, etc.) are reported on LOL-v1/v2 or SICEv2, nor is residual error characterized in textureless or residual-noise regions. Without this, Tables 3–5 show that “some depth signal helps” but do not establish that accurate scene geometry—not residual texture cues in R or generic multi-scale attention—is what drives the structural gains claimed for DAAF (Eqs. 15–18). Please add quantitative depth evaluation (and, if possible, a controlled noisy/biased-depth ablation) or temper the geometry-aware wording accordingly.
  2. Section 2.3 and contribution list: Lin et al. [24] already introduce hierarchical depth feature extraction/fusion for geometric-aware low-light enhancement. The present paper’s novelty relative to that line of work is not sharply delineated (Retinex-fronted depth estimation vs. their depth pipeline; encoder-only DAAF vs. their fusion modules; image-only vs. video). A direct quantitative and qualitative comparison to [24] (or a clear statement why it is inapplicable), plus an explicit novelty paragraph, is needed for the contribution claims (1)–(3) to hold at journal standard.
  3. Tables 1–2 and Section 4.2: Reported gains are often small (e.g., LOL-v2-real PSNR 22.032 vs. strong baselines in the low 21s; mixed ranking on LOL-v1 and synthetic SSIM). There are no error bars, multiple-run statistics, or significance tests, and several compared methods use different training protocols/data. For the claim of “superior restoration … while improving structural preservation,” please (i) report mean±std over seeds or folds where feasible, (ii) clarify training-data parity for each baseline, and (iii) add a structure-focused metric or analysis (edge/gradient fidelity, depth-consistency of enhanced outputs, or boundary IoU on depth discontinuities) beyond SSIM/LPIPS, which conflate appearance and structure.

Circularity Check

0 steps flagged

No significant circularity: supervised empirical architecture evaluated on external paired/unpaired benchmarks with frozen off-the-shelf depth estimator; no equation or claim reduces by construction to its own inputs.

full rationale

DMSA-Net is an end-to-end trainable encoder-decoder that obtains reflectance via a standard Retinex pretext (Eqs. 1-8), freezes an external monocular model (DepthAnythingV2) to produce depth (Eq. 9), and fuses via multi-scale attention (MbAM/DAAF, Eqs. 10-18) under ordinary supervised losses (Eq. 19) against ground-truth normal-light images. All quantitative claims (Tables 1-6) are measured by independent external metrics (PSNR/SSIM/LPIPS/NIQE/etc.) on public LOL/SICEv2/DICM/LIME/MEF splits; ablations simply ablate modules rather than re-label fitted parameters as predictions. The authors' LOL-D construction is an auxiliary contribution that does not enter the reported enhancement scores. No self-definitional loop, no fitted-input-as-prediction, and no load-bearing self-citation uniqueness theorem appears in the derivation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

The central claim rests on standard Retinex decomposition, a frozen foundation depth model, hand-chosen loss weights, and the architectural hypothesis that encoder-only cross-attention fusion of depth is superior. No new physical constants or particles are invented; free parameters are ordinary deep-learning hyper-parameters.

