Pith. sign in

REVIEW 3 major objections 5 minor 61 references

ForCenNet: Foreground-Centric Network for Document Image Rectification

T0 review · 3 major / 5 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 The paper claims that a network trained with foreground-centric labels—masks and line points from text, tables, and graphics—and a curvature consistency loss achieves state-of-the-art document dewarping on DocUNet, DIR300, WarpDoc, and…

desk verdict Solid foreground-aware dewarping recipe with a credible DOC3D-only variant, but the headline SOTA numbers are contaminated by training on the same clean images used to build the DocUNet and DIR300 tests. read the letter →

arxiv 2507.19804 v1 pith:M37JSULY submitted 2025-07-26 cs.CV

Peng Cai , Qiang Li , Kaicheng Yang , Dong Guo , Jia Li , Nan Zhou , Xiang An , Ninghua Yang
show 1 more author
Jiankang Deng
This is my paper · ORCID
classification cs.CV
keywords documentimagerectificationgeometricdistortioncorrectionforegroundsegmentationcurvatureconsistencylossmask-guidedtransformerOCRreadabilitydewarpingbenchmarksyntheticdeformationfield
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 tries to establish that document image rectification is improved when the model is explicitly told where the readable content is. It proposes ForCenNet, which generates foreground labels from undistorted images, deforms them synthetically, and uses a mask-guided transformer plus a curvature consistency loss to predict the backward mapping. On four real-world benchmarks—DocUNet, DIR300, WarpDoc, and DocReal—the authors report the best published numbers, including a CER of 0.136 and LD of 4.82 on DocUNet. If correct, the method matters because it turns clean reference images, which are easier to collect than paired distorted and ground-truth images, into a training signal that preserves readability and layout.

What carries the argument

The machinery is a foreground-centric training pipeline in three parts. A label generator uses a fine-tuned Hi-SAM segmenter for foreground masks, PaddleOCR midlines for text lines, and an LSD-based filter for table lines, then applies random backward mappings from DOC3D to synthesize distorted images, masks, and line sets. The transformer decoder injects a differentiable predicted foreground mask into self-attention as an additive term, concentrating attention on readable regions. A curvature consistency loss samples points every four pixels along extracted lines, maps them through the predicted and ground-truth backward fields, computes discrete curvature at each point, and penalizes the difference, targeting thin structures that pixel-wise L1 handles poorly.

What would settle it

Check whether the 365 undistorted images used to train the main model (Section 4.1) are the ground-truth reference images of the DocUNet and DIR300 test sets; if they are, retrain without them and see whether the reported MS-SSIM of 0.582 and 0.713 falls to the ForCenNet-DOC3D level.

Watch

Extended reading notes

Core claim

The central claim is that foreground structure—text, table lines, and graphics—is the right carrier of deformation information for dewarping. ForCenNet extracts character-level foreground masks and line-element points from undistorted images, forward-maps them through random distortions to create training pairs, and optimizes three objectives: an L1 foreground mask loss, an L1 backward-map regression loss, and a curvature consistency loss that aligns predicted and ground-truth curvature of sampled control lines. The authors report state-of-the-art results on DocUNet (MS-SSIM 0.582, LD 4.82, CER 0.136) and DIR300 (MS-SSIM 0.713, LD 4.65, CER 0.138), with competitive cross-domain numbers on WarpDoc and DocReal. A DOC3D-only trained variant also beats previous foreground-supervised methods on DocUNet, which the authors take as evidence that the foreground-centric design, not the training data alone, drives the gains.

Load-bearing premise

The reported top results on the two main benchmarks assume the model is allowed to train on clean versions of the same photographs it is later tested on; if that overlap is disallowed, the clean evidence comes only from the separately trained DOC3D variant.

Editorial extensions

If this is right

  • On DocUNet, the method reports MS-SSIM 0.582, LD 4.82, and CER 0.136, improving over prior methods and lowering OCR error.
  • On DIR300, it reports MS-SSIM 0.713 and LD 4.65, the best numbers the paper compares against.
  • Training with only DOC3D undistorted images still beats all prior foreground-supervised methods on DocUNet, suggesting the foreground labels are the source of gain.
  • The label-generation pipeline lets practitioners adapt dewarping to new document domains by supplying undistorted reference images, without paired distorted and clean data.
  • The predicted foreground mask also enables a downstream illumination-enhancement application, whitening background regions while preserving foreground color, with MS-SSIM 0.6712 on DocUNet enhancement.

