Pith. sign in

REVIEW 3 major objections 4 minor 81 references

Uni-DocDiff: A Unified Document Restoration Model Based on Diffusion

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Uni-DocDiff is a single diffusion-based network that handles six document restoration tasks—dewarping, deblurring, deshadowing, illumination rectification, binarization, and handwriting removal—claiming performance on par with or better tha

desk verdict A credible unified document-restoration diffusion model whose six-task results hold up, but whose headline 'seamless adaptation' claim rests on one underspecified experiment. read the letter →

arxiv 2508.04055 v1 pith:3WFFAE3B submitted 2025-08-06 cs.CV

classification cs.CV
keywords documentimagerestorationall-in-onemodeldiffusionmulti-tasklearningtaskinterferencedewarpingdeshadowingpriorpool
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

Uni-DocDiff is a single diffusion-based network that restores damaged documents across six tasks: dewarping, deblurring, deshadowing, illumination rectification, binarization, and handwriting removal. The paper's central claim is that this one unified model performs comparably to or better than task-specific expert models on each task, while remaining easy to extend to new tasks with minimal retraining. To make that possible, the authors replace handcrafted prompts with a learnable task embedding, a Prior Pool of classical image features, and a Prior Fusion Module that adaptively selects the relevant features at every reconstruction stage, suppressing interference between tasks. If the claim holds, a document processing system no longer needs a stack of specialist models; one trainable, extensible unit can cover the main restoration jobs. The paper reports top or second-best results on most benchmarks and shows that adding a new task by freezing everything except the small fusion module outperforms an earlier prompt-based unified model.

What carries the argument

The Prior Pool and the Prior Fusion Module (PFM). The Prior Pool is a bank of cheap classical features - Sobel and Canny edge maps for high-frequency content, median- and Gaussian-filtered images and a Discrete Cosine Transform for low-frequency illumination - extracted from the degraded document. The PFM, inserted at each reconstruction stage of the diffusion denoiser, computes content and task embeddings, uses them to weight the pool, and adds the weighted priors into the feature stream, so the network can emphasise edges for deblurring and background shading for deshadowing. This selection mechanism is claimed to be what lets one diffusion backbone handle several restoration tasks without

What would settle it

Collect 200 real camera-captured documents with paired clean references under varied lighting (e.g., an extension of the RealDAE test set), run Uni-DocDiff and a single-task illumination expert such as GCDRNet on each, and compare PSNR/SSIM; if the expert wins on a majority of images, the claimed parity/superiority for illumination rectification fails.

Watch

Extended reading notes

Core claim

Uni-DocDiff is a dual-stream architecture. A Pixel Prediction Branch (PPB) runs an image-conditional diffusion denoiser that directly predicts the clean document image, conditioned on the degraded image, a learnable task embedding, and a Prior Pool; a Coordinate Prediction Branch (CPB) predicts a sparse backward map for dewarping. The Prior Pool contains high-frequency features from Sobel and Canny edge operators and low-frequency features from median filtering, Gaussian filtering, and the Discrete Cosine Transform, all computed from the degraded image. The Prior Fusion Module (PFM), placed between Res blocks and upsampling stages, converts the pool features to the current resolution, derive

Load-bearing premise

The illumination-rectification training data is synthesized by overlaying lighting patterns from DocShade onto clean document images, but the paper does not describe how the layers are combined, so the model's real-world illumination gains depend on an untested assumption that the synthetic shadows and lighting resemble camera-captured conditions.

Editorial extensions

If this is right

  • Document processing systems can retire separate models for dewarping, deblurring, deshadowing, illumination correction, binarization, and handwriting removal, replacing them with a single diffusion network.
  • New restoration tasks can be added by training only the Prior Fusion Module, keeping the diffusion backbone frozen, which lowers the cost and risk of extending the system.
  • The frequency-aware loss separates tasks into high-frequency (deblurring, binarization, handwriting removal) and low-frequency (deshadowing, illumination rectification) groups, a grouping that can be reused in other multi-task image restoration settings.
  • The dual-stream design lets dewarping be trained or fine-tuned independently of pixel restoration, so geometric and photometric degradations do not have to be optimised together.
  • The Prior Pool's cheap classical features (Sobel, Canny, median/Gaussian filtering, DCT) provide task-relevant conditioning that a multi-task network can exploit with no learned preprocessing.

