Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Low-Light Enhancement via Encoder-Decoder Network with Illumination Guidance

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read EDNIG, a U-Net guided by a Bright Channel Prior illumination map, reports the best PSNR and SSIM on the LOL dataset among thirteen compared methods while using only 1.737 million parameters.

desk verdict A compact, plausible low-light enhancer that adapts the authors' dehazing network, but the 'surpasses all' claim rests on a cross-publication table with thin margins and an unstated test-time resolution. read the letter →

arxiv 2507.13360 v1 pith:BZGQ4EAI submitted 2025-07-04 cs.CV

Le-Anh Tran , Chung Nguyen Tran , Ngoc-Luu Nguyen , Nhan Cach Dang , Jordi Carrabina , David Castells-Rufas , Minh Son Nguyen This is my paper · ORCID
classification cs.CV
keywords low-lightimageenhancementilluminationguidanceBrightChannelPriorU-Netencoder-decodernetworkgenerativeadversarialSpatialPyramidPoolingLOLdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes EDNIG, a U-Net-style encoder-decoder network that enhances low-light images by taking the dark RGB image together with an estimated illumination map as its input. The illumination map, computed from the Bright Channel Prior and smoothed by guided filtering, marks where the scene is underexposed and steers the network's corrections toward those regions. The architecture adds multi-scale Spatial Pyramid Pooling, Swish activations, and a GAN-like composite loss combining adversarial, pixel-wise MSE, and perceptual terms. The paper reports that EDNIG reaches a PSNR of 21.512 and an SSIM of 0.8313 on the LOL dataset, the highest among the thirteen methods compared, while using only 1.737M parameters and running in 0.042 seconds. The paper concludes that a compact, physically guided encoder-decoder can hold its own against much larger restoration models, which matters for real-world deployment on constrained devices.

What carries the argument

The load-bearing object is the illumination guidance map $M$, defined by the Bright Channel Prior as the maximum intensity across the R, G, B channels in each local patch, then refined by guided filtering to align with image edges. Concatenating $M$ with the low-light image creates a four-channel input that tells the encoder where illumination is missing. The encoder-decoder stages use Swish activations for smoother gradients, and a Spatial Pyramid Pooling block with $5\times5$, $9\times9$, and $13\times13$ pooling kernels at the bottleneck gathers fine, medium, and broad context. The composite generator loss $L_{Gen}=\lambda_1 L_{Adv}+\lambda_2 L_{MSE}+\lambda_3 L_{Per}$ balances adversarial realism, pixel fidelity, and perceptual similarity, and the critic loss $L_{Cri}$ trains the discriminator with $\lambda_4=1$. These pieces together carry the claimed quality-versus-efficiency result.

What would settle it

Re-run every method in Table I on the identical 15-image LOL validation split with the same preprocessing and training protocol; if any baseline reaches or exceeds a PSNR of 21.512 with an SSIM of 0.8313, the paper's top-ranking claim is refuted. Running the released EDNIG code on the same split and failing to reproduce 21.512 and 0.8313 would also settle the question.

Watch

Extended reading notes

Core claim

The central claim is that replacing the transmission map of a dehazing U-Net with an illumination map derived from the Bright Channel Prior turns the same architecture into a leading low-light enhancer. The network consumes a four-channel tensor $\{R,G,B,M\}$, where $M$ is the BCP illumination map refined by guided filtering, and is trained with $L_{Gen}=\lambda_1 L_{Adv}+\lambda_2 L_{MSE}+\lambda_3 L_{Per}$ using $\lambda_1=\lambda_2=\lambda_3=100$ and $\lambda_4=1$ for the critic. On the LOL dataset, EDNIG scores PSNR 21.512 and SSIM 0.8313, ahead of all twelve comparison methods in Table I, with 1.737M parameters. On realistic no-reference images it ranks second behind PQP under NIQE and BRISQUE. The paper interprets these results as evidence that prior-guided, lightweight generative training yields strong enhancement quality with little computational cost.

Load-bearing premise

The load-bearing premise is that the PSNR and SSIM values listed for the twelve competing methods were produced on the same 15-image LOL validation split and under training conditions comparable to EDNIG's, so that Table I is a fair controlled comparison rather than a mix of protocols.

