Pith. sign in

REVIEW 4 major objections 5 minor 44 references

CMIG-Net uses conditional mutual information to quantify chrominance's extra contribution beyond intensity, outperforming CIDNet by up to 0.619 dB PSNR on low-light benchmarks.

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 ↗

CMIG-Net adaptively fuses chrominance and intensity features with a learned conditional mutual information map, improving PSNR by up to 0.6 dB over CIDNet.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Incremental LLIE gains from a dual-branch architecture, but the CMI framing is unsupported: L_p-fit trains the intensity-only reference against the full-information distribution, collapsing the very quantity the paper claims to measure. the 4 major comments →

arxiv 2608.01886 v1 pith:OKY3JMNU submitted 2026-08-03 cs.CV

Beyond Illumination: A Conditional Mutual Information-Guided Network for Low-Light Image Enhancement

Ya-nan Guan , Shaonan Zhang , Tao Dai , Tianqu Zhuang , Yongchao Qiao , Zhensen Chen , Shu-Tao Xia , Hang Guo This is my paper
classification cs.CV
keywords low-light image enhancementconditional mutual informationHVI color spacedual-branch networkchrominance-intensity interactioncross-branch attentionimage restorationKL divergence
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 argues that in low-light image enhancement, chrominance is not merely a supplement to intensity: it carries information that intensity alone cannot provide, and the size of that contribution changes from region to region. The authors measure this contribution as a conditional mutual information map—the KL divergence between a full-information cluster predictor and an intensity-only reference predictor—and feed this map into two new modules. One recalibrates the chrominance input at each location; the other dynamically gates how much intensity and chrominance features exchange across restoration stages. On paired benchmarks, the resulting network outperforms the current state-of-the-art dual-branch method by up to 0.619 dB PSNR, including a 0.382 dB gain on the Sony-Total-Dark dataset. The result suggests that modeling cross-component dependence explicitly, rather than decoupling them, is a more effective design principle for LLIE.

Core claim

The central claim is that the additional contribution of chrominance (HV) given intensity (I) can be quantified per pixel by I(HV; C | I), estimated as D_KL(q(C|HV,I) || p(C|I)), where q and p are learned softmax predictors over a latent cluster variable C. Using this quantity, the paper constructs a Conditional Mutual Information Calibration module that recalibrates the HV input by a scale derived from the cluster-soft-routed calibration strength, and a Dynamic Dual-branch Information Restoration module whose Conditional Information-Guided Attention uses the CMI map together with current branch states to modulate both the value and output of cross-branch attention. The paper reports that th

What carries the argument

The central object is the conditional mutual information map M_cmi(u) = D_KL(q_phi(C|HV,I) || p_psi(C|I)) computed at every spatial location, together with the two modules it drives: CMIC, which uses M_cmi and the cluster distribution Q to produce a spatially varying calibration scale for HV, and D2IR with CIGA, which concatenates M_cmi with channel-averaged branch features to generate per-stage and per-location gating coefficients for cross-branch value and output modulation. The KL map is what converts an abstract information-theoretic quantity into actionable, spatially adaptive control signals for the restoration network.

Load-bearing premise

The intensity-only reference distribution p(C|I) is a faithful baseline of what can be inferred from intensity alone, despite being trained to mimic the full-information distribution Q.

What would settle it

On an image where all chrominance is artificially removed (HV set to neutral gray), the CMI map should be near zero everywhere; if it is not, the map does not measure chrominance's contribution.

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

If this is right

  • If correct, the CMI map gives a principled, interpretable way to decide where chrominance should be trusted or suppressed, replacing fixed heuristic injection strengths.
  • The D2IR gating mechanism generalizes the dual-branch HVI architecture to a controllable information-flow setting, which could be extended to other component-decoupled restoration tasks such as reflectance/illumination or detail/base decomposition.
  • The L_p-fit reference fitting loss stabilizes the intensity-only baseline, making the CMI map reliable during training; this training strategy is a reusable recipe for any CMI-style network.
  • The reported gains on unpaired datasets suggest the learned conditional prior transfers beyond paired supervision.

