pith. machine review for the scientific record. sign in

arxiv: 2604.17390 · v2 · submitted 2026-04-19 · 💻 cs.CV · cs.AI· cs.GR

Recognition: unknown

MESA: A Training-Free Multi-Exemplar Deep Framework for Restoring Ancient Inscription Textures

Vasileios Toulatzis , Sofia Theodoridou , Ioannis Fudos
Authors on Pith no claims yet

Pith reviewed 2026-05-10 06:17 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.GR
keywords ancient inscriptionsimage restorationmulti-exemplar synthesistexture transferVGG featuresepigraphystyle-aware restorationtraining-free
0
0 comments X

The pith

MESA restores damaged ancient inscription images by guiding synthesis with textures from multiple well-preserved exemplars without training.

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

The paper introduces MESA as a way to restore missing or corrupted parts of ancient inscription images without any training. It draws on well-preserved exemplar inscriptions from the same monument or similar styles to guide the reconstruction of textures and strokes. By encoding features from a pretrained network as Gram matrices and selecting matches based on displacement, plus weighting by estimated character widths, the method focuses synthesis on damaged areas only. Scholars would benefit if this produces legible text that matches the original carving style, aiding historical analysis where physical artifacts are incomplete.

Core claim

MESA encodes VGG19 convolutional features as Gram matrices to capture exemplar texture, style, and stroke structure. For each neural network layer, it selects the exemplar minimizing Mean-Squared Displacement to the damaged input. Layer-wise contribution weights come from Optical Character Recognition-estimated character widths in the exemplar set to bias filters toward scales matching letter geometry. A training mask preserves intact regions so synthesis is restricted to damaged areas.

What carries the argument

MESA framework using Gram matrices of VGG19 features for multi-exemplar texture capture, with mean-squared displacement selection and OCR-derived weights for layer contributions.

Load-bearing premise

Suitable well-preserved exemplars with matching letterforms and material are available and the mean-squared displacement selection plus OCR widths will pick the right ones.

What would settle it

A controlled test on inscriptions with known original intact versions after applying artificial damage, verifying if MESA reconstructs the correct textures and letter shapes.

Figures

Figures reproduced from arXiv: 2604.17390 by Ioannis Fudos, Sofia Theodoridou, Vasileios Toulatzis.