Editorial extensions

If this is right

  • EDNIG's design offers a template for adapting dehazing networks to other restoration tasks by swapping the physical guidance map (transmission for haze, illumination for darkness).
  • At 1.737M parameters and 0.042 seconds per image, the model is light enough for mobile or embedded low-light enhancement, not just server-side processing.
  • The combination of adversarial, MSE, and perceptual losses, with weights copied from the dehazing predecessor, appears to transfer across restoration domains without re-tuning.
  • On paired benchmarks EDNIG leads all compared methods, while on unpaired realistic images it places second to PQP, suggesting its comparative advantage is clearest when ground truth is available.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: an ablation study that removes the BCP map, the SPP block, or the GAN loss individually would reveal which component actually drives the 21.512 PSNR, since the paper does not isolate them.
  • Beyond the paper: because the illumination map is computed per image by a fixed prior, the same guidance could be extended to video by temporally smoothing $M$ across frames; this is a testable route to flicker-free enhancement.
  • Beyond the paper: the reported PSNR lead over the nearest rival (about 0.2 dB over KinD++ and IDR) may be below the threshold of perceptual notice; a forced-choice human study would test whether the metric advantage translates to visible improvement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes EDNIG, an encoder-decoder (U-Net-derived) network for low-light image enhancement that concatenates a Bright Channel Prior illumination map with the RGB input, uses Spatial Pyramid Pooling and Swish activations, and is trained in a GAN framework with adversarial, MSE, and perceptual losses. Experiments on the LOL dataset report a PSNR of 21.512 dB and SSIM of 0.8313 on the 15-image validation set, outperforming all listed baselines in Table I with only 1.737M parameters; additional no-reference results on 10 realistic images and runtime comparisons are also reported. The central claim is that this compact, illumination-guided architecture achieves state-of-the-art quantitative performance while remaining efficient.

Significance. If the reported numbers are reproducible, EDNIG would be a meaningful combination of known components: a lightweight U-Net-style generator, illumination prior guidance, multi-scale pooling, and GAN-based training. The paper's strengths include a public code link, explicit training hyperparameters, parameter-count and runtime comparisons, and evaluation against a broad set of recent baselines. The main weakness is that the headline 'surpasses all' claim rests on cross-paper baseline numbers without a shared evaluation protocol, a small validation set without error bars, and no ablation isolating the contribution of each design choice; these issues are fixable but currently prevent the claim from being accepted at face value.

major comments (4)
  1. [§IV.A, §IV.B, Table I] The claim that EDNIG 'surpasses all competing methods' is load-bearing, but Table I is assembled from values reported in the original papers rather than from re-runs under a common protocol. Since EDNIG is trained and evaluated at 512×512 while the LOL dataset contains 400×600 images, and since the margins over KinD++ (0.212 dB) and IDR (0.172 dB) are small, differences in resizing, cropping, boundary handling, or score computation can reverse the ranking. Please re-run the baselines with the same preprocessing and evaluation code, or provide evidence that the cited numbers were obtained under identical conditions; in either case, report per-image scores or standard deviations for the 15-image validation set.
  2. [§III.B, §III.C, Table I] No ablation study isolates the contributions of the illumination guidance, SPP module, Swish activation, or the GAN/composite loss. Because the generator and loss weighting are inherited from the authors' prior EDN-GTM work, the quantitative gain cannot currently be attributed to any specific component. Please add ablations that remove or replace each component (e.g., without the fourth input channel, without SPP, without the adversarial term) and state the BCP patch size and guided-filter parameters used in Eq. (1).
  3. [§IV.B, Table II] The realistic-image evaluation uses 'a set of 10 standard realistic test images' that is never enumerated or sourced. NIQE and BRISQUE are no-reference metrics sensitive to image content, resolution, and color handling, so without a precise list of images and preprocessing steps the second-place result is not reproducible and the generalization claim in Section V is not testable. Please specify the image set, preprocessing, and the evaluation script.
  4. [§IV.D, Table III] The runtime comparison is presented as evidence of efficiency, but Table III compares implementations in different frameworks (TensorFlow vs. PyTorch) and reports a single average without input resolution, warm-up, or variance. Please specify the evaluation conditions or soften the claim that EDNIG is among the most efficient methods.
minor comments (6)
  1. [§I] The introduction refers to the proposed method as ENDIG twice, while the rest of the manuscript uses EDNIG; please make the name consistent.
  2. [§III.B, §IV.A] There are duplicated passages in the compiled text: the paragraph beginning 'To boost efficiency...' appears twice at the start of Section III.B, and the experimental settings paragraph in Section IV.A is also repeated; the Figure 2 caption appears twice as well.
  3. [§III.C, Eq. (3)] Equation (3) writes G_i(z) even though the generator input is called I in Eq. (2); please use G(I)_i for clarity.
  4. [§IV.A] The dataset citation in Section IV.A (LOL dataset '[16]') points to the LIME reference [16] rather than to a LOL dataset reference; please correct the citation.
  5. [§III.A, §III.C] The loss weights and guided-filter hyperparameters are free parameters; a short sensitivity analysis or at least explicit reporting of their values would strengthen reproducibility.
  6. [Table II] Table II drops several methods from Table I without explanation; please state why only those baselines are included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EDNIG's core empirical claims rest on external benchmarks and end-to-end training, not on a fitted parameter or self-referential definition.