free parameters (5)
  • loss weights λ1..λ4
    Set by hand to 1.0, 0.5, 0.1, 0.3 (Eq. 19); directly control the trade-off among reconstruction, perceptual, SSIM and gradient terms that determine reported metrics.
  • Retinex cycle weights χ1, χ2
    Fixed at 0.1 and 0.5 (Eq. 8) without sensitivity analysis; affect quality of the reflectance fed to the depth estimator.
  • DAAF learnable scales η̂, μ̂
    Learned residual gates (Eq. 18) that balance depth-enhanced features against original RGB features; fitted during training.
  • Top-K sparsity parameters t1=3, t2=5
    Hard-coded sparsity levels inside MFMA (Eq. 12); control which attention scores survive.
  • learning rate, batch size, epochs, image size
    5e-5, batch 4, 200 epochs, 256×256; standard but still free choices that affect final numbers.
axioms (4)
  • domain assumption An image I equals reflectance R composed with illumination L (Retinex model, Eq. 1).
    Classic imaging model assumed throughout Section 3.2; not re-derived.
  • domain assumption Reflectance of the same scene under different lighting is identical (R1 = R2).
    Used to justify cycle-consistency training of the decomposition network.
  • ad hoc to paper DepthAnythingV2 applied to reflectance yields a usable geometric prior under low light.
    Core premise of the front-end module (Eq. 9); validated only qualitatively.
  • ad hoc to paper Encoder features are more compatible with depth priors than decoder features.
    Design choice justified by citation to Lin et al. and confirmed by ablation (Table 4).
invented entities (3)
  • Depth-Aware Attentional Feature Fusion (DAAF) module no independent evidence
    purpose: Cross-modal attention that injects depth geometry into RGB features without propagating depth noise as a value stream.
    New architectural block introduced in Section 3.3.2; no independent physical existence outside the network.
  • Multi-branch Attention Module (MbAM) with MFMA no independent evidence
    purpose: Extract multi-scale appearance features and perform sparse cross-scale modulation before depth fusion.
    New composite block (Fig. 4); engineering construct rather than a discovered natural entity.
  • LOL-D dataset (LOLv1-D / LOLv2-D) no independent evidence
    purpose: Provide paired low-light / normal-light images together with estimated depth maps for geometry-aware research.
    Constructed by the authors using their own depth pipeline; useful resource but dependent on the same depth estimator.

pith-pipeline@v1.1.0-grok45 · 26491 in / 3068 out tokens · 24404 ms · 2026-07-11T10:13:16.370954+00:00 · methodology

0 comments
read the original abstract

Low-light degradation reduces image visibility and weakens structural cues that are important for visual representation and scene understanding. Existing low-light image enhancement methods mainly focus on appearance restoration, while insufficiently exploiting scene geometry to preserve structural consistency. To address this limitation, this paper proposes a Depth-guided Multi-scale Attention Network (DMSA-Net) for geometry-aware low-light image enhancement. DMSA-Net introduces depth-related structural priors into low-light representation learning through reflectance-geometry interaction. A Retinex-based decomposition module is first used to obtain illumination-invariant reflectance representations, from which depth cues are inferred to characterize scene structure under degraded illumination. A multi-scale depth-guided fusion strategy is then embedded into a hierarchical encoder-decoder architecture, where depth-aware attention adaptively integrates geometric and appearance features. Experiments on several benchmark datasets show that DMSA-Net achieves effective low-light restoration while improving structural preservation. Moreover, we construct LOL-D, a depth-augmented low-light dataset, to facilitate research on geometry-aware low-light vision.

Figures

Figures reproduced from arXiv: 2607.05005 by Cheng Wu, Fang Gao, Hanbo Zheng, Jiabao Wang, Jingfeng Tang, Jiongkai Qin, Ming Cheng, Qingbao Huang.

Figure 2
Figure 2. Figure 2: Overview of the proposed Depth-aware Multi-scale Attentional Network (DMSA-Net) for low-light image enhancement. The network consists of a front-end Retinex-based decomposition module that separates the input image into reflectance and illumination components, followed by robust depth estimation from the illumination-invariant reflectance. Extracted depth maps are integrated with the RGB features produced … view at source ↗
Figure 3
Figure 3. Figure 3: An overview of the pretext task for extracting depth information. The task consists of two steps: (1) A Retinex-based decomposition network  is trained to obtain the reflectance component 𝑅and the illumination component 𝐿; (2) The network  parameters are frozen, and depth information with physical invariance is extracted from the 𝑅 component using the DepthAnythingV2 network, ultimately generating the de… view at source ↗
Figure 3
Figure 3. Figure 3: The above process can be described as follows: [PITH_FULL_IMAGE:figures/full_fig_p009_3.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

69 extracted references · 3 linked inside Pith

  1. [1]