Where Pith is reading between the lines

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

  • Our reading: because the intensity-only reference P is trained with L_p-fit to imitate the full-information Q, the KL map may partly reflect how much capacity the intensity branch lacks rather than pure chrominance information; a variant that restricts P's architecture or training objective would distinguish these.
  • A direct falsifier already accessible in any paired dataset: replace HV with neutral gray and recompute M_cmi; if high values persist, the map is not actually tied to chrominance content.
  • The conditional-prior idea transfers: any two-stream restoration network (e.g., reflectance/illumination or detail/base decomposition) could use a similar per-pixel CMI map between streams to decide where and when to exchange information.
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

4 major / 5 minor

Summary. The paper proposes CMIG-Net for low-light image enhancement in the HVI color space. Its central claim is that the additional contribution of chrominance (HV) given intensity (I) can be quantified as a conditional mutual information I(HV;C|I), approximated by a per-pixel KL divergence M_cmi between two learned distributions q(C|HV,I) and p(C|I). This map is used in a CMIC module to calibrate chrominance and in D2IR modules to dynamically gate cross-branch information. Extensive paired benchmarks (LOLv1, LOLv2-Real, LOLv2-Synthetic, SICE, SID) and unpaired datasets report consistent gains over CIDNet, e.g., 28.256 vs 27.890 dB PSNR on LOLv1 and 23.286 vs 22.904 dB on SID.

Significance. The paper has clear empirical strengths: careful paired and unpaired evaluations, a systematic ablation, and consistent PSNR/LPIPS improvements over the strong CIDNet baseline. The proposed architecture may be useful as an empirical dual-branch gating mechanism. However, the scientific contribution rests entirely on the claim that M_cmi is a principled conditional mutual information measure. That claim is not supported by the implementation. The reference distribution in Eq. (13) is trained to match the full-information distribution, directly contradicting the required semantics of p(C|I); and Eq. (6) replaces the CMI expectation with a per-pixel KL divergence between learned softmax outputs. The reported gains therefore cannot be attributed to conditional mutual information guidance in the information-theoretic sense.

major comments (4)
  1. [§3, Eq. (6)] Equation (2) states the standard definition of CMI as an expectation over the joint distribution. Equation (6) replaces this expectation with a per-location KL divergence between the softmax outputs of two learned networks. No argument is given for why the spatial average of this quantity estimates I(HV;C|I), nor for why the latent cluster variable C has the required semantics. As written, M_cmi is a heuristic discrepancy map, not a conditional mutual information. This is load-bearing because every subsequent module uses M_cmi as an information-theoretic prior.
  2. [§4, Eq. (13)] The reference distribution P is optimized by minimizing KL(Q||P) with Q = q(C|HV,I). If this loss is effective, P approaches Q and M_cmi = KL(Q||P) is driven toward zero; if it is ineffective, P is an arbitrary output of the same network. In neither case is P a faithful intensity-only reference. The text's assertion that 'P faithfully captures the cluster states that can be inferred from intensity alone when chrominance information is unavailable' is contradicted by the loss definition, since Q is conditioned on HV. Table 4 shows that L_p-fit changes PSNR, but this only demonstrates that the regularizer affects optimization; it does not validate the CMI interpretation.
  3. [§3, Figure 2] The empirical motivation is not independently validated. Figure 2 reports 'conditional contribution modeling' and 'cluster-wise contribution statistics' but gives no estimation procedure, no ground-truth information-theoretic quantity, and no comparison with a direct CMI estimate. Both q and p are learned by the same model, and p is trained to match q, so the clusters and statistics are model constructs rather than measurements of image statistics. This weakens the claimed motivation for the CMI metric.
  4. [§5, Table 4] The ablations do not test whether M_cmi actually quantifies chrominance contribution. They only test whether adding the KL-based term and calibration changes PSNR/SSIM/LPIPS. A direct validation would require, for example, demonstrating that M_cmi is invariant when HV is perturbed in ways that do not affect C, or comparing M_cmi against an independent non-learned estimate of I(HV;C|I). Without such evidence, the paper's central scientific framing is unsupported even though the architecture may still be an effective empirical method.
minor comments (5)
  1. [§1, Figure 1] The sentence 'as Figure 1)' has a typo; should read 'as shown in Figure 1'.
  2. [§3, Eq. (1)] The notation 'HV I' is not defined explicitly; it should be written as the joint variable [HV, I] for clarity.
  3. [§4, Eq. (13)] The hyperparameters K, alpha, and lambda_fit are introduced without specifying their values or a sensitivity analysis. Since the CMI map depends on K and the calibration strength depends on alpha, this omission is notable.
  4. [§5, Table 2] On SICE, CMIG-Net has a lower PSNR than CIDNet (13.207 vs 13.435), although the text correctly states that it achieves the highest SSIM. This partial trade-off should be acknowledged in the main text rather than only in the table.
  5. [General] The paper does not mention code availability. Given the empirical nature of the contributions, releasing the implementation would strengthen reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the CMI map is trained via a standard variational projection, and the central claim is externally benchmarked.