full rationale

The paper's derivation chain is not circular. The illumination guidance map is computed from Bright Channel Prior, an externally published statistical prior, and the network is trained end-to-end on LOL pairs with a composite loss; the reported PSNR/SSIM values are empirical results of that training, not quantities manufactured from the model's inputs. The self-citations to EDN-GTM are real but not load-bearing in a circular sense: they supply the architectural template, the SPP module, and the loss balancing weights (λ1 = λ2 = λ3 = 100, λ4 = 1, "Following [8]"), but these are design and hyperparameter choices, not parameters fitted to the target evaluation metric. Nothing in the paper defines the illumination map in terms of the enhanced output, and no equation equates the claimed PSNR 21.512 / SSIM 0.8313 with a fitted input. Concerns about cross-paper protocol parity for Table I (input resolution 512×512 vs. native LOL resolution, missing error bars, and baselines quoted from their original papers) are legitimate experimental-rigor risks, but they are correctness concerns rather than circularity; they do not make the prediction equivalent to its inputs by construction. The evaluation is self-contained against independent external baselines, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central result depends on a stack of inherited choices: BCP and guided filtering for guidance, EDN-GTM as backbone, and loss weights from prior work. None of these are derived or ablated in the paper, though none are new physical postulates.

free parameters (5)
  • Loss weighting coefficients lambda_1, lambda_2, lambda_3, lambda_4 = lambda_1=lambda_2=lambda_3=100, lambda_4=1
    Copied from the authors' prior EDN-GTM work; no sensitivity analysis is reported, so the loss balance is a hand-set input to the result.
  • First-layer channel count = 12 (doubled per downsampling stage)
    Reduced from 64 in EDN-GTM to cut parameters; no ablation shows this width is optimal.
  • SPP pooling kernel sizes = 5x5, 9x9, 13x13
    Chosen to capture fine, medium, and broad context; not varied or ablated.
  • Guided filter parameters (patch radius and regularization) = Not reported
    The illumination map M depends on these settings, but the paper states only that guided filtering is applied.
  • Training schedule (learning rate, epochs, batch size) = 1e-4 decaying to 0 over 200 epochs, batch size 1
    Standard settings; they affect results but are not swept.
assumptions (5)
  • domain assumption Bright Channel Prior: in a natural well-lit image, at least one color channel in every local patch has high intensity.
    Used in Eq. (1) to build the initial illumination map.
  • domain assumption Guided filtering improves the illumination estimate by denoising while preserving edges.
    Section III.A, after Eq. (1), adopted without quantification.
  • domain assumption LOL paired images are reliable ground truth for supervised enhancement.
    Training and evaluation in Section IV.A use 485/15 pairs.
  • domain assumption The EDN-GTM encoder-decoder is a transferable backbone for restoration besides dehazing.
    Section III.B reuses the architecture from the authors' prior dehazing papers.
  • domain assumption A weighted sum of adversarial, MSE, and perceptual losses improves enhancement.
    Section III.C assumes this hybrid loss structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-Light Enhancement via Encoder-Decoder Network with Illumination Guidance." pith.science (2026). https://pith.science/paper/BZGQ4EAI

@misc{pith2026250713360,
  author       = {Pith},
  title        = {Pith review of: Low-Light Enhancement via Encoder-Decoder Network with Illumination Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZGQ4EAI}},
  note         = {Machine review of arXiv:2507.13360}
}
read the original abstract

This paper introduces a novel deep learning framework for low-light image enhancement, named the Encoder-Decoder Network with Illumination Guidance (EDNIG). Building upon the U-Net architecture, EDNIG integrates an illumination map, derived from Bright Channel Prior (BCP), as a guidance input. This illumination guidance helps the network focus on underexposed regions, effectively steering the enhancement process. To further improve the model's representational power, a Spatial Pyramid Pooling (SPP) module is incorporated to extract multi-scale contextual features, enabling better handling of diverse lighting conditions. Additionally, the Swish activation function is employed to ensure smoother gradient propagation during training. EDNIG is optimized within a Generative Adversarial Network (GAN) framework using a composite loss function that combines adversarial loss, pixel-wise mean squared error (MSE), and perceptual loss. Experimental results show that EDNIG achieves competitive performance compared to state-of-the-art methods in quantitative metrics and visual quality, while maintaining lower model complexity, demonstrating its suitability for real-world applications. The source code for this work is available at https://github.com/tranleanh/ednig.