    Learning multi -scale photo exposure correction,

    M. Afifi, K. G. Derpanis, B. Ommer , and M. S. Brown, “Learning multi -scale photo exposure correction,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Nashville, TN, USA, 2021, pp. 9157–9167

    2021

  2. [2]

    NeuTac: Zero-shot sim2real measurement for neuromorphic vision-based tactile sensors,

    M. Salah, et al., “NeuTac: Zero-shot sim2real measurement for neuromorphic vision-based tactile sensors,” IEEE Trans. Instrum. Meas., vol. 73, Sep. 2024, Art. no. 5031315

    2024

  3. [3]

    Image classification using convolution -al neural networks,

    M. A. Hasan et al., “Image classification using convolution -al neural networks,” Int. J. Mech. Eng. Res. Technol., vol. 16, no. 2, pp. 155 –162, Apr. 2024

    2024

  4. [4]

    A novel low light object detection method based on the YOLOv5 fusion feature enhancement,

    D. Peng, W. Ding, and T. Zhen, “A novel low light object detection method based on the YOLOv5 fusion feature enhancement,” Sci. Rep., vol. 14, no. 1, Feb. 2024, Art. no. 4486

    2024

  5. [5]

    Low-resolution self-attention for semantic segmentation,

    Y.-H. Wu et al., “Low-resolution self-attention for semantic segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 47, no. 9, pp. 8180 - 8192, Jun. 2025

    2025

  6. [6]

    A dynamic histogram equalization for image contrast enhancement,

    M. Abdullah-Al-Wadud, M. H. Kabir, M. A. A. Dewan, and O. Chae, “A dynamic histogram equalization for image contrast enhancement,” IEEE Trans. Consum. Electron., vol. 53, no. 2, pp. 593–600, May. 2007

    2007

  7. [7]

    Brightness preserving dynamic histogram equalization for image contrast enhancement,

    H. Ibrahim and N. S. P. Kong, “Brightness preserving dynamic histogram equalization for image contrast enhancement,” IEEE Trans. Consum. Electron., vol. 53, no. 4, pp. 1752–1758, Nov. 2007

    2007

  8. [8]

    The retinex theory of color vision,

    E. H. Land, “The retinex theory of color vision,” Sci. Am., vol. 237, no. 6, pp. 108–129, Dec. 1977

    1977

  9. [9]

    A probabilistic method for image enhancement with simultaneous illumination and reflectance estimation,

    X. Fu, Y. Liao, D. Zeng, Y. Huang, X. Zhang, and X. Ding, “A probabilistic method for image enhancement with simultaneous illumination and reflectance estimation,” IEEE Trans. Image Process., vol. 24, no. 12, pp. 4965–4977, Aug. 2015

    2015

  10. [10]

    Lime: Low-light image enhancement via illumination map estimation,

    X. Guo, Y. Li, and H. Ling, “Lime: Low-light image enhancement via illumination map estimation,” IEEE Trans. Image Process., vol. 26, no. 2, pp. 982–993, Feb. 2017. 25

    2017

  11. [11]

    Structure -revealing low-light image enhancement via robust retinex model,

    M. Li, J. Liu, W. Yang, X. Sun, and Z. Guo, “Structure -revealing low-light image enhancement via robust retinex model,” IEEE Trans. Image Process., vol. 27, no. 6, pp. 2828–2841, Feb. 2018

    2018

  12. [12]

    Low -light image enhancement using variational optimization-based retinex model,

    S. Park, S. Yu, B. Moon, S. Ko, and J. Paik, “Low -light image enhancement using variational optimization-based retinex model,” IEEE Trans. Consum. Electron., vol. 63, no. 2, pp. 178–184, May. 2017

    2017

  13. [13]

    Naturalness preserved enhancement algorithm for non -uniform illumination images,

    S. Wang, J. Zheng, H. -M. Hu, and B. Li, “Naturalness preserved enhancement algorithm for non -uniform illumination images,” IEEE Trans. Image Process., vol. 22, no. 9, pp. 3538–3548, Sep. 2013