Reading between the lines

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

  • A testable extension is to apply the mask-guided attention and curvature loss to scene-text rectification or curved text-line straightening, where thin structures dominate; the same loss may transfer directly.
  • The ablation plateau in the dataset-scaling table suggests the bottleneck is not more synthetic samples but foreground label quality or deformation-field diversity, so a follow-up could vary distortion types rather than sample count.
  • Since the method trains on undistorted images, one could bootstrap from a large unlabeled web corpus by using the model's own mask predictions as pseudo-labels, though the paper does not test self-training.
  • The reported bias analysis, which selects a 40% sampling ratio for forward-backward mapping, gives a quantitative recipe usable in other geometric warping tasks.
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 / 5 minor

Summary. ForCenNet is a document image rectification network that trains on undistorted images only, using a synthetic distortion field pipeline. The method extracts foreground masks and line elements from undistorted images, applies random forward-map deformations to create training pairs, and trains a mask-guided transformer decoder with a curvature consistency loss. The paper reports state-of-the-art results on four real-world benchmarks: DocUNet, DIR300, WarpDoc, and DocReal, and shows quantitative gains in MS-SSIM, local distortion, and OCR error rates. The central claim is that foreground-centric supervision improves geometric rectification and readability.

Significance. If the results hold, the training-from-undistorted-images paradigm is practically valuable because it avoids the need for paired distorted/undistorted data, which is expensive to acquire. The DOC3D-trained variant and the WarpDoc/DocReal generalization results are credible and suggest the method is competitive with prior art. The design of a foreground-centric mask mechanism and a curvature consistency loss is well motivated and the paper includes several useful ablations. However, the headline SOTA claims on DocUNet and DIR300 are weakened by the fact that the main model is trained on undistorted images from those same benchmarks, so the strength of the central claim depends on a clean retraining.

major comments (3)
  1. [Section 4.1, Tables 1-2] The main ForCenNet model is trained on 365 undistorted images from DocUNet and DIR300, which are exactly the source documents of the distorted test images in those benchmarks. Because the test inputs are generated from the same document content, content memorization can inflate all metrics (MS-SSIM, LD, ED, CER). Consequently, the gains of ForCenNet over ForCenNet-DOC3D (e.g., 0.582 vs 0.579 MS-SSIM and 0.136 vs 0.158 CER on DocUNet) do not establish a clean state-of-the-art on these two benchmarks. Please retrain the model without these images or report the DOC3D-trained variant as the primary result for DocUNet and DIR300, and re-frame the abstract's SOTA claim accordingly.
  2. [Section 4.4] The ablation studies are performed on the DocUNet benchmark using 65 undistorted images from DocUNet to construct all training groups, and the evaluation set is derived from the same document content. The relative improvements reported in Table 5 (e.g., the mask-guided decoder raising MS-SSIM from 0.558 to 0.571) may therefore reflect overfitting to the test content rather than a genuinely transferable effect. Please rerun the ablations on a held-out set (or using DOC3D-based training data) to validate that the mask-guided decoder and curvature loss generalize to unseen document content.
  3. [Equation (6)] The curvature consistency loss is defined as Lk = (1/(N-1)) * sum_{i}(k_i_hat - k_i). As written, this is a signed mean, so positive and negative curvature errors cancel, making the loss zero even for large opposite errors. This is not a valid regression objective. The intended form is likely the mean absolute error or mean squared error. Please correct the equation and confirm that the implementation matches the corrected form.
minor comments (5)
  1. [Equation (2)] The softmax expression in Eq. (2) is ambiguous about the dimension over which the softmax is computed; since M has Cseg=2 channels, please clarify that the softmax is over the channel dimension and that the subscript i ranges over {0,1}.
  2. [Algorithm 1] The text says the algorithm 'eliminates non-horizontal and non-vertical lines', but the condition slope < alpha or slope > beta actually keeps near-horizontal and near-vertical lines; the wording should be corrected to match the algorithm's behavior.
  3. [Table 4] The description of the x1000 group as '1,000 randomly generated deformation fields to the same image' is inconsistent with the surrounding text that says 65 undistorted images are used; please specify how many unique source images each group contains.
  4. [References] Reference [52] has the venue 'W ACV', which should be 'WACV'; other references have minor formatting inconsistencies (e.g., duplicated 'Neural Information Processing Systems' in [42]).
  5. [Section 2.1] DR-Net [27] is described as using undistorted images as direct supervisory signals, which closely resembles the proposed approach; the authors should explicitly state the difference in supervision methodology to avoid confusion.