Figure 1
Figure 1. Figure 1: Restoring Ancient Inscription Textures: overview of our MESA training free approach. On the left is the inscription [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training and Feed-Forward Phases for Inscription Restoration of well-known Network Models (DRUNET, ESRGAN [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Restoring Ancient Inscriptions: 𝐺 denotes the Gram matrix of feature maps in layer 𝑙, depending on the number of filters 𝐹𝑙 . The network structure is based on VGG19 [18], using AvgPooling instead of MaxPooling layers. Networks generated with PlotNeuralNet (https://github.com/HarisIqbal88/PlotNeuralNet) and then modified. Masking the image, both at initialization and at each training step in either of the … view at source ↗
Figure 4
Figure 4. Figure 4: Best-fit distributions of letter widths for the three datasets used in our experiments. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Tesseract [19] letter widths detection. Each rectangle captures some letters and we divide each width by the number of letters detected within it to define an average width of letters inside it. MESA is particularly suitable when the damaged inscription is part of a larger text containing intact regions that can serve as exemplars, or when a dataset of inscriptions in the same or a similar style is availab… view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of our method’s outputs per test image using Dataset A images as exemplars with other well-known [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of our method’s outputs per test image using Dataset B images as exemplars with other well-known [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of our method’s outputs per test image using Dataset C images as exemplars with other well-known [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Levenshtein Distance evaluation [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Text Recovery Score (TRS) evaluation [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Log-scaled Levenshtein Similarity evaluation as described in 4.1 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: LPIPS evaluation with Alex network usage [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: LPIPS evaluation with VGG network usage Method Dataset A Dataset B Dataset C Average DIP-DENOISE TF 0.8269 0.7059 0.6922 0.7416 DIP-INPAINTING TF 0.8834 0.7548 0.8212 0.8198 DNCNN TN 0.9590 0.9066 0.6169 0.8275 DNCNN TS 0.9458 0.8855 0.9094 0.9136 DRUNET TN 0.8393 0.8579 0.9182 0.8718 DRUNET TS 0.8174 0.8703 0.9040 0.8639 ESRGAN TN 0.9607 0.9133 0.9203 0.9314 ESRGAN TS 0.9482 0.8938 0.9071 0.9164 MESA TF … view at source ↗
Figure 14
Figure 14. Figure 14: SSIM evaluation Method Dataset A Dataset B Dataset C Averages DIP-DENOISE TF 29.6638 29.9005 29.7345 29.7663 DIP-INPAINTING TF 31.2086 31.1740 31.3163 31.2330 DNCNN TN 38.4847 36.8845 31.3225 35.5639 DNCNN TS 38.3920 36.8139 36.0718 37.0926 DRUNET TN 30.8336 32.4036 36.2790 33.1720 DRUNET TS 31.0895 33.4126 35.7140 33.4054 ESRGAN TN 39.3524 37.5612 36.0042 37.6393 ESRGAN TS 39.1888 36.8152 36.6121 37.5387… view at source ↗
Figure 15
Figure 15. Figure 15: PSNR evaluation 5 Discussion, Conclusions & Future Work In this work, we introduced MESA, a training-free, multi-exemplar, style-aware method for the restoration of degraded inscriptions. Unlike conventional supervised approaches such as DRUNET, DnCNN, and ESRGAN, our method does not require paired ruined/clean datasets or extensive pretraining. Instead, it leverages clean exemplars from the same or simil… view at source ↗
Figure 16
Figure 16. Figure 16: Dataset A used as exemplars in our method or as ground truth in techniques that need supervised learning. [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Dataset B used as exemplars in our method or as ground truth in techniques that need supervised learning. [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Dataset C used as exemplars in our method or as ground truth in techniques that need supervised learning. [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Comparison of the output of our method per test image using Dataset A images as exemplars with baseline methods. [PITH_FULL_IMAGE:figures/full_fig_p029_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Comparison of the output of our method per test image using Dataset B images as exemplars with baseline methods. [PITH_FULL_IMAGE:figures/full_fig_p030_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Comparison of the output of our method per test image using Dataset C images as exemplars with baseline methods. [PITH_FULL_IMAGE:figures/full_fig_p031_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Different setups for restoring an inscription showing each network layer’s contribution and performance weight in [PITH_FULL_IMAGE:figures/full_fig_p032_22.png] view at source ↗
read the original abstract

Ancient inscriptions frequently suffer missing or corrupted regions from fragmentation, erosion, or other damage, hindering reading, and analysis. We review prior image restoration methods and their applicability to inscription image recovery, then introduce MESA (Multi-Exemplar, Style-Aware) -an image-level restoration method that uses well-preserved exemplar inscriptions (from the same epigraphic monument, material, or similar letterforms) to guide reconstruction of damaged text. MESA encodes VGG19 convolutional features as Gram matrices to capture exemplar texture, style, and stroke structure; for each neural network layer it selects the exemplar minimizing Mean-Squared Displacement (MSD) to the damaged input. Layer-wise contribution weights are derived from Optical Character Recognition-estimated character widths in the exemplar set to bias filters toward scales matching letter geometry, and a training mask preserves intact regions so synthesis is restricted to damaged areas. We also summarize prior network architectures and exemplar and single-image synthesis, inpainting, and Generative Adversarial Network (GAN) approaches, highlighting limitations that MESA addresses. Comparative experiments demonstrate the advantages of MESA. Finally, we provide a practical roadmap for choosing restoration strategies given available exemplars and metadata.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper proposes MESA, a training-free image-level restoration framework for damaged ancient inscriptions. It encodes VGG19 features as Gram matrices from multiple well-preserved exemplars (same monument, material, or letterforms), selects per-layer the exemplar minimizing Mean-Squared Displacement (MSD) to the damaged input, derives layer weights from OCR-estimated character widths to bias toward matching scales, and applies an intact-region mask to restrict synthesis to damaged areas. The work reviews prior inpainting, style transfer, and GAN methods, claims comparative experiments show advantages, and provides a roadmap for choosing strategies based on available exemplars.