    2013

  14. [14]

    Retinexformer: One -stage retinex -based transformer for low -light image enhancement,

    Y. Cai, H. Bian, J. Lin, H. Wang, R. Timofte, and Y. Zhang, “Retinexformer: One -stage retinex -based transformer for low -light image enhancement,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Paris, France, 2023, pp. 12504–12513

    2023

  15. [15]

    Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans,

    Y.-S. Chen, Y.-C. Wang, M.-H. Kao, and Y.-Y. Chuang, “Deep photo enhancer: Unpaired learning for image enhancement from photographs with gans,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Salt Lake City, UT, USA, 2018, pp. 6306–6314

    2018

  16. [16]

    SNRD-Net: SNR-aware dual enhancement network for low-light images,

    M. Z. U. Abideen et al., “SNRD-Net: SNR-aware dual enhancement network for low-light images,” Comput. Vis. Image Underst., vol. 268, Art. no. 104776, 2026

    2026

  17. [17]

    An efficient three-stage network via Multi-Scale Orthogonal Complementary Transformer for low-light image enhancement,

    J. Tan, S. Pei, L. Huang, and K. Cong, “An efficient three-stage network via Multi-Scale Orthogonal Complementary Transformer for low-light image enhancement,” Comput. Vis. Image Underst., vol. 268, Art. no. 104778, 2026

    2026

  18. [18]

    Progressive dual-branch network for low-light image enhancement,

    H. Cui, et al., “Progressive dual-branch network for low-light image enhancement,” IEEE Trans. Instrum. Meas., vol. 71, 2022, Art. no. 2520318

    2022

  19. [19]

    LLIE -Face: A multi-modal dataset for low-light facial image enhancement,

    H. Sun, D. Gao, P. He, X. Li, and F. Wang, “LLIE -Face: A multi-modal dataset for low-light facial image enhancement,” Comput. Vis. Image Underst., vol. 263, Art. no. 104576, 2026

    2026

  20. [20]

    SNR -aware low-light image enhancement,

    X. Xu, R. Wang, C. -W. Fu, and J. Jia, “SNR -aware low-light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), New Orleans, LA, USA, 2022, pp. 17714–17724

    2022

  21. [21]

    JSF: A joint spatial-frequency domain network for low-light image enhancement,

    Y. Wu, F. Liu, and R. Wang, “JSF: A joint spatial-frequency domain network for low-light image enhancement,” Comput. Vis. Image Underst., vol. 261, Art. no. 104496, 2025

    2025

  22. [22]

    Towards Scale -Aware Low-Light Enhancement via Structure -Guided Transformer Design,

    W. Dong et al., “Towards Scale -Aware Low-Light Enhancement via Structure -Guided Transformer Design,” in Proc. Comput. Vis. Pattern Recognit. Conf. (CVPR), Nashville City, TN, USA, 2025, pp. 1460-1470

    2025

  23. [23]

    Low-light image enhancement via generative perceptual priors,

    H. Zhou et al., “Low-light image enhancement via generative perceptual priors,” in Proc. AAAI Conf. Artif. Intell. (AAAI), vol. 39, no. 10, 2025, pp. 10752-10760

    2025

  24. [24]

    Geometric -aware low-light image and video enhancement via depth guidance,

    Y. Lin et al., “Geometric -aware low-light image and video enhancement via depth guidance,” IEEE Trans. Image Process., vol. 34, pp. 5442 - 5457, Aug. 2025

    2025

  25. [25]

    Contrast enhancement based on layered difference representation of 2d histograms,

    C. Lee, C. Lee, and C.-S. Kim, “Contrast enhancement based on layered difference representation of 2d histograms,” IEEE Trans. Image Process., vol. 22, no. 12, pp. 5372–5384, Sep. 2013

    2013

  26. [26]

    Lightennet: a convolutional neural network for weakly illuminated image enhancement,