Circularity Check

1 steps flagged · score 4.0 of 10

Main SOTA numbers on DocUNet/DIR300 are produced by training on the same benchmarks' clean source images, a partial fit-to-prediction overlap; the DOC3D variant and cross-domain tables remain independent.

  1. fitted input called prediction [Section 4.1 Implementation Details; Section 4.4 Dataset Scaling; Tables 1 and 2]
    "We train our model on two distinct undistorted datasets. The first, referred to as ForCenNet, comprises 365 images from DocUNet [29] and DIR300 [11]. ... A total of 65 undistorted images from DocUNet [29] are used to construct experimental groups of varying sizes. ... Result comparisons between our proposed with existing methods on the DocUNet Benchmark [29]."

    The main ForCenNet model is trained on the undistorted source images of the two evaluation benchmarks it then reports as state-of-the-art. The DocUNet benchmark evaluates on distorted images generated from the same clean source documents, and Section 4.4 confirms that those 65 DocUNet undistorted images are used as training images; DIR300 is also included in the 365-image training set. Thus the Table 1-2 DocUNet/DIR300 numbers for ForCenNet are not predictions on unseen content: the network has been optimized to output the very images that serve as evaluation targets, so the reported MS-SSIM, LD, and CER gains over the DOC3D-only variant partly measure memorization of training content rather than generalizable dewarping.

full rationale

The method's internal derivation is not circular: foreground labels are extracted from undistorted images and warped by DOC3D-derived deformation fields, and the mask loss, backward-map L1, and curvature loss all supervise against externally generated ground truth (DOC3D BM and warped masks/lines), not against the benchmark metrics. The SPW attention module is self-cited [17] but is a complexity-reduction component rather than the source of the SOTA claim, and no uniqueness theorem is imported from the authors' prior work. The one load-bearing overlap is the training/evaluation dataset identity: Section 4.1 trains ForCenNet on 365 images from DocUNet and DIR300, and Section 4.4 confirms the 65 DocUNet undistorted images are used for training, while Tables 1-2 evaluate on those same benchmarks whose ground truths are these clean source documents. This makes the main model's DocUNet/DIR300 numbers partly forced by content memorization. However, ForCenNet-DOC3D, trained exclusively on DOC3D, also surpasses prior methods on both benchmarks, and the WarpDoc/DocReal results are cross-domain, so the central claim retains independent support. A separate text bug (Eq. 6 writes a signed mean of curvature differences rather than an absolute value) is an implementation/documentation inconsistency, not a circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper contributes an empirical deep learning method rather than a derivation. The listed free parameters are hand-set hyperparameters; no physical constants are fitted. The central assumptions are the realism of DOC3D warps and the reliability of Hi-SAM labels. No new physical or categorical entities are introduced.

free parameters (6)
  • Mask smoothing coefficient gamma = 0.8
    Chosen by hand in Table 6; controls softmax sharpness of predicted foreground mask and affects mask guidance.
  • Scaling factor sigma for mask attention = 0.005
    Hand-set in Table 6; scales the mask term in self-attention.
  • Curvature denominator epsilon = 0.0001
    Added to avoid division by zero in curvature computation; hand-set.
  • Line sampling interval = 4 pixels
    Selected via sweep (2,4,8,16,32) as trade-off between fidelity and efficiency; reported in supplementary Section 6.
  • Forward-map sampling ratio = 40%
    Chosen from bias analysis in supplementary Section 9 to balance mapping bias.
  • LSD line filtering thresholds (epsilon_s, delta, alpha, beta) = not specified
    Algorithm 1 requires slope and intercept tolerances; exact values are not given, which affects reproducibility.
assumptions (4)
  • domain assumption DOC3D backward maps are a representative prior for real-world document warps
    All training deformations are sampled from DOC3D's 100,000 backward maps (Section 3.1); if this prior underrepresents real distortions, the synthetic training will not transfer. WarpDoc/DocReal cross-domain results are the only direct evidence for transfer.
  • domain assumption Fine-tuned Hi-SAM segmentation of undistorted images is accurate enough for label generation
    Foreground masks and line elements are extracted from clean images (Section 3.1); errors in these labels propagate into the distorted training labels and mask supervision.
  • ad hoc to paper Curvature consistency loss improves geometric fidelity
    Proposed in Section 3.3 to supervise thin structures; the ablation (Table 5) supports this empirically, but the mechanism is a modeling choice rather than a derived guarantee.
  • standard math Bilinear interpolation and central differences correctly model point projection and curvature
    Used in Eq. 4-5; standard and differentiable, but the discrete derivative approximation at boundaries introduces small errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ForCenNet: Foreground-Centric Network for Document Image Rectification." pith.science (2026). https://pith.science/paper/M37JSULY