Reading between the lines

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

  • If the scalability result transfers, document restoration could become a single continuously growing model: each new degradation is added by collecting data and training a small adapter, rather than retraining or composing several specialist networks. This is the paper's implicit promise but is tested on only one added task.
  • The unspecified procedure for synthesizing the illumination-rectification training set (combining DocShade lighting with clean documents) is the clearest risk to the real-world claims; a natural follow-up is to publish the composition protocol and test on a new camera-captured benchmark.
  • The Prior Pool's operators are individually cheap and interpretable, so one could ablate each one (Sobel vs Canny vs DCT) to learn which features carry the inter-task interference reduction, and potentially replace them with learned filters for even better scalability.
  • The paper's own conclusion notes that only one new task (handwriting removal) was tested for scalability; adding two or three tasks at once with only PFM training would reveal whether the interference suppression generalizes or whether pairwise interactions accumulate.
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

3 major / 4 minor

Summary. The paper proposes Uni-DocDiff, a diffusion-based unified document restoration model that handles six tasks in a single network: deblurring, deshadowing, illumination rectification, dewarping, binarization, and handwriting removal. It introduces a Prior Pool built from classical image processing operators, a Prior Fusion Module (PFM) that combines task-specific and content-specific weights, and a dual-stream architecture with a Coordinate Prediction Branch for dewarping. The method is evaluated against task-specific state-of-the-art models and the unified DocRes baseline, with ablations supporting the Prior Pool, PFM, and frequency-aware loss, plus a scalability experiment that adapts the model to handwriting removal while freezing the backbone except for PFM.

Significance. If the results hold, the paper makes a useful contribution: it would be among the first diffusion-based unified document restoration models and could replace several specialist models with one network while remaining extensible to new tasks. The quantitative comparisons in Table 1 cover six tasks and multiple established benchmarks, and the ablations in Table 2 and Figure 7 provide evidence that the proposed components reduce task interference. The central limitations are that the scalability claim rests on a single, underspecified new-task experiment, and that several implementation details needed for reproducibility are missing.

major comments (3)
  1. [§3.4, Eq. (7); §3.2 Prior Pool] The frequency-aware loss is load-bearing: Table 2 shows that removing it degrades results on most tasks. However, the low-pass and high-pass filters φ_L and φ_H in Eq. (7) are never specified: no filter type, cutoff frequency, kernel size, or implementation is given. Likewise, the Prior Pool construction in §3.2 names Sobel, Canny, median filtering, Gaussian filtering, and DCT, but omits kernel sizes, thresholds, and how DCT coefficients are retained. These details are necessary to reproduce the method and to verify that the reported gains are not artifacts of particular filter choices.
  2. [§4.5, Eq. (4)] The task-scalability claim is not yet established. The text says all parameters except PFM are frozen when adding handwriting removal, but Eq. (4) uses task as input to an MLP, and the paper never defines the task representation. If task is a learned embedding, it is unclear how a brand-new task's embedding is initialized and trained; if it is a fixed one-hot or random vector, the mechanism by which the PFM can adapt to a novel task is unexplained. Section 4.5 also calls PFM the 'Prompt Feature Module' while §3.2 defines it as 'Prior Fusion Module,' which adds further ambiguity. Only one new task is tested (Table 3), and the training data, protocol, and dataset for that experiment are not described. The conclusion itself concedes that broader verification is needed, so the abstract's 'seamless adaptation to new tasks' claim should be softened or supported by a second held-out task and a
  3. [§4.1, Illumination Rectification] The training set for illumination rectification is synthesized by 'combining illumination images from DocShade with collected clean document images,' but the combination procedure is not specified. How are shadows and lighting overlaid? Is the illumination multiplied, added, or blended with spatially varying masks? What post-processing is applied? Since the reported gains on the real-world RealDAE and DocUNet benchmarks depend on this synthetic distribution matching camera-captured documents, the current description is insufficient to judge whether the results would transfer or to reproduce the experiments.
minor comments (4)
  1. [Table 2, DIBCO'18 row] The pFM value for the 'w/o PFM' column is 0.9250, while all other pFM values in the table are on a 0–100 scale (e.g., 91.59 and 93.84). This appears to be a units/typographical error and should be corrected, since it obscures the comparison.
  2. [Eq. (1)] The definition of α_t is circular: the text says 'α_t is a hyperparameter' and then defines α_t as a product over α_i with α_0=1. Please clarify whether α_t denotes the noise schedule and ᾱ_t the cumulative product, and fix the notation.
  3. [General] No code or trained models are provided. Given the number of underspecified components (filters, task embeddings, data synthesis), releasing code would substantially increase confidence in the results.
  4. [Figures and text] There are minor typos and inconsistencies: 'recitification' in Figure 1, 'Uni-Docdiff' capitalization in the contribution bullet, and the Figure 6 caption says 'each pair of adjacent images' where the figure shows triplets. Also, no error bars or statistical significance tests are reported for any of the quantitative comparisons; for a multi-task comparison paper this would strengthen the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Uni-DocDiff's central restoration claims are benchmarked against external models and held-out tests; self-citations are not load-bearing.