    C. Li, J. Guo, F. Porikli, and Y. Pang, “Lightennet: a convolutional neural network for weakly illuminated image enhancement,” Pattern Recognit. Lett., vol. 104, pp. 15–22, Mar. 2018

    2018

  27. [27]

    Contextual and variational contrast enhancement,

    T. Celik and T. Tjahjadi, “Contextual and variational contrast enhancement,” IEEE Trans. Image Process., vol. 20, no. 12, pp. 3431–3441, Dec. 2011. 26

    2011

  28. [28]

    A histogram modification framework and its application for image contrast enhancement,

    T. Arici, S. Dikbas, and Y. Altunbasak, “A histogram modification framework and its application for image contrast enhancement,” IEEE Trans. Image Process., vol. 18, no. 9, pp. 1921–1935, Sep. 2009

    1921

  29. [29]

    Efficient contrast enhancement using adaptive gamma correction with weighting distribution,

    S.-C. Huang, F.-C. Cheng, and Y.-S. Chiu, “Efficient contrast enhancement using adaptive gamma correction with weighting distribution,” IEEE Trans. Image Process., vol. 22, no. 3, pp. 1032–1041, Oct. 2012

    2012

  30. [30]

    An adaptive gamma correction for image enhancement,

    S. Rahman, M. M. Rahman, M. A. -A. Wadud, G. D. Al -Quaderi, and M. Shoyaib, “An adaptive gamma correction for image enhancement,” EURASIP J. Image Video Process., 2016, article number 35, Oct. 2016

    2016

  31. [31]

    Osmamba: Omnidirectional spectral mamba with dual-domain prior generator for exposure correction,

    G. Li, B. Chen, C. Zhao, et al., “Osmamba: Omnidirectional spectral mamba with dual-domain prior generator for exposure correction,” in Proc. Comput. Vis. Pattern Recognit. Conf. (CVPR), Nashville City, TN, USA, 2025, pp. 7480–7490

    2025

  32. [32]

    A weighted variational model for simultaneous reflectance and illumination estimation,

    X. Fu, D. Zeng, Y. Huang, X. Zhang, and X. Ding, “A weighted variational model for simultaneous reflectance and illumination estimation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Las Vegas, NV, USA, 2016, pp. 2782–2790

    2016

  33. [33]

    Star: A structure and texture aware retinex model,

    J. Xu, Y. Hou, D. Ren, L. Liu, F. Zhu, M. Yu, H. Wang, and L. Shao, “Star: A structure and texture aware retinex model,” IEEE Trans. Image Process., vol. 29, pp. 5022–5037, Mar. 2020

    2020

  34. [34]

    Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,

    R. Liu, L. Ma, J. Zhang, et al., “Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Nashville, TN, USA, 2021, pp. 10561–10570

    2021

  35. [35]

    Low -light image enhancement with normalizing flow,

    Y. Wang, R. Wan, W. Yang, et al., “Low -light image enhancement with normalizing flow,” in Proc. AAAI Conf. Artif. Intell. (AAAI), 2022, vol. 36, no. 3, pp. 2604–2612

    2022

  36. [36]

    Real -time exposure correction via collaborative transformations and adaptive sampling,

    Z. Li, F. Zhang, M. Cao, et al., “Real -time exposure correction via collaborative transformations and adaptive sampling,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle City, WA, USA, 2024, pp. 2984–2994

    2024

  37. [37]

    Zero-reference deep curve estimation for low-light image enhancement,

    C. Guo, C. Li, J. Guo, et al., “Zero-reference deep curve estimation for low-light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle City, WA, USA, 2020, pp. 1780–1789

    2020

  38. [38]

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

    C. Li, C. Guo, and C. C. Loy, “Learning to enhance low-light image via zero-reference deep curve estimation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 8, pp. 4225–4238, Mar. 2021

    2021

  39. [39]

    Beyond brightening low-light images,

    Y. Zhang, X. Guo, J. Ma, et al., “Beyond brightening low-light images,” Int. J. Comput. Vis., vol. 129, no. 4, pp. 1013–1037, Jan. 2021