@misc{pith2026250719804,
  author       = {Pith},
  title        = {Pith review of: ForCenNet: Foreground-Centric Network for Document Image Rectification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M37JSULY}},
  note         = {Machine review of arXiv:2507.19804}
}
read the original abstract

Document image rectification aims to eliminate geometric deformation in photographed documents to facilitate text recognition. However, existing methods often neglect the significance of foreground elements, which provide essential geometric references and layout information for document image correction. In this paper, we introduce Foreground-Centric Network (ForCenNet) to eliminate geometric distortions in document images. Specifically, we initially propose a foreground-centric label generation method, which extracts detailed foreground elements from an undistorted image. Then we introduce a foreground-centric mask mechanism to enhance the distinction between readable and background regions. Furthermore, we design a curvature consistency loss to leverage the detailed foreground labels to help the model understand the distorted geometric distribution. Extensive experiments demonstrate that ForCenNet achieves new state-of-the-art on four real-world benchmarks, such as DocUNet, DIR300, WarpDoc, and DocReal. Quantitative analysis shows that the proposed method effectively undistorts layout elements, such as text lines and table borders. The resources for further comparison are provided at https://github.com/caipeng328/ForCenNet.

Figures

Figures reproduced from arXiv: 2507.19804 by the authors.

Figure 1
Figure 1. Visualization of Deformation Degree. We compute the deformation displacement for each pixel and visualize it using a heatmap. Blue represents minor deformation, while red indicates significant displacement. a significant challenge in the field of document analysis and recognition. With advances in deep learning, data-driven ap￾proaches [7, 9, 43] have been developed for geometric cor￾rection of distorted document im… view at source ↗
Figure 2
Figure 2. The overview architecture of the proposed ForCenNet. M is the predicted foreground mask. k is the curvature value calculated from line elements, with kˆi as the predicted value and ki as the ground truth. BMˆ is the predicted backward mapping field. layers. We adopt overlapping patch embeddings [40] and employ a kernel size of 3 and a stride of 2 to preserve fea￾ture information at text boundaries while reducing com… view at source ↗
Figure 3
Figure 3. Qualitative Comparison with Prior Methods on DocUNet and DIR300 Benchmarks. Red arrows highlight the differences. Additional visualizations are available in the appendix. ements like text and images, table lines occupy fewer pixels, making it challenging for the network to capture line distor￾tion. This imbalance weakens the supervisory effect of the L1 loss. Furthermore, although L1 loss governs pixel-level distort… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Visualization of Foreground Segmentation. From left to right: the distorted input image, segmentation results from the frozen model and the differentiable model, followed by the corre￾sponding dewarped outputs generated by each model. iterative training with domain-spe…
Figure 5
Figure 5. Figure 5: Visualization of Foreground Results on Warp￾Doc [49] and DocReal [53] benchmarks. Columns 1-3: de￾tection results of line elements. Columns 4-6: detection results of text elements. Highlighted colors indicate detected regions, while red arrows mark differences [PITH_F…
Figure 6
Figure 6. Figure 6: Quantitative evaluation of straight-line rectification. The first image displays results on the DocReal [53] dataset, while the second image presents results on the WarpDoc [49] dataset. performance is further improved. Ablation on Different Segmentation Models. To ass…
Figure 9
Figure 9. Figure 9: Exploration of enhancement tasks. 9. Bias Analysis In the label preprocessing module, the forward map (FM) is derived by proportionally sampling anchor points on the backward map (BM) and constructing an augmentation matrix. However, a certain bias arises due to incomp…
Figure 10
Figure 10. Figure 10: Statistic analysis of bias. 10. More visualizations We present additional comparisons of the model’s dewarp￾ing results in [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Visualization comparison on the DocUNet dataset [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Visualization comparison on the DIR300 dataset [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Visualization comparison on the DocReal dataset [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Visualization comparison on the WarpDoc dataset [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 55 canonical work pages