full rationale

Uni-DocDiff's central claim—unified diffusion-based restoration across six tasks—is supported by quantitative comparisons with external task-specific experts and DocRes on standard benchmarks (Table 1), not by a derivation from its own assumptions. The model is trained on public or disclosed training sets (TDD, SD7K/RDD, the RealDAE training split, DIBCO collections, EnsExam training split, Doc3D/UVDoc) and evaluated on distinct test sets (TDD eval, Jung's/Kligler's/OSR, RealDAE test, DIR300/DocUNet, DIBCO2018, EnsExam test), so the reported numbers are not fitted inputs renamed as predictions. The Prior Pool and PFM are architectural components whose contributions are tested by ablations (Table 2); the frequency-aware losses in Eq. (7) are supervised losses against ground truth, not circular self-definitions. No uniqueness theorem is imported from the authors' prior work, and no load-bearing claim reduces to a self-citation: citations to the authors' own papers (e.g., [26,52,68]) are generic references for diffusion/text capabilities, while the key baselines DocRes and DocDiff are external. At Sec. 5 the authors concede 'further experimental verification on a broader range of tasks is still needed'; this is an honest limitation of the scalability evidence, not a circular step. Similarly, the under-specified illumination-rectification synthesis in Sec. 4.1 is a reproducibility and data-distribution concern, not a circular reduction. Accordingly, no step in the derivation is equivalent to its input by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 3 invented entities

The central claim relies on a small number of hand-chosen hyperparameters (beta1, beta2), several standard diffusion assumptions, and three newly introduced architectural components. The handcrafted prior pool is a particularly notable assumption because it posits that classical image filters are sufficient to characterize task-relevant information across all six tasks. The dual-stream and PFM assumptions are supported only by ablations, not by a theoretical analysis.

free parameters (2)
  • beta1 (low-frequency loss weight) = 1
    Chosen by hand to balance the low-frequency-aware loss for deshadowing and illumination rectification.
  • beta2 (high-frequency loss weight) = 0.1
    Chosen by hand to balance the high-frequency-aware loss for deblurring, binarization, and handwriting removal.
assumptions (5)
  • standard math Standard DDPM forward and reverse processes (Eq. 1 and Eq. 3) are valid for conditional image generation.
    The diffusion formulation is taken directly from Ho et al. (DDPM) and used without proof.
  • domain assumption Predicting x0 rather than noise improves conditional generation quality for document restoration.
    The paper relies on DocDiff's argument that predicting x0 reduces diversity but enhances quality in conditional settings.
  • domain assumption Handcrafted features (Sobel, Canny, median, Gaussian, DCT) provide useful task-relevant priors that complement learned features.
    The Prior Pool design assumes these fixed operators capture the high-frequency and low-frequency cues needed by all tasks.
  • domain assumption A dual-stream architecture decoupling coordinate prediction (dewarping) from pixel prediction reduces task interference.
    The paper asserts this decoupling reduces interference and learning cost, but provides no formal analysis.
  • domain assumption Frequency-aware losses (Eq. 7-8) improve restoration by emphasizing task-relevant frequency bands.
    The loss design assumes that low-pass and high-pass filtering of the prediction error steers the model toward the desired output.
invented entities (3)
  • Prior Pool independent evidence
    purpose: A set of fixed handcrafted prior features (edges, illumination, frequency components) extracted from the degraded image and used to condition the diffusion model.
    Ablation in Table 2 shows performance drops when the Prior Pool is replaced with learnable parameters, providing in-paper falsification.
  • Prior Fusion Module (PFM) independent evidence
    purpose: Adaptively selects and combines task-specific and content-specific priors at each reconstruction stage to mitigate task interference.
    Ablation in Table 2 (w/o PFM) shows a clear performance drop, supporting its role.
  • Coordinate Prediction Branch (CPB) independent evidence
    purpose: A lightweight branch that predicts a sparse backward map for dewarping, using frozen encoder features from the diffusion model.
    Dewarping results on DocUNet and DIR300 validate the branch's effectiveness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uni-DocDiff: A Unified Document Restoration Model Based on Diffusion." pith.science (2026). https://pith.science/paper/3WFFAE3B