full rationale

The paper's central derivation is self-contained and not circular. Eq. 6 defines the CMI map as the KL divergence Mcmi = KL(q(C|HV,I) || p(C|I)), and Eq. 13 defines L_p-fit as the batch-averaged KL(Q||P). Although L_p-fit uses Q as the target for P, P is constrained by Eq. 5 to be a function of intensity only: P = Softmax(f_p(I)). Therefore, minimizing L_p-fit over P does not drive P to equal Q at each pixel; it drives P to the conditional marginal E_{HV|I}[Q(C|HV,I)], and the minimum of the loss is exactly the conditional mutual information. Thus Mcmi does not collapse to zero; it estimates the intended quantity provided Q is a faithful full-information model. The method is empirically validated against external benchmarks (LOL, SICE, SID, etc.), so the performance claim is not forced by construction. The only self-citation is to Zhuang et al. (2025) for a standard chain-rule identity, which is not load-bearing. The paper's wording in Sec. 4 ('aligns the reference distribution P with Q') is loose, but the formal equations clarify that P is the intensity-only projection of Q, not Q itself. Overall, no prediction reduces to a fitted parameter or to a self-citation chain; any concern about the semantic validity of the learned latent cluster variable C is an assumption about the method's interpretability, not a circularity in the derivation.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central claim rests on a learned latent variable C and on treating a per-pixel KL divergence as conditional mutual information. These are not independently validated. The loss L_p-fit explicitly aligns the intensity-only reference P with the full-information Q, so the 'measurement' is fitted rather than principled.

free parameters (3)
  • K
    Number of latent cluster states for the conditional-information variable C; a hyperparameter set by hand, not reported in the main text.
  • alpha
    Maximum modulation amplitude in the CMIC calibration scale s(u)=1+alpha*delta(u); hand-chosen hyperparameter.
  • lambda_fit
    Weight of the reference distribution fitting loss L_p-fit in Eq. 14; hand-chosen.
axioms (5)
  • standard math Chain rule of mutual information I(HV;C|I)=I(HV I;C)-I(I;C)
    Used in Eq. 1 to decompose conditional mutual information; valid.
  • domain assumption HVI color space decomposition faithfully separates intensity and chrominance
    Adopted from CIDNet (Yan et al. 2025); the paper builds on this without re-justifying.
  • ad hoc to paper Learnable proxies q and p accurately approximate the true conditional distributions
    Eq. 3 introduces q and p as approximations; no assurance they are calibrated to the true conditionals.
  • ad hoc to paper Pixel-wise KL divergence approximates the conditional mutual information
    Eqs. 4-6 replace an expectation over the joint distribution with a per-pixel KL sum; this is not a valid CMI estimator.
  • domain assumption The latent cluster variable C captures information relevant to restoration
    C is not tied to any ground-truth clustering or task variable; its states are emergent.
invented entities (2)
  • Conditional information cluster variable C no independent evidence
    purpose: Latent variable whose distribution, conditioned on HV and I, is used to define the CMI map; guides calibration and gating.
    No external validation; C is a learned per-pixel categorical variable.
  • CMI map Mcmi no independent evidence
    purpose: Spatially varying weight for chrominance calibration and cross-branch attention.
    Computed from the network's own learned distributions; no ground-truth information value.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Illumination: A Conditional Mutual Information-Guided Network for Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/OKY3JMNU