  1. [1]

    Multi-label cluster discrimination for visual representation learning

    Xiang An, Kaicheng Yang, Xiangzi Dai, Ziyong Feng, and Jiankang Deng. Multi-label cluster discrimination for visual representation learning. In ECCV, pages 428–444. Springer,

    work page

  2. [2]

    A gated and bifurcated stacked u-net module for document image dewarping

    Hmrishav Bandyopadhyay, Tanmoy Dasgupta, Nibaran Das, and Mita Nasipuri. A gated and bifurcated stacked u-net module for document image dewarping. In 2020 25th Inter- national Conference on Pattern Recognition (ICPR) , pages 10548–10554. IEEE, 2021. 2

    work page 2020

  3. [3]

    Brown and W.B

    M.S. Brown and W.B. Seales. Document restoration using 3d shape: a general deskewing algorithm for arbitrarily warped documents. In Proceedings Eighth IEEE International Con- ference on Computer Vision. ICCV 2001, 2002. 2

    work page 2001

  4. [4]

    Rectify- ing the bound document image captured by the camera: a model based approach

    Huaigu Cao, Xiaoqing Ding, and Changsong Liu. Rectify- ing the bound document image captured by the camera: a model based approach. In Seventh International Conference on Document Analysis and Recognition, 2003. Proceedings., page 71–75, 2004. 2

    work page 2003

  5. [5]

    Numerical meth- ods for engineers

    Steven C Chapra, Raymond P Canale, et al. Numerical meth- ods for engineers. Mcgraw-hill New York, 2011. 5

    work page 2011

  6. [6]

    MataDoc: Margin and Text Aware Document Dewarping for Arbitrary Boundary

    Beiya Dai, Qunyi Xie, Yulin Li, Xiameng Qin, Chengquan Zhang, Kun Yao, Junyu Han, et al. Matadoc: margin and text aware document dewarping for arbitrary boundary. arXiv preprint arXiv:2307.12571, 2023. 6

    work page Pith review arXiv 2023

  7. [7]

    Dewarpnet: Single-image document unwarping with stacked 3d and 2d regression networks

    Sagnik Das, Ke Ma, Zhixin Shu, Dimitris Samaras, and Roy Shilkrot. Dewarpnet: Single-image document unwarping with stacked 3d and 2d regression networks. In ICCV, pages 131–140, 2019. 1, 2, 3, 6

    work page 2019

  8. [8]

    End- to-end piece-wise unwarping of document images

    Sagnik Das, Kunwar Yashraj Singh, Jon Wu, Erhan Bas, Vi- jay Mahadevan, Rahul Bhotika, and Dimitris Samaras. End- to-end piece-wise unwarping of document images. In ICCV, pages 4268–4277, 2021. 2, 6

    work page 2021

Show all 61 references
  1. [9]

    Doctr: Document image transformer for geo- metric unwarping and illumination correction.arXiv preprint arXiv:2110.12942, 2021

    Hao Feng, Yuechen Wang, Wengang Zhou, Jiajun Deng, and Houqiang Li. Doctr: Document image transformer for geo- metric unwarping and illumination correction.arXiv preprint arXiv:2110.12942, 2021. 1, 2, 4, 6, 7

    2021 arXiv

  2. [10]

    Docscanner: Robust document image rectification with progressive learning

    Hao Feng, Wengang Zhou, Jiajun Deng, Qi Tian, and Houqiang Li. Docscanner: Robust document image rectification with progressive learning. arXiv preprint arXiv:2110.14968, 2021. 2

    2021 arXiv

  3. [11]