Figures

Figures reproduced from arXiv: 2507.13360 by the authors.

Figure 1
Figure 1. The EDNIG framework. performance is sensitive to parameter tuning and struggles with complex scenes. Optimization-based techniques, such as LIME [16], estimate illumination maps using hand-crafted priors and refine them via regularization. Although effective in some cases, these methods lack the generalization ability required for diverse low-light conditions. Currently, deep learning has transformed low-light image… view at source ↗
Figure 2
Figure 2. Typical visual comparisons of various approaches on the LOL validation dataset. Fig. 2: Typical visual comparisons of various approaches on the LOL validation dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Typical visual comparisons of various approaches on realistic image data. Fig. 3: Typical visual comparisons of various approaches on realistic image data. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [1]

    Low light enhancement al- gorithm for color images using intuitionistic fuzzy sets with histogram equalization,

    J. R. Jebadass and P. Balasubramaniam, “Low light enhancement al- gorithm for color images using intuitionistic fuzzy sets with histogram equalization,” Multimedia Tools and Applications , vol. 81, no. 6, pp. 8093–8106, 2022. 1

    work page 2022

  2. [2]

    An optimization-based approach to gamma correc- tion parameter estimation for low-light image enhancement,

    I. Jeong and C. Lee, “An optimization-based approach to gamma correc- tion parameter estimation for low-light image enhancement,” Multimedia Tools and Applications, vol. 80, pp. 18 027–18 042, 2021. 1

    work page 2021

  3. [3]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems , vol. 27, 2014. 1, 3

    work page 2014

  4. [4]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . Springer, 2015, pp. 234–241. 1, 2

    work page 2015

  5. [5]

    Robust u-net-based road lane markings detection for autonomous driving,

    L.-A. Tran and M.-H. Le, “Robust u-net-based road lane markings detection for autonomous driving,” in 2019 International Conference on System Science and Engineering (ICSSE) . IEEE, 2019, pp. 62–66. 1

    work page 2019

  6. [6]

    Lightweight image dehazing networks based on soft knowledge distillation,

    L.-A. Tran and D.-C. Park, “Lightweight image dehazing networks based on soft knowledge distillation,” The Visual Computer , pp. 1–20, 2024. 1

    work page 2024

  7. [7]

    A novel encoder-decoder network with guided transmission map for single image dehazing,

    L.-A. Tran, S. Moon, and D.-C. Park, “A novel encoder-decoder network with guided transmission map for single image dehazing,” Procedia Computer Science, vol. 204, pp. 682–689, 2022. 1

    work page 2022

  8. [8]

    Encoder–decoder network with guided transmission map: Robustness and applicability,

    L.-A. Tran and D.-C. Park, “Encoder–decoder network with guided transmission map: Robustness and applicability,” in International Sym- posium on Intelligent Informatics . Springer, 2022, pp. 41–54. 1, 3

    work page 2022

Show all 35 references
  1. [9]

    Encoder-decoder networks with guided transmission map for effective image dehazing,

    ——, “Encoder-decoder networks with guided transmission map for effective image dehazing,” The Visual Computer , pp. 1–24, 2024. 1, 2, 3

    2024

  2. [10]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE transactions on pattern analysis and machine intelligence, vol. 33, no. 12, pp. 2341–2353, 2010. 1, 2

    2010

  3. [11]

    Unsupervised low-light image en- hancement using bright channel prior,

    H. Lee, K. Sohn, and D. Min, “Unsupervised low-light image en- hancement using bright channel prior,” IEEE Signal Processing Letters , vol. 27, pp. 251–255, 2020. 1, 2

    2020

  4. [12]

    Realization of the contrast limited adaptive histogram equalization (clahe) for real-time image enhancement,

    A. M. Reza, “Realization of the contrast limited adaptive histogram equalization (clahe) for real-time image enhancement,” Journal of VLSI signal processing systems for signal, image and video technology , vol. 38, pp. 35–44, 2004. 1

    2004

  5. [13]

    The retinex theory of color vision,

    E. H. Land, “The retinex theory of color vision,” Scientific american , vol. 237, no. 6, pp. 108–129, 1977. 1, 2

    1977

  6. [14]

    Properties and performance of a center/surround retinex,

    D. J. Jobson, Z.-u. Rahman, and G. A. Woodell, “Properties and performance of a center/surround retinex,” IEEE transactions on image processing, vol. 6, no. 3, pp. 451–462, 1997. 1