    2021

  40. [40]

    LLNet: A deep autoencoder approach to natural low-light image enhancement,

    K. G. Lore, A. Akintayo, and S. Sarkar, “LLNet: A deep autoencoder approach to natural low-light image enhancement,” Pattern Recognit., vol. 61, pp. 650–662, Jan. 2017

    2017

  41. [41]

    Deep retinex decomposition for low-light enhancement,

    C. Wei et al., “Deep retinex decomposition for low-light enhancement,” arXiv preprint arXiv:1808.04560, 2018

    Pith/arXiv arXiv 2018

  42. [42]

    MBLLEN: Low-Light Image/Video Enhancement Using CNNs,

    F. Lv et al., “MBLLEN: Low-Light Image/Video Enhancement Using CNNs,” in Br. Mach. Vis. Conf., London, UK, 2018, p. 4

    2018

  43. [43]

    Difflight: integrating content and detail for low -light image enhancement,

    Y. Feng et al., “Difflight: integrating content and detail for low -light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle City, WA, USA, 2024, pp. 6143-6152

    2024

  44. [44]

    Deep color consistent network for low -light image enhancement,

    Z. Zhang et al., “Deep color consistent network for low -light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), New Orleans City, LA, USA, 2022, pp. 1899-1908

    2022

  45. [45]

    Learning a simple low -light image enhancer from paired low -light instances,

    Z. Fu et al., “Learning a simple low -light image enhancer from paired low -light instances,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Vancouver City, BC, CA, 2023, pp. 22252-22261. 27

    2023

  46. [46]

    Cycle -retinex: Unpaired low-light image enhancement via retinex-inline cyclegan,

    K. Wu et al., “Cycle -retinex: Unpaired low-light image enhancement via retinex-inline cyclegan,” IEEE Trans. Multimed., vol. 26, pp. 1213 – 1228, May. 2023

    2023

  47. [47]

    Toward fast, flexible, and robust low-light image enhancement,

    L. Ma et al., “Toward fast, flexible, and robust low-light image enhancement,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), New Orleans City, LA, USA, 2022, pp. 5637-5646

    2022

  48. [48]

    Sparse gradient regularized deep retinex network for robust low -light image enhancement,

    W. Yang et al., “Sparse gradient regularized deep retinex network for robust low -light image enhancement,” IEEE Trans. Image Process., vol. 30, pp. 2072–2086, Jan. 2021

    2072

  49. [49]

    Darkvisionnet: Low-light imaging via RGB-NIR fusion with deep inconsistency prior,

    S. Jin et al., “Darkvisionnet: Low-light imaging via RGB-NIR fusion with deep inconsistency prior,” in Proc. AAAI Conf. Artif. Intell. (AAAI), vol. 36, no. 1, 2022, pp. 1104-1112

    2022

  50. [50]

    Implicit neural representation for cooperative low-light image enhancement,

    S. Yang et al., “Implicit neural representation for cooperative low-light image enhancement,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Paris, FR, 2023, pp. 12918-12927

    2023

  51. [51]

    Low-light image enhancement via structure modeling and guidance,

    X. Xu, R. Wang, and J. Lu, “Low-light image enhancement via structure modeling and guidance,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Vancouver City, BC, CA, 2023, pp. 9893-9903

    2023

  52. [52]

    Depth anything v2,

    L. Yang et al., “Depth anything v2,” in Adv. Neural Inf. Process. Syst., vol. 37, 2024, pp. 21875–21911

    2024

  53. [53]

    UniDepth: Universal monocular metric depth estimation,

    L. Piccinelli et al., “UniDepth: Universal monocular metric depth estimation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle City, WA, USA, 2024, pp. 10106-10116

    2024

  54. [54]

    Depth pro: Sharp monocular metric depth in less than a second,

    A. Bochkovskii et al., “Depth pro: Sharp monocular metric depth in less than a second,” arXiv preprint arXiv:2410.02073, 2024