    Geometric representation learning for document image rectification

    Hao Feng, Wengang Zhou, Jiajun Deng, Yuechen Wang, and Houqiang Li. Geometric representation learning for document image rectification. In ECCV, pages 475–492. Springer, 2022. 2, 3, 4, 6, 7

    2022

  4. [12]

    Accurate, large minibatch sgd: Training imagenet in 1 hour

    Priya Goyal, Piotr Doll ´ar, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6

    2017 arXiv

  5. [13]

    Breaking the modality barrier: Universal embedding learning with multimodal llms

    Tiancheng Gu, Kaicheng Yang, Ziyong Feng, Xingjun Wang, Yanzhao Zhang, Dingkun Long, Yingda Chen, Wei- dong Cai, and Jiankang Deng. Breaking the modality barrier: Universal embedding learning with multimodal llms. arXiv preprint arXiv:2504.17432, 2025. 2

    2025

  6. [14]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 1

    2017

  7. [15]

    Template-guided illumination correction for document images with imperfect geometric reconstruction

    Felix Hertlein and Alexander Naumann. Template-guided illumination correction for document images with imperfect geometric reconstruction. In ICCV, pages 904–913, 2023. 2

    2023

  8. [16]

    Densely connected convolutional net- works

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kil- ian Q Weinberger. Densely connected convolutional net- works. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 1

    2017

  9. [17]

    A transformer-based architecture for high-resolution stereo matching

    Di Jia, Peng Cai, Qian Wang, and Ninghua Yang. A transformer-based architecture for high-resolution stereo matching. IEEE Transactions on Computational Imaging ,

  10. [18]

    Revisiting document image dewarp- ing by grid regularization

    Xiangwei Jiang, Rujiao Long, Nan Xue, Zhibo Yang, Cong Yao, and Gui-Song Xia. Revisiting document image dewarp- ing by grid regularization. InCVPR, pages 4543–4552, 2022. 2, 3, 6

    2022

  11. [19]

    Bilinear interpola- tion

    Earl J Kirkland and Earl J Kirkland. Bilinear interpola- tion. Advanced computing in electron microscopy , pages 261–263, 2010. 5

    2010

  12. [20]

    Active contours network to straighten distorted text lines

    Olivier Lavialle, X Molines, Franck Angella, and Pierre Bay- lou. Active contours network to straighten distorted text lines. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205) , pages 748–751. IEEE, 2001. 2

    2001

  13. [21]

    Binary coors capable or ‘correcting dele- tions, insertions, and reversals

    VI Lcvenshtcin. Binary coors capable or ‘correcting dele- tions, insertions, and reversals. In Soviet physics-doklady,

  14. [22]

    Foreground and text-lines aware document image rectifica- tion

    Heng Li, Xiangping Wu, Qingcai Chen, and Qianjin Xiang. Foreground and text-lines aware document image rectifica- tion. In ICCV, pages 19574–19583, 2023. 2, 3, 6

    2023

  15. [23]

    Layout-aware single-image document flattening

    Pu Li, Weize Quan, Jianwei Guo, and Dong-Ming Yan. Layout-aware single-image document flattening. ACM Transactions on Graphics, 43(1):1–17, 2023. 2, 3, 6

    2023

  16. [24]

    Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V . Sander. Docu- ment rectification and illumination correction using a patch- based cnn. ACM Transactions on Graphics , page 1–11,

  17. [25]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision , pages 740–755. Springer, 2014. 6

    2014

  18. [26]

    Sift flow: Dense correspondence across scenes and its applications

    Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE transactions on pattern analysis and machine intelligence , 33(5):978–994, 2010. 6

    2010

  19. [27]

    Rethinking su- pervision in document unwarping: A self-consistent flow- free approach

    Shaokai Liu, Hao Feng, and Wengang Zhou. Rethinking su- pervision in document unwarping: A self-consistent flow- free approach. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 2, 6

    2023

  20. [28]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Cornell University - arXiv,Learning, 2017. 6

    2017

  21. [29]

    Docunet: Document image unwarping via a stacked u- net

    Ke Ma, Zhixin Shu, Xue Bai, Jue Wang, and Dimitris Sama- ras. Docunet: Document image unwarping via a stacked u- net. In CVPR, pages 4700–4709, 2018. 2, 6, 7

    2018

  22. [30]