    1997

  7. [15]

    Multi-scale retinex for color image enhancement,

    Z.-u. Rahman, D. J. Jobson, and G. A. Woodell, “Multi-scale retinex for color image enhancement,” in Proceedings of 3rd IEEE international conference on image processing , vol. 3. IEEE, 1996, pp. 1003–1006. 1

    1996

  8. [16]

    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 Transactions on image processing , vol. 26, no. 2, pp. 982–993, 2016. 2

    2016

  9. [17]

    Deep retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” arXiv preprint arXiv:1808.04560 , 2018. 2, 3, 4, 5

    2018 arXiv

  10. [18]

    Gladnet: Low-light enhance- ment network with global awareness,

    W. Wang, C. Wei, W. Yang, and J. Liu, “Gladnet: Low-light enhance- ment network with global awareness,” in 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018) . IEEE, 2018, pp. 751–755. 2, 4, 5

    2018

  11. [19]

    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,” inProceedings of the 27th ACM international conference on multimedia , 2019, pp. 1632–1640. 2, 4, 5

    2019

  12. [20]

    Beyond brightening low-light images,

    Y . Zhang, X. Guo, J. Ma, W. Liu, and J. Zhang, “Beyond brightening low-light images,” International Journal of Computer Vision , vol. 129, pp. 1013–1037, 2021. 2, 4, 5

    2021

  13. [21]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 17 683–17 693. 2, 4

    2022

  14. [22]

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models,

    O. ¨Ozdenizci and R. Legenstein, “Restoring vision in adverse weather conditions with patch-based denoising diffusion models,” IEEE Trans- actions on Pattern Analysis and Machine Intelligence , vol. 45, no. 8, pp. 10 346–10 357, 2023. 2, 4

    2023

  15. [23]

    Zero-reference low-light enhancement via physical quadruple priors,

    W. Wang, H. Yang, J. Fu, and J. Liu, “Zero-reference low-light enhancement via physical quadruple priors,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 26 057–26 066. 2, 4, 5

    2024

  16. [24]

    Guided image filtering,

    K. He, J. Sun, and X. Tang, “Guided image filtering,” IEEE transactions on pattern analysis and machine intelligence , vol. 35, no. 6, pp. 1397– 1409, 2012. 2

    2012

  17. [25]

    Deblurgan: Blind motion deblurring using conditional adversarial net- works,

    O. Kupyn, V . Budzan, M. Mykhailych, D. Mishkin, and J. Matas, “Deblurgan: Blind motion deblurring using conditional adversarial net- works,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8183–8192. 3

    2018

  18. [26]

    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 Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part II 14 . Springer, 2016, pp. 694–711. 3

    2016

  19. [27]

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

    C. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, and R. Cong, “Zero- reference deep curve estimation for low-light image enhancement,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1780–1789. 4

    2020

  20. [28]

    Band representation- based semi-supervised low-light image enhancement: Bridging the gap between signal fidelity and perceptual quality,

    W. Yang, S. Wang, Y . Fang, Y . Wang, and J. Liu, “Band representation- based semi-supervised low-light image enhancement: Bridging the gap between signal fidelity and perceptual quality,” IEEE Transactions on Image Processing, vol. 30, pp. 3461–3473, 2021. 4

    2021

  21. [29]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739. 4

    2022

  22. [30]

    Image restoration with mean-reverting stochastic differential equations,

    Z. Luo, F. K. Gustafsson, Z. Zhao, J. Sj ¨olund, and T. B. Sch ¨on, “Image restoration with mean-reverting stochastic differential equations,” arXiv preprint arXiv:2301.11699, 2023. 4

    2023 arXiv

  23. [31]

    Ingredient-oriented multi-degradation learning for image restoration,

    J. Zhang, J. Huang, M. Yao, Z. Yang, H. Yu, M. Zhou, and F. Zhao, “Ingredient-oriented multi-degradation learning for image restoration,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 5825–5835. 4

    2023

  24. [32]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014. 3

    2014 arXiv

  25. [33]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,” IEEE Signal processing letters , vol. 20, no. 3, pp. 209–212, 2012. 4

    2012

  26. [34]

    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 Transactions on image processing, vol. 21, no. 12, pp. 4695–4708, 2012. 4

    2012

  27. [35]

    Distilled pooling transformer encoder for efficient realistic image dehazing,

    L.-A. Tran and D.-C. Park, “Distilled pooling transformer encoder for efficient realistic image dehazing,” Neural Computing and Applications , vol. 37, no. 6, pp. 5203–5221, 2025. 5

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.