    Pith/arXiv arXiv 2024

  55. [55]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei -Fei, “Perceptual losses for real-time style transfer and super-resolution,” in Eur. Conf. Comput. Vis. (ECCV), Cham, Switzerland: Springer Int. Publ., 2016, pp. 694–711

    2016

  56. [56]

    Memory-efficient Transformers via Top-k Attention,

    A. Gupta et al., “Memory-efficient Transformers via Top-k Attention,” arXiv preprint arXiv:2106.06899, 2021

    Pith/arXiv arXiv 2021

  57. [57]

    Transfer learning using VGG-16 with deep convolutional neural network for classifying images,

    T. S. Tammina, “Transfer learning using VGG-16 with deep convolutional neural network for classifying images,” Int. J. Sci. Res. Publ. (IJSRP), vol. 9, no. 10, pp. 143–150, Oct. 2019

    2019

  58. [58]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004

    2004

  59. [59]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, et al., “The unreasonable effectiveness of deep features as a perceptual metric,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Salt Lake City, UT, USA, 2018, pp. 586–595

    2018

  60. [60]

    Enlightengan: Deep light enhancement without paired supervision,

    Y. Jiang, X. Gong, D. Liu, et al., “Enlightengan: Deep light enhancement without paired supervision,” IEEE Trans. Image Process., vol. 30, pp. 2340–2349, Jan. 2021

    2021

  61. [61]

    Kindling the darkness: A practical low-light image enhancer,

    Y. Zhang, J. Zhang, and X. Guo, “Kindling the darkness: A practical low-light image enhancer,” in Proc. 27th ACM Int. Conf. Multimedia, Nice City, FR, 2019, pp. 1632–1640

    2019

  62. [62]

    Llformer: An efficient and real -time lidar lane detection method based on transformer,

    J. He, X. Zuo, J. Gao, et al., “Llformer: An efficient and real -time lidar lane detection method based on transformer,” in Proc. 5th Int. Conf. Pattern Recognition and Intelligent Systems, 2023, pp. 18–23

    2023

  63. [63]

    Glow in the dark: Low-light image enhancement with external memory,

    D. Ye, Z. Ni, W. Yang, et al., “Glow in the dark: Low-light image enhancement with external memory,” IEEE Trans. Multimedia, vol. 26, pp. 2148–2163, 2023. 28

    2023

  64. [64]

    A review on computational low -light image enhancement models: Challenges, benchmarks, and perspectives,

    P. Singh and A. K. Bhandari, “A review on computational low -light image enhancement models: Challenges, benchmarks, and perspectives,” Arch. Comput. Methods Eng., vol. 32, no. 5, pp. 2853–2885, 2025

    2025

  65. [65]

    FIPNet: Self-supervised low-light image enhancement combining feature and illumination priors,

    J. Cao, et al., “FIPNet: Self-supervised low-light image enhancement combining feature and illumination priors,” Neurocomputing, vol. 623, Art. no. 129426, 2025

    2025

  66. [66]

    Making a ‘completely blind’ image quality analyzer,

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a ‘completely blind’ image quality analyzer,” IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212, Mar. 2013

    2013

  67. [67]

    No -reference image quality assessment in the spatial domain,

    A. Mittal, A. K. Moorthy, and A. C. Bovik, “No -reference image quality assessment in the spatial domain,” IEEE Trans. Image Process ., vol. 21, no. 12, pp. 4695–4708, Dec. 2012

    2012

  68. [68]

    A fusion -based enhancing method for weak illumination images,

    Z. Ying, G. Li, Y. Ren, R. Wang, and W. Wang, “A fusion -based enhancing method for weak illumination images,” in Proc. IEEE Int. Conf. Image Processing (ICIP), Beijing, China, 2017, pp. 1–5

    2017

  69. [69]

    An underwater color image quality evaluation metric,

    M. Yang and A. Sowmya, “An underwater color image quality evaluation metric,” IEEE Trans. Image Process., vol. 24, no. 12, pp. 6062–6071, Dec. 2015

    2015