Significance. If the central claims hold, MESA offers a practical, training-free alternative to data-hungry methods for epigraphic restoration when suitable exemplars exist, leveraging standard components (VGG19, Gram matrices, OCR) in a style-aware pipeline. The parameter-free exemplar selection and mask-based preservation are notable strengths for reproducibility. However, significance is limited by the absence of quantitative metrics or dataset details in the provided text, and the approach's reliability depends on unproven robustness of MSD selection under damage.

major comments (2)
  1. [Method (exemplar selection and MSD computation)] The exemplar selection step (described in the method overview) relies on per-layer MSD minimization between Gram matrices (or features) of the damaged input and each exemplar. Because MSD is evaluated on the full damaged image, erosion, fragmentation, and artifacts can distort local activations and global statistics; no analysis shows that the argmin still recovers the underlying letterform style rather than an exemplar matching the damage pattern. The downstream OCR-derived weights and intact mask operate after selection and cannot correct an upstream mismatch. This is load-bearing for the claim that MESA reliably guides reconstruction.
  2. [Abstract and Experimental section] The abstract and conclusion state that 'comparative experiments demonstrate the advantages of MESA,' yet no quantitative results, error metrics (e.g., PSNR, SSIM, perceptual scores), dataset sizes, baseline details, or ablation tables are referenced. Without these, the performance claims cannot be verified and the cross-method superiority remains unsupported.
minor comments (2)
  1. [Method notation] Clarify the precise definition and computation of 'Mean-Squared Displacement' (MSD) when applied to Gram matrices or VGG features, as this term is non-standard in style-transfer literature (typically MSE or Frobenius distance is used).
  2. [Abstract] The abstract would be strengthened by including at least one key quantitative result or metric to support the 'demonstrate advantages' claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the presentation and support for our claims.

read point-by-point responses

  1. Referee: [Method (exemplar selection and MSD computation)] The exemplar selection step (described in the method overview) relies on per-layer MSD minimization between Gram matrices (or features) of the damaged input and each exemplar. Because MSD is evaluated on the full damaged image, erosion, fragmentation, and artifacts can distort local activations and global statistics; no analysis shows that the argmin still recovers the underlying letterform style rather than an exemplar matching the damage pattern. The downstream OCR-derived weights and intact mask operate after selection and cannot correct an upstream mismatch. This is load-bearing for the claim that MESA reliably guides reconstruction.

    Authors: We acknowledge the concern that MSD computed over the full image could potentially be influenced by damage patterns rather than underlying style. Gram matrices from VGG19 layers are intended to capture global second-order statistics of texture and stroke structure, which are less sensitive to localized erosion or fragmentation than raw features. Nevertheless, the manuscript does not include explicit robustness analysis of the selection step. In the revised version we will add a new subsection with controlled experiments: we will apply synthetic damage masks of varying severity to intact exemplars, re-run the per-layer MSD selection, and demonstrate that the chosen exemplar consistently matches the original letterform style (verified via visual inspection and OCR consistency) rather than the damage configuration. This analysis will directly address the load-bearing nature of the selection mechanism. revision: yes

  2. Referee: [Abstract and Experimental section] The abstract and conclusion state that 'comparative experiments demonstrate the advantages of MESA,' yet no quantitative results, error metrics (e.g., PSNR, SSIM, perceptual scores), dataset sizes, baseline details, or ablation tables are referenced. Without these, the performance claims cannot be verified and the cross-method superiority remains unsupported.

    Authors: We agree that the current experimental section relies on qualitative visual comparisons and does not report numerical metrics, dataset statistics, or ablations, which limits verifiability of the superiority claims. The manuscript presents side-by-side results against representative inpainting and style-transfer baselines, but these are not quantified. In the revision we will expand the experimental section to include: (1) a clear description of the evaluation dataset (number of images, sources, and damage characteristics), (2) quantitative metrics (PSNR, SSIM, and a perceptual metric such as LPIPS) computed against ground-truth intact regions where available, (3) details of the baseline implementations, and (4) ablation tables isolating the contributions of per-layer MSD selection, OCR-derived weights, and the intact-region mask. These additions will be placed before the conclusion and will directly support the statements in the abstract. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper presents MESA as a pipeline that encodes VGG19 features into Gram matrices, selects exemplars via MSD minimization per layer, derives layer weights from OCR-estimated character widths, and applies an intact-region mask for synthesis. These steps operate on external inputs (damaged image, exemplar set, pre-trained VGG19, OCR) without any equation or selection rule that reduces the restored output to a fitted parameter, self-defined quantity, or self-citation chain. No load-bearing premise collapses to prior author work by construction, and the method remains self-contained against the listed external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on established computer vision primitives rather than new postulates; the main additions are algorithmic choices for exemplar selection and weighting.