@misc{pith2026250804055,
  author       = {Pith},
  title        = {Pith review of: Uni-DocDiff: A Unified Document Restoration Model Based on Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WFFAE3B}},
  note         = {Machine review of arXiv:2508.04055}
}
read the original abstract

Removing various degradations from damaged documents greatly benefits digitization, downstream document analysis, and readability. Previous methods often treat each restoration task independently with dedicated models, leading to a cumbersome and highly complex document processing system. Although recent studies attempt to unify multiple tasks, they often suffer from limited scalability due to handcrafted prompts and heavy preprocessing, and fail to fully exploit inter-task synergy within a shared architecture. To address the aforementioned challenges, we propose Uni-DocDiff, a Unified and highly scalable Document restoration model based on Diffusion. Uni-DocDiff develops a learnable task prompt design, ensuring exceptional scalability across diverse tasks. To further enhance its multi-task capabilities and address potential task interference, we devise a novel \textbf{Prior \textbf{P}ool}, a simple yet comprehensive mechanism that combines both local high-frequency features and global low-frequency features. Additionally, we design the \textbf{Prior \textbf{F}usion \textbf{M}odule (PFM)}, which enables the model to adaptively select the most relevant prior information for each specific task. Extensive experiments show that the versatile Uni-DocDiff achieves performance comparable or even superior performance compared with task-specific expert models, and simultaneously holds the task scalability for seamless adaptation to new tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 72 canonical work pages

  1. [1]

    Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. 2021. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126 (2021)

  2. [2]

    Risab Biswas, Swalpa Kumar Roy, Ning Wang, Umapada Pal, and Guang-Bin Huang. 2023. DocBinFormer: A Two-Level Transformer Network for Effective Document Image Binarization. arXiv preprint arXiv:2312.03568 (2023)

  3. [3]

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. 2022. Simple baselines for image restoration. InEuropean conference on computer vision. Springer, 17–33

  4. [4]

    Marcos V Conde, Gregor Geigle, and Radu Timofte. 2024. High-quality image restoration following human instructions. arXiv preprint arXiv:2401.16468 (2024)

  5. [5]

    Sagnik Das, Ke Ma, Zhixin Shu, Dimitris Samaras, and Roy Shilkrot. 2019. De- warpnet: Single-image document unwarping with stacked 3d and 2d regression networks. In Proceedings of the IEEE/CVF international conference on computer vision. 131–140

  6. [6]

    Sagnik Das, Hassan Ahmed Sial, Ke Ma, Ramon Baldrich, Maria Vanrell, and Dimitris Samaras. 2020. Intrinsic decomposition of document images in-the-wild. arXiv preprint arXiv:2011.14447 (2020)

  7. [7]

    Fanbo Deng, Zheng Wu, Zheng Lu, and Michael S Brown. 2010. Binarizationshop: a user-assisted software suite for converting old documents to black-and-white. In Proceedings of the 10th annual joint conference on Digital libraries . 255–258

  8. [8]