@misc{pith2026260801886,
  author       = {Pith},
  title        = {Pith review of: Beyond Illumination: A Conditional Mutual Information-Guided Network for Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OKY3JMNU}},
  note         = {Machine review of arXiv:2608.01886}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Low-light image enhancement (LLIE) seeks to restore structural fidelity, natural color rendition, and proper exposure from images captured under inadequate lighting conditions. Recent state-of-the-art approaches, such as CIDNet, adopt a dual-branch architecture comprising a chrominance (HV) branch and an intensity (I) branch to separately model decoupled chromatic and luminance information within the HVI color space. However, these methods overlook the mutual interaction between intensity and chrominance components, which inherently limits their representational capacity and leads to suboptimal enhancement performance. To address this limitation, we propose the Conditional Mutual Information-Guided Network (CMIG-Net), which leverages conditional mutual information as a principled metric to quantitatively assess the contribution of chrominance features conditioned on the available intensity information. In particular, we design a Conditional Mutual Information Calibration (CMIC) module that generates a conditional information map, enabling region-adaptive recalibration of chrominance representations according to local illumination statistics. Furthermore, we introduce a Dynamic Dual-branch Information Restoration (D2IR) module, which adaptively governs bidirectional information flow between the intensity and chrominance branches, guided by both the conditional prior and the instantaneous restoration state. Extensive experiments on paired LLIE benchmarks demonstrate that CMIG-Net consistently outperforms CIDNet, achieving up to a 0.619 dB gain in PSNR, with a 0.382 dB improvement specifically on the challenging Sony-Total-Dark dataset.

Figures

Figures reproduced from arXiv: 2608.01886 by Hang Guo, Shaonan Zhang, Shu-Tao Xia, Tao Dai, Tianqu Zhuang, Ya-nan Guan, Yongchao Qiao, Zhensen Chen.

Figure 1
Figure 1. Figure 1: Effects of HV chrominance injection. Weak HV [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical analysis of chrominance contribution in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of our CMIG-Net, which mainly consists of the Conditional Mutual Information Calibration [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of color and tone restoration by different LLIE methods on the LOL dataset. The numbers [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of detail recovery and noise suppression by different LLIE methods on the LOL dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual analysis of conditional information cali [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparisons on unpaired low-light [PITH_FULL_IMAGE:figures/full_fig_p007_6.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

44 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Proceedings of the 27th ACM international conference on multimedia , pages=

    Kindling the darkness: A practical low-light image enhancer , author=. Proceedings of the 27th ACM international conference on multimedia , pages=

  2. [2]

    British Machine Vision Conference , year=

    Deep Retinex Decomposition for Low-Light Enhancement , author =. British Machine Vision Conference , year=

  3. [3]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Retinexformer: One-stage retinex-based transformer for low-light image enhancement , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Zero-reference deep curve estimation for low-light image enhancement , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Low-light image enhancement with normalizing flow , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  6. [6]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Snr-aware low-light image enhancement , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  7. [7]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Learning a simple low-light image enhancer from paired low-light instances , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  8. [8]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Global structure-aware diffusion process for low-light image enhancement , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Cwnet: Causal wavelet network for low-light image enhancement , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  11. [11]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  12. [12]

    Proceedings of the computer vision and pattern recognition conference , pages=

    Hvi: A new color space for low-light image enhancement , author=. Proceedings of the computer vision and pattern recognition conference , pages=

  13. [13]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Darkir: Robust low-light image restoration , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  14. [14]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  15. [15]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    ICLR: Inter-Chrominance and Luminance Interaction for Natural Color Restoration in Low-Light Image Enhancement , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  16. [16]

    Proceedings of the 31st ACM international conference on multimedia , pages=

    Mutual information-driven triple interaction network for efficient image dehazing , author=. Proceedings of the 31st ACM international conference on multimedia , pages=

  17. [17]

    1999 , publisher=

    Elements of Information Theory , author=. 1999 , publisher=

    1999

  18. [18]

    arXiv preprint arXiv:1808.06670 , year=

    Learning deep representations by mutual information estimation and maximization , author=. arXiv preprint arXiv:1808.06670 , year=

    Pith/arXiv arXiv

  19. [19]

    Frontiers in Neuroscience , volume=

    Mutual information measure of visual perception based on noisy spiking neural networks , author=. Frontiers in Neuroscience , volume=. 2023 , publisher=

    2023

  20. [20]