axioms (2)
  • domain assumption VGG19 convolutional features encoded as Gram matrices capture texture, style, and stroke structure relevant to inscription restoration
    Invoked in the description of how exemplars guide reconstruction; standard assumption from neural style transfer literature.
  • domain assumption OCR-estimated character widths in the exemplar set provide reliable layer-wise contribution weights that match letter geometry
    Used to bias filters toward appropriate scales; assumes OCR performs accurately on the chosen exemplars.

pith-pipeline@v0.9.0 · 5522 in / 1321 out tokens · 54826 ms · 2026-05-10T06:17:08.699918+00:00 · methodology

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

30 extracted references · 17 canonical work pages · 2 internal anchors

  1. [1]

    Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murra...

    2015

  2. [2]

    Yannis Assael, Thea Sommerschield, Alison Cooley, Brendan Shillingford, John Pavlopoulos, Priyanka Suresh, Bailey Herms, Justin Grayston, Benjamin Maynard, Nicholas Dietrich, Robbe Wulgaert, Jonathan Prag, Alex Mullen, and Shakir Mohamed. 2025. Contextual- izing ancient texts with generative neural networks.Nature(07 2025), 1–7. doi:10.1038/s41586-025-09292-5

    work page doi:10.1038/s41586-025-09292-5 2025

  3. [3]

    Yannis Assael, Thea Sommerschield, and J. Prag. 2019. Restoring ancient text using deep learning: a case study on Greek epigraphy. In Conference on Empirical Methods in Natural Language Processing. https://api.semanticscholar.org/CorpusID:202781916

    2019

  4. [4]

    Yannis Assael, Thea Sommerschield, and Jonathan Prag. 2019. Restoring ancient text using deep learning: a case study on Greek epigraphy. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Kentaro Inui, Jing Jiang, Vincent Ng, and...

    work page doi:10.18653/v1/d19-1668 2019

  5. [5]

    Yannis Assael*, Thea Sommerschield*, Brendan Shillingford, Mahyar Bordbar, John Pavlopoulos, Marita Chatzipanagiotou, Ion Androut- sopoulos, Jonathan Prag, and Nando de Freitas. 2022. Restoring and attributing ancient texts using deep neural networks.Nature(2022). doi:10.1038/s41586-022-04448-z

    work page doi:10.1038/s41586-022-04448-z 2022

  6. [6]

    Tai-Yin Chiu. 2019. Understanding Generalized Whitening and Coloring Transform for Universal Style Transfer. In2019 IEEE/CVF International Conference on Computer Vision (ICCV). 4451–4459. doi:10.1109/ICCV.2019.00455

    work page doi:10.1109/iccv.2019.00455 2019

  7. [7]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, et al. 2024. The LLaMA 3 Herd of Models. arXiv:2407.21783 [cs.AI] doi:10.48550/ arXiv.2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    Gatys, Alexander S

    Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. 2016. Image Style Transfer Using Convolutional Neural Networks. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2414–2423. doi:10.1109/CVPR.2016.265

    work page doi:10.1109/cvpr.2016.265 2016

  9. [9]

    Shuhao Guan, Moule Lin, Cheng Xu, Xinyi Liu, Jinman Zhao, Jiexin Fan, Qi Xu, and Derek Greene. 2025. PreP-OCR: A Complete Pipeline for Document Image Restoration and Enhanced OCR Accuracy. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and M...

    work page doi:10.18653/v1/2025.acl-long.749 2025

  10. [10]

    Xun Huang and Serge Belongie. 2017. Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization. In2017 IEEE International Conference on Computer Vision (ICCV). 1510–1519. doi:10.1109/ICCV.2017.167

    work page doi:10.1109/iccv.2017.167 2017

  11. [11]

    Huynh-Thu and M

    Q. Huynh-Thu and M. Ghanbari. 2008. Scope of validity of PSNR in image/video quality assessment.Electronics letters44, 13 (2008), 800–801

    2008

  12. [12]

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. 2017. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 105–114. doi:10.110...

    work page doi:10.1109/cvpr.2017.19 2017

  13. [13]

    V. I. Levenshtein. 1966. Binary Codes Capable of Correcting Deletions, Insertions and Reversals.Soviet Physics Doklady10 (Feb. 1966), 707

    1966

  14. [14]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. SwinIR: Image Restoration Using Swin Transformer. In2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). 1833–1844. doi:10.1109/ICCVW54120. 2021.00210

    work page doi:10.1109/iccvw54120 2021

  15. [15]