    Hao Feng, Shaokai Liu, Jiajun Deng, Wengang Zhou, and Houqiang Li. 2023. Deep unrestricted document image rectification. IEEE Transactions on Multimedia 26 (2023), 6142–6154

Show all 81 references
  1. [9]

    Hao Feng, Yuechen Wang, Wengang Zhou, Jiajun Deng, and Houqiang Li. 2021. Doctr: Document image transformer for geometric unwarping and illumination correction. arXiv preprint arXiv:2110.12942 (2021)

  2. [10]

    Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. 2022. Geometric Representation Learning for Document Image Rectification. In Euro- pean Conference on Computer Vision . Springer, 475–492

  3. [11]

    Basilis Gatos, Konstantinos Ntirogiannis, and Ioannis Pratikakis. 2009. ICDAR 2009 document image binarization contest (DIBCO 2009). In 2009 10th Interna- tional conference on document analysis and recognition . IEEE, 1375–1382

  4. [12]

    Rachid Hedjam, Hossein Ziaei Nafchi, Reza Farrahi Moghaddam, Margaret Kalac- ska, and Mohamed Cheriet. 2015. Icdar 2015 contest on multispectral text extrac- tion (ms-tex 2015). In 2015 13th International Conference on Document Analysis and Recognition (ICDAR). IEEE, 1181–1185

  5. [13]

    Felix Hertlein and Alexander Naumann. 2023. Template-guided illumination correction for document images with imperfect geometric reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 904–913

  6. [14]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

  7. [15]

    Michal Hradiš, Jan Kotera, Pavel Zemcık, and Filip Šroubek. 2015. Convolutional neural networks for direct text deblurring. In Proceedings of BMVC, Vol. 10

  8. [16]

    Liufeng Huang, Bangdong Chen, Chongyu Liu, Dezhi Peng, Weiying Zhou, Yaqiang Wu, Hui Li, Hao Ni, and Lianwen Jin. 2023. EnsExam: A Dataset for Handwritten Text Erasure on Examination Papers. In International Conference on Document Analysis and Recognition. Springer, 470–485

  9. [17]

    Seungjun Jung, Muhammad Abul Hasan, and Changick Kim. 2018. Water-filling: An efficient algorithm for digitized document shadow removal. In Asian Confer- ence on Computer Vision . Springer, 398–414

  10. [18]

    Netanel Kligler, Sagi Katz, and Ayellet Tal. 2018. Document enhancement using visibility detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2374–2382

  11. [19]

    Ashutosh Kulkarni, Prashant W Patil, Subrahmanyam Murala, and Sunil Gupta

  12. [20]

    Gengluo Li, Huawen Shen, and Yu Zhou. 2025. Beyond Cropped Regions: New Benchmark and Corresponding Baseline for Chinese Scene Text Retrieval in Diverse Layouts. In ICML

  13. [21]

    Heng Li, Xiangping Wu, Qingcai Chen, and Qianjin Xiang. 2023. Foreground and text-lines aware document image rectification. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 19574–19583

  14. [22]

    Pu Li, Weize Quan, Jianwei Guo, and Dong-Ming Yan. 2023. Layout-aware single-image document flattening. ACM Transactions on Graphics 43, 1 (2023), 1–17

  15. [23]

    Shuai Li, Xiaolong Zheng, Kewen Lan, Ji Hu, Guangqin Wu, and Lihuan Shao. 2024. Scene handwritten text erasure based on multi-scale feature fusion. International Journal on Document Analysis and Recognition (IJDAR) (2024), 1–16

  16. [24]

    Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. 2019. Document rectification and illumination correction using a patch-based CNN. ACM Transactions on Graphics (TOG) 38, 6 (2019), 1–11

  17. [25]

    Zinuo Li, Xuhang Chen, Chi-Man Pun, and Xiaodong Cun. 2023. High-resolution document shadow removal via a large-scale real-world dataset and a frequency- aware shadow erasing net. In2023 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 12415–12424

  18. [26]

    Zhenhang Li, Yan Shu, Weichao Zeng, Dongbao Yang, and Yu Zhou. 2024. First creating backgrounds then rendering texts: A new paradigm for visual text blending. In ECAI 2024. IOS Press, 346–353