    Information and Control , volume=

    A definition of conditional mutual information for arbitrary ensembles , author=. Information and Control , volume=. 1978 , publisher=

    1978

  21. [21]

    2025 IEEE International Symposium on Information Theory (ISIT) , pages=

    Conditional mutual information based diffusion posterior sampling for solving inverse problems , author=. 2025 IEEE International Symposium on Information Theory (ISIT) , pages=. 2025 , organization=

    2025

  22. [22]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Conditional mutual information constrained deep learning for classification , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  23. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Learning degradation-independent representations for camera isp pipelines , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [24]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    InfoBridge: Balanced Multimodal Integration through Conditional Dependency Modeling , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  25. [25]

    The Thirteenth International Conference on Learning Representations , year=

    Stealthy shield defense: A conditional mutual information-based approach against black-box model inversion attacks , author=. The Thirteenth International Conference on Learning Representations , year=

  26. [26]

    Pattern Recognition , volume=

    LLNet: A deep autoencoder approach to natural low-light image enhancement , author=. Pattern Recognition , volume=. 2017 , publisher=

    2017

  27. [27]

    IEEE transactions on image processing , volume=

    Enlightengan: Deep light enhancement without paired supervision , author=. IEEE transactions on image processing , volume=. 2021 , publisher=

    2021

  28. [28]

    Computers & Graphics , pages=

    HSVNet: A dual-branch collaborative network for low-light image enhancement in the HSV color space , author=. Computers & Graphics , pages=. 2025 , publisher=

    2025

  29. [29]

    British Machine Vision Conference 2018,

    Chen Wei and Wenjing Wang and Wenhan Yang and Jiaying Liu , title =. British Machine Vision Conference 2018,. 2018 , url =

    2018

  30. [30]

    IEEE Transactions on Image Processing , volume=

    Sparse gradient regularized deep retinex network for robust low-light image enhancement , author=. IEEE Transactions on Image Processing , volume=. 2021 , publisher=

    2021

  31. [31]

    IEEE transactions on image processing , volume=

    Learning a deep single image contrast enhancer from multi-exposure images , author=. IEEE transactions on image processing , volume=. 2018 , publisher=

    2018

  32. [32]

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

    Shen Zheng and Yiling Ma and Jinqian Pan and Changjie Lu and Gaurav Gupta , title =. CoRR , volume =. 2022 , url =. doi:10.48550/ARXIV.2212.10772 , eprinttype =. 2212.10772 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.10772 2022

  33. [33]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Learning to see in the dark , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  34. [34]

    IEEE transactions on image processing , volume=

    Contrast enhancement based on layered difference representation of 2D histograms , author=. IEEE transactions on image processing , volume=. 2013 , publisher=

    2013

  35. [35]

    IEEE Transactions on image processing , volume=

    LIME: Low-light image enhancement via illumination map estimation , author=. IEEE Transactions on image processing , volume=. 2016 , publisher=

    2016

  36. [36]

    IEEE Transactions on Image Processing , volume=

    Perceptual quality assessment for multi-exposure image fusion , author=. IEEE Transactions on Image Processing , volume=. 2015 , publisher=

    2015

  37. [37]

    IEEE transactions on image processing , volume=

    Naturalness preserved enhancement algorithm for non-uniform illumination images , author=. IEEE transactions on image processing , volume=. 2013 , publisher=

    2013

  38. [38]

    Multimedia Tools and Applications , volume=

    On the evaluation of illumination compensation algorithms , author=. Multimedia Tools and Applications , volume=. 2018 , publisher=

    2018

  39. [39]

    Kingma and Jimmy Ba , editor =

    Diederik P. Kingma and Jimmy Ba , editor =. Adam:. 3rd International Conference on Learning Representations,. 2015 , url =

    2015

  40. [40]

    2017 , url=

    Ilya Loshchilov and Frank Hutter , booktitle=. 2017 , url=

    2017

  41. [41]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  42. [42]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  43. [43]

    completely blind

    Making a “completely blind” image quality analyzer , author=. IEEE Signal processing letters , volume=. 2012 , publisher=

    2012

  44. [44]

    IEEE Transactions on image processing , volume=

    No-reference image quality assessment in the spatial domain , author=. IEEE Transactions on image processing , volume=. 2012 , publisher=

    2012

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.