    Dong C Liu and Jorge Nocedal. 1989. On the limited memory BFGS method for large scale optimization.Mathematical programming45, 1-3 (1989), 503–528

    1989

  16. [16]

    Reda, Kevin J

    Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. 2018. Image Inpainting for Irregular Holes Using Partial Convolutions. InProceedings of the European Conference on Computer Vision (ECCV)(Munich, Germany). Springer-Verlag, Berlin, Heidelberg, 89–105. doi:10.1007/978-3-030-01252-6_6

    work page doi:10.1007/978-3-030-01252-6_6 2018

  17. [17]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Nassir Navab, Joachim Hornegger, William M. Wells, and Alejandro F. Frangi (Eds.). Springer International Publishing, Cham, 234–241

    2015

  18. [18]

    Karen Simonyan and Andrew Zisserman. 2014. Very Deep Convolutional Networks for Large-Scale Image Recognition.CoRR abs/1409.1556 (2014). http://arxiv.org/abs/1409.1556

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  19. [19]

    Ray Smith. 2007. An Overview of the Tesseract OCR Engine. InProceedings of the International Conference on Document Analysis and Recognition (ICDAR), Vol. 2. IEEE, 629–633

    2007

  20. [20]

    Lempitsky

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor S. Lempitsky. 2021. Resolution-robust Large Mask Inpainting with Fourier Convolutions.2022 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)(2021), 3172–3182. https://api.sem...

    2021

  21. [21]

    Vasilis Toulatzis and Ioannis Fudos. 2019. Deep Terrain Expansion: Terrain Texture Synthesis with Deep Learning. InComputer Graphics and Visual Computing (CGVC), Franck P. Vidal, Gary K. L. Tam, and Jonathan C. Roberts (Eds.). The Eurographics Association. doi:10.2312/cgvc.20191262

    work page doi:10.2312/cgvc.20191262 2019

  22. [22]

    Vasilis Toulatzis and Ioannis Fudos. 2021. Deep Tiling: Texture Tile Synthesis Using a Constant Space Deep Learning Approach. In Advances in Visual Computing, George Bebis, Vassilis Athitsos, Tong Yan, Manfred Lau, Frederick Li, Conglei Shi, Xiaoru Yuan, Christos Mousas, and Gerd Bruder (Eds.). Springer International Publishing, Cham, 414–426

    2021

  23. [23]

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. 2018. Deep Image Prior. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2018

  24. [24]

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. 2018. ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks. InComputer Vision – ECCV 2018 Workshops: Munich, Germany, September 8-14, 2018, Proceedings, Part V(Munich, Germany). Springer-Verlag, Berlin, Heidelberg, 63–79. doi:10.1007/978-3-030-11021-5_5

    work page doi:10.1007/978-3-030-11021-5_5 2018

  25. [25]

    Zhou Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity. Trans. Img. Proc.13, 4 (April 2004), 600–612. doi:10.1109/TIP.2003.819861

    work page doi:10.1109/tip.2003.819861 2004

  26. [26]

    Zhenhua Yang, Dezhi Peng, Yongxin Shi, Yuyi Zhang, Chongyu Liu, and Lianwen Jin. 2025. Predicting the Original Appearance of Damaged Historical Documents.Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)(2025). 34•Vasilis Toulatzis, Sofia Theodoridou, and Ioannis Fudos

    2025

  27. [27]

    Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Gool, and Radu Timofte. 2021. Plug-and-Play Image Restoration With Deep Denoiser Prior.IEEE Transactions on Pattern Analysis and Machine IntelligencePP (06 2021), 1–1. doi:10.1109/TPAMI.2021.3088914

    work page doi:10.1109/tpami.2021.3088914 2021

  28. [28]

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. 2016. Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising.IEEE Transactions on Image ProcessingPP (08 2016). doi:10.1109/TIP.2017.2662206

    work page doi:10.1109/tip.2017.2662206 2016

  29. [29]

    Zhang, Wangmeng Zuo, and Lei Zhang

    K. Zhang, Wangmeng Zuo, and Lei Zhang. 2017. FFDNet: Toward a Fast and Flexible Solution for CNN-Based Image Denoising.IEEE Transactions on Image Processing27 (2017), 4608–4622. https://api.semanticscholar.org/CorpusID:10514149

    2017

  30. [30]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 586–595

    2018