  19. [27]

    Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. 2025. Diffbir: Toward blind image restora- tion with generative diffusion prior. In European Conference on Computer Vision . Springer, 430–448

  20. [28]

    Yun-Hsuan Lin, Wen-Chin Chen, and Yung-Yu Chuang. 2020. Bedsr-net: A deep shadow removal network from a single document image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 12905–12914

  21. [29]

    Chongyu Liu, Lianwen Jin, Yuliang Liu, Canjie Luo, Bangdong Chen, Fengjun Guo, and Kai Ding. 2022. Don’t forget me: accurate background recovery for text removal via modeling local-global context. In European Conference on Computer Vision. Springer, 409–426

  22. [30]

    Shaokai Liu, Hao Feng, and Wengang Zhou. 2023. Rethinking Supervision in Document Unwarping: A Self-Consistent Flow-Free Approach. IEEE Transactions on Circuits and Systems for Video Technology 34, 6 (2023), 4817–4828

  23. [31]

    Yihao Liu, Xiangyu Chen, Xianzheng Ma, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. 2023. Unifying image processing as visual prompting question answering. arXiv preprint arXiv:2310.10513 (2023)

  24. [32]

    Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. 2024. Diff-Plugin: Revitalizing Details for Diffusion-based Low-level Tasks. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4197–4208

  25. [33]

    Jiahao Lyu, Wei Wang, Dongbao Yang, Jinwen Zhong, and Yu Zhou. 2025. Arbi- trary reading order scene text spotter with local semantics guidance. In Proceed- ings of the AAAI Conference on Artificial Intelligence , Vol. 39. 5919–5927

  26. [34]

    Jiaqi Ma, Tianheng Cheng, Guoli Wang, Qian Zhang, Xinggang Wang, and Lefei Zhang. 2023. Prores: Exploring degradation-aware visual prompt for universal image restoration. arXiv preprint arXiv:2306.13653 (2023)

  27. [35]

    Ke Ma, Sagnik Das, Zhixin Shu, and Dimitris Samaras. 2022. Learning from documents in the wild to improve document unwarping. InACM SIGGRAPH 2022 Conference Proceedings. 1–9

  28. [36]

    Ke Ma, Zhixin Shu, Xue Bai, Jue Wang, and Dimitris Samaras. 2018. DocUNet: Document Image Unwarping via a Stacked U-Net. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4700–4709

  29. [37]

    Bharat Mamidibathula and Prabir Kumar Biswas. 2019. SVDocNet: Spatially vari- ant U-Net for blind document deblurring. In Workshop on Document Intelligence at NeurIPS 2019

  30. [38]

    Jianhan Mei, Ziming Wu, Xiang Chen, Yu Qiao, Henghui Ding, and Xudong Jiang

  31. [39]

    Hossein Ziaei Nafchi, Seyed Morteza Ayatollahi, Reza Farrahi Moghaddam, and Mohamed Cheriet. 2013. An efficient ground truthing tool for binarization of historical manuscripts. In2013 12th International Conference on Document Analysis and Recognition. IEEE, 807–811

  32. [40]

    Konstantinos Ntirogiannis, Basilis Gatos, and Ioannis Pratikakis. 2014. ICFHR2014 competition on handwritten document image binarization (H-DIBCO 2014). In 2014 14th International conference on frontiers in handwriting recognition . IEEE, 809–813

  33. [41]

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan

  34. [42]

    Ioannis Pratikakis, Basilis Gatos, and Konstantinos Ntirogiannis. 2010. H-DIBCO 2010-handwritten document image binarization competition. In 2010 12th Inter- national Conference on Frontiers in Handwriting Recognition . IEEE, 727–732

  35. [43]

    Ioannis Pratikakis, Basilis Gatos, and Konstantinos Ntirogiannis. 2012. ICFHR 2012 competition on handwritten document image binarization (H-DIBCO 2012). In 2012 international conference on frontiers in handwriting recognition . IEEE, 817–822

  36. [44]

    Ioannis Pratikakis, Basilis Gatos, and Konstantinos Ntirogiannis. 2013. ICDAR 2013 document image binarization contest (DIBCO 2013). In 2013 12th Interna- tional Conference on Document Analysis and Recognition . IEEE, 1471–1476