    Learning from documents in the wild to improve document unwarping

    Ke Ma, Sagnik Das, Zhixin Shu, and Dimitris Samaras. Learning from documents in the wild to improve document unwarping. In SIGGRAPH, pages 1–9, 2022. 1, 2, 6

    2022

  23. [31]

    Metric rectification of curved document images

    Gaofeng Meng, Chunhong Pan, Shiming Xiang, Jiangyong Duan, and Nanning Zheng. Metric rectification of curved document images. IEEE transactions on pattern analysis and machine intelligence, 34(4):707–722, 2011. 2

    2011

  24. [32]

    From wer and ril to mer and wil: improved evaluation mea- sures for connected speech recognition

    Andrew Cameron Morris, Viktoria Maier, and Phil Green. From wer and ril to mer and wil: improved evaluation mea- sures for connected speech recognition. In Interspeech 2004,

    2004

  25. [33]

    Paddleocr: A practical and easy-to-use ocr tool for multilingual scenarios

    PaddlePaddle. Paddleocr: A practical and easy-to-use ocr tool for multilingual scenarios. https://github.com/ PaddlePaddle/PaddleOCR, 2020. 3

    2020

  26. [34]

    Automatic differentiation in pytorch

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Al- ban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 6

    2017

  27. [35]

    B´ezier and B-spline techniques

    Hartmut Prautzsch, Wolfgang Boehm, and Marco Paluszny. B´ezier and B-spline techniques. Springer, 2002. 2

    2002

  28. [36]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  29. [37]

    Smith and Nicholay Topin

    Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, 2019. 6

    2019

  30. [38]

    An overview of the tesseract ocr engine

    Ray Smith. An overview of the tesseract ocr engine. In Ninth international conference on document analysis and recogni- tion (ICDAR 2007), pages 629–633. IEEE, 2007. 8

    2007

  31. [39]

    Raft: Recurrent all-pairs field transforms for optical flow (extended abstract)

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow (extended abstract). In Proceed- ings of the Thirtieth International Joint Conference on Arti- ficial Intelligence, 2021. 2

    2021

  32. [40]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In International conference on machine learning , pages 10347–10357. PMLR, 2021. 4

    2021

  33. [41]

    Document image dewarping using robust estimation of curled text lines

    Adrian Ulges, Christoph H Lampert, and Thomas M Breuel. Document image dewarping using robust estimation of curled text lines. InEighth International Conference on Doc- ument Analysis and Recognition (ICDAR’05) , pages 1001–

  34. [42]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, AidanN. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Neural Information Processing Systems,Neural Information Processing Systems,

  35. [43]

    Uvdoc: Neural grid-based document unwarping

    Floor Verhoeven, Tanguy Magne, and Olga Sorkine- Hornung. Uvdoc: Neural grid-based document unwarping. In SIGGRAPH, pages 1–11, 2023. 1, 2, 3, 6, 7

    2023

  36. [44]

    Lsd: A fast line seg- ment detector with a false detection control

    Rafael Grompone V on Gioi, Jeremie Jakubowicz, Jean- Michel Morel, and Gregory Randall. Lsd: A fast line seg- ment detector with a false detection control. IEEE trans- actions on pattern analysis and machine intelligence, 32(4): 722–732, 2008. 3

    2008

  37. [45]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision , p...

    2021

  38. [46]

    Wang, A.C

    Z. Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Im- age quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing , page 600–612, 2004. 6

    2004

  39. [47]

    Document image de- warping for text/graphics recognition

    Changhua Wu and Gady Agam. Document image de- warping for text/graphics recognition. In Structural, Syn- tactic, and Statistical Pattern Recognition: Joint IAPR Inter- national Workshops SSPR 2002 and SPR 2002 Windsor, On- tario, Canada, August 6–9, 2002 Proceedings , pages 348–

    2002

  40. [48]

    Document dewarping with control points

    Guo-Wang Xie, Fei Yin, Xu-Yao Zhang, and Cheng-Lin Liu. Document dewarping with control points. In ICDAR, pages 466–480. Springer, 2021. 2, 3, 6

    2021

  41. [49]

    Fourier document restoration for robust document dewarping and recognition

    Chuhui Xue, Zichen Tian, Fangneng Zhan, Shijian Lu, and Song Bai. Fourier document restoration for robust document dewarping and recognition. In CVPR, pages 4573–4582,