  37. [45]

    Ioannis Pratikakis, Konstantinos Zagori, Panagiotis Kaddas, and Basilis Gatos

  38. [46]

    Ioannis Pratikakis, Konstantinos Zagoris, George Barlas, and Basilis Gatos. 2017. ICDAR2017 competition on document image binarization (DIBCO 2017). In 2017 14Th IAPR international conference on document analysis and recognition (ICDAR) , Vol. 1. IEEE, 1395–1403

  39. [47]

    Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. 2025. SPIRE: Semantic Prompt-Driven Image MM ’25, October 27–31, 2025, Dublin, Ireland Zhao et al. Restoration. In European Conference on Computer Vision . Springer, 446–464

  40. [48]

    Zhi Qiao, Yu Zhou, Dongbao Yang, Yucan Zhou, and Weiping Wang. 2020. SEED: Semantics enhanced encoder-decoder framework for scene text recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13528–13537

  41. [49]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

  42. [50]

    Huawen Shen, Xiang Gao, Jin Wei, Liang Qiao, Yu Zhou, Qiang Li, and Zhanzhan Cheng. 2023. Divide rows and conquer cells: towards structure recognition for large tables. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (Macao, P.R.Ch...

  43. [51]

    Huawen Shen, Gengluo Li, Jinwen Zhong, and Yu Zhou. 2025. LDP: Generalizing to multilingual visual information extraction by language decoupled pretraining. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 6805–6813

  44. [52]

    Yan Shu, Weichao Zeng, Fangmin Zhao, Zeyu Chen, Zhenhang Li, Xiaomeng Yang, Yu Zhou, Paolo Rota, Xiang Bai, Lianwen Jin, et al . 2025. Visual text processing: A comprehensive review and unified evaluation. arXiv preprint arXiv:2504.21682 (2025)

  45. [53]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

  46. [54]

    Mohamed Ali Souibgui and Yousri Kessentini. 2020. De-gan: A conditional generative adversarial network for document enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence 44, 3 (2020), 1180–1191

  47. [55]

    Hao Tang, Junyuan Guo, Teng Wang, Yanwei Yu, and Chao Wang. 2024. Efficient Joint Rectification of Photometric and Geometric Distortions in Document Images. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 3690–3694

  48. [56]

    Floor Verhoeven, Tanguy Magne, and Olga Sorkine-Hornung. 2023. UVDoc: Neural Grid-based Document Unwarping. In SIGGRAPH Asia 2023 Conference Papers. 1–11

  49. [57]

    Bingshu Wang and CL Philip Chen. 2020. Local water-filling algorithm for shadow detection and removal of document images. Sensors 20, 23 (2020), 6929

  50. [58]

    Biao Wang, Jiayi Li, Xin Jin, and Qiong Yuan. 2022. CHENet: image to image Chinese handwriting eraser. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 40–51

  51. [59]

    Ruilu Wang, Yang Xue, and Lianwen Jin. 2024. DocNLC: A Document Image En- hancement Framework with Normalized and Latent Contrastive Representation for Multiple Degradations. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 5563–5571

  52. [60]

    Wei Wang, Yu Zhou, Jiahao Lv, Dayan Wu, Guoqing Zhao, Ning Jiang, and Weipinng Wang. 2022. TPSNet: Reverse thinking of thin plate splines for arbitrary shape scene text representation. In Proceedings of the 30th ACM International Conference on Multimedia. 5014–5025

  53. [61]

    Yonghui Wang, Wengang Zhou, Zhenbo Lu, and Houqiang Li. 2022. Udoc-gan: Unpaired document illumination correction with background light prior. In Proceedings of the 30th ACM International Conference on Multimedia . 5074–5082

  54. [62]

    Zhou Wang, Eero P Simoncelli, and Alan C Bovik. 2003. Multi-scale Structural Similarity for Image Quality Assessment. In The Thrity-Seventh Asilomar Confer- ence on Signals, Systems & Computers, 2003 , Vol. 2. Ieee, 1398–1402

  55. [63]

    Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. 2021. Document Dewarping with Control Points. In Document Analysis and Recognition–ICDAR 2021: 16th International Conference, Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part I 16 . Springer, 466–480

  56. [64]

    Xiaomeng Yang, Zhi Qiao, and Yu Zhou. 2025. IPAD: Iterative, parallel, and diffusion-based network for scene text recognition. International Journal of Computer Vision (2025), 1–21

  57. [65]

    Zongyuan Yang, Baolin Liu, Yongping Xiong, and Guibin Wu. 2024. GDB: gated convolutions-based document binarization. Pattern Recognition 146 (2024), 109989

  58. [66]

    Zongyuan Yang, Baolin Liu, Yongping Xxiong, Lan Yi, Guibin Wu, Xiaojun Tang, Ziqi Liu, Junjie Zhou, and Xing Zhang. 2023. Docdiff: Document enhancement via residual diffusion models. In Proceedings of the 31st ACM international conference on multimedia. 2795–2806

  59. [67]

    Shaodi You, Yasuyuki Matsushita, Sudipta Sinha, Yusuke Bou, and Katsushi Ikeuchi. 2017. Multiview Rectification of Folded Documents. IEEE transactions on pattern analysis and machine intelligence 40, 2 (2017), 505–511

  60. [68]

    Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. 2024. TextCtrl: Diffusion-based scene text editing with prior guidance control.Advances in Neural Information Processing Systems 37 (2024), 138569–138594

  61. [69]

    Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. 2023. All- in-one multi-degradation image restoration network via hierarchical degradation representation. In Proceedings of the 31st ACM international conference on multi- media. 2285–2293

  62. [70]

    Demin Zhang, Jiahao Lyu, Zhijie Shen, and Yu Zhou. 2025. Class-Agnostic Region- of-Interest Matching in Document Images.arXiv preprint arXiv:2506.21055 (2025)

  63. [71]

    Jiaxin Zhang, Bangdong Chen, Hiuyi Cheng, Fengjun Guo, Kai Ding, and Lianwen Jin. 2023. DocAligner: Annotating real-world photographic document images by simply taking pictures. arXiv preprint arXiv:2306.05749 (2023)

  64. [72]

    Jiaxin Zhang, Lingyu Liang, Kai Ding, Fengjun Guo, and Lianwen Jin. 2023. Appearance enhancement for camera-captured document images in the wild. IEEE Transactions on Artificial Intelligence (2023)

  65. [73]

    Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. 2024. DocRes: A Generalist Model Toward Unifying Document Image Restoration Tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15654–15664

  66. [74]

    Ling Zhang, Yinghao He, Qing Zhang, Zheng Liu, Xiaolong Zhang, and Chunxia Xiao. 2023. Document image shadow removal guided by color-aware background. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. 1818–1827

  67. [75]

    Weiguang Zhang, Qiufeng Wang, and Kaizhu Huang. 2023. Polar-Doc: one-stage document dewarping with multi-scope constraints under polar representation. arXiv preprint arXiv:2312.07925 (2023)

  68. [76]

    Tianlun Zheng, Zhineng Chen, Shancheng Fang, Hongtao Xie, and Yu-Gang Jiang. 2024. CDistNet: Perceiving multi-domain character distance for robust text recognition. International Journal of Computer Vision 132, 2 (2024), 300–318

  69. [77]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks. In 2017 IEEE International Conference on Computer Vision (ICCV) . 2242–2251. doi:10. 1109/ICCV.2017.244

  70. [2018]

    In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR)

    ICFHR 2018 Competition on Handwritten Document Image Binarization (H-DIBCO 2018). In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR). IEEE, 489–493

  71. [2019]

    Multimedia tools and applications 78 (2019), 18869–18885

    Deepdeblur: text image recovery from blur to sharp. Multimedia tools and applications 78 (2019), 18869–18885

  72. [2022]

    IEEE Transactions on Multime- dia 25 (2022), 7686–7698

    Unified multi-weather visibility restoration. IEEE Transactions on Multime- dia 25 (2022), 7686–7698

  73. [2024]

    Advances in Neural Information Processing Systems 36 (2024)

    Promptir: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems 36 (2024)

Pith tools

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