  42. [50]

    Clip-cid: Efficient clip distillation via cluster-instance discrimination

    Kaicheng Yang, Tiancheng Gu, Xiang An, Haiqiang Jiang, Xiangzi Dai, Ziyong Feng, Weidong Cai, and Jiankang Deng. Clip-cid: Efficient clip distillation via cluster-instance discrimination. In AAAI, pages 21974–21982, 2025. 2

    2025

  43. [51]

    Hi-sam: Marrying segment anything model for hierarchical text segmentation

    Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, and Dacheng Tao. Hi-sam: Marrying segment anything model for hierarchical text segmentation. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 47(03):1431–1447, 2025. 3

    2025

  44. [52]

    Multiview rectification of folded documents

    Shaodi You, Yasuyuki Matsushita, Sudipta Sinha, Yusuke Bou, and Katsushi Ikeuchi. Multiview rectification of folded documents. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, page 505–511, 2018. 6

    2018

  45. [53]

    Docreal: Robust document dewarping of real-life images via attention-enhanced control point prediction

    Fangchen Yu, Yina Xie, Lei Wu, Yafei Wen, Guozhi Wang, Shuai Ren, Xiaoxin Chen, Jianfeng Mao, and Wenye Li. Docreal: Robust document dewarping of real-life images via attention-enhanced control point prediction. InWACV, pages 665–674, 2024. 7, 8

    2024

  46. [54]

    Marior: Margin removal and iterative content rec- tification for document dewarping in the wild.arXiv preprint arXiv:2207.11515, 2022

    Jiaxin Zhang, Canjie Luo, Lianwen Jin, Fengjun Guo, and Kai Ding. Marior: Margin removal and iterative content rec- tification for document dewarping in the wild.arXiv preprint arXiv:2207.11515, 2022. 2, 6

    2022 arXiv

  47. [55]

    Docres: A generalist model toward unifying document image restoration tasks

    Jiaxin Zhang, Dezhi Peng, Chongyu Liu, Peirong Zhang, and Lianwen Jin. Docres: A generalist model toward unifying document image restoration tasks. In CVPR, pages 15654– 15664, 2024. 2, 3, 6, 7, 8

    2024

  48. [56]

    Pub- laynet: largest dataset ever for document layout analysis

    Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Pub- laynet: largest dataset ever for document layout analysis. In 2019 International conference on document analysis and recognition (ICDAR), pages 1015–1022. IEEE, 2019. 2 ForCenNet: Foreground-Centric Network for Document Im...

    2019

  49. [57]

    To investigate the impact of sam- pling density on geometric fidelity, we evaluate line sam- pling intervals of 2, 4, 8, 16, and 32 pixels on the DocUNet dataset

    Experimental Details In the Tab.6, we present the detailed experimental setup and model hyperparameters. To investigate the impact of sam- pling density on geometric fidelity, we evaluate line sam- pling intervals of 2, 4, 8, 16, and 32 pixels on the DocUNet dataset. The corre...

  50. [58]

    Data augmentation Figure 8. Visualization of the cropping process To enhance the diversity of distortion relationships, we apply minor cropping to the generated distorted images and compute the correspondences of the deformation field be- fore and after cropping. Fig. 8 shows ...

  51. [59]

    This task aims to restore a pristine appearance similar to that produced by a scanner or digital PDF file, without being limited to specific degra- dation types

    Downstream tasks We conduct an exploratory investigation into a significant downstream task, namely appearance enhancement, also known as illumination correction. This task aims to restore a pristine appearance similar to that produced by a scanner or digital PDF file, without...

  52. [60]

    However, a certain bias arises due to incomplete sampling

    Bias Analysis In the label preprocessing module, the forward map ( F M) is derived by proportionally sampling anchor points on the backward map ( BM) and constructing an augmentation matrix. However, a certain bias arises due to incomplete sampling. To quantify this bias, we a...

  53. [61]

    11 , 12 ,13 and 14, which effectively demonstrate the superiority of our approach

    More visualizations We present additional comparisons of the model’s dewarp- ing results in Fig. 11 , 12 ,13 and 14, which effectively demonstrate the superiority of our approach. Red arrows highlight the differences. Figure 11. Visualization comparison on the DocUNet dataset....

Pith tools

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