Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Why Stop at Words? Unveiling the Bigger Picture through Line-Level OCR

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that recognizing whole text lines instead of individual words improves end-to-end OCR accuracy by 5.4% and speeds up inference four-fold.

desk verdict Useful benchmark and a plausible idea, but the headline 5.4% gain is confounded by training-data mismatch and an order-insensitive metric; should go to review with a demand for controlled comparisons. read the letter →

arxiv 2508.21693 v1 pith:7I3VEE42 submitted 2025-08-29 cs.CV cs.AIcs.CLcs.LG

Shashank Vempati , Nishit Anand , Gaurav Talebailkar , Arpan Garai , Chetan Arora This is my paper
classification cs.CVcs.AIcs.CLcs.LG
keywords line-levelOCRdocumenttextrecognitionworddetectionPARSeqpermutedautoregressivemodelend-to-endbenchmarksynthetictrainingdataflexiblecharacteraccuracy
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

The paper argues that the main remaining bottleneck in document OCR is word segmentation, not character recognition, and proposes skipping word detection entirely by recognizing full text lines with a sequence model. By fine-tuning PARSeq on synthetically generated line images and pairing it with the Kraken line detector, the authors report a Flexible Character Accuracy of 97.62% on their new 251-page English benchmark, versus 92.15% for the strongest word-based pipeline (CRAFT + PARSeq), along with a four-fold speedup. They also contribute a line-annotated page-image dataset to enable end-to-end page-level evaluation. The headline accuracy is measured with an order-independent metric; the paper openly shows that order-sensitive metrics make the gain much smaller unless lines are reordered correctly.

What carries the argument

The central object is PARSeq_line: the PARSeq permuted-autoregressive sequence model, originally built for word recognition, re-trained on synthetic line-level images (32×400 pixels) with a charset that includes spaces. Its permutation language modeling and iterative bidirectional refinement let the decoder use both left and right context within a line, which is what drives the accuracy gain. Kraken provides line detection, and the FCA metric of Clausner et al. is the evaluation device that makes the order-independent comparison possible.

What would settle it

Recompute both pipelines' accuracy using strict order-sensitive CRR on a page-level benchmark with reading-order ground truth, e.g., by matching recognized lines against ground-truth lines ordered by human reading flow. The paper's own Table 3 shows the line pipeline's CRR swings from 85.76 (default) to 96.27 (oracle order), so this experiment would settle whether the 5.4% FCA gain reflects true end-to-end superiority or an artifact of ignoring order.

Watch

Extended reading notes

Core claim

The central claim is that OCR should operate at the line level: detect text lines, then feed each line image directly to a permuted-autoregressive sequence model that outputs the full line of text, eliminating the word-detection stage. The authors show that this approach outperforms every tested word-level pipeline on their end-to-end benchmark: with Kraken line detection and a line-fine-tuned PARSeq, they achieve 97.62 Flexible Character Accuracy versus 92.15 for CRAFT+PARSeq, and 0.53s versus 2.11s inference per page. The mechanism is that the recognizer conditions each character prediction on the whole line, leveraging sentence-level context, and the pipeline no longer inherits errors fro

Load-bearing premise

The headline 5.4% accuracy gain is measured with Flexible Character Accuracy, which ignores reading order; if reading order is part of what end-to-end OCR must get right, the gain is far smaller and the default-order CRR (85.76) is actually below Tesseract's 88.16.

Editorial extensions

If this is right

  • Word detection models could be removed from document OCR pipelines without an accuracy penalty; the bottleneck shifts to line detection and line recognition.
  • The four-fold speedup makes line-level OCR attractive for edge devices and high-throughput document processing.
  • Line-level context improves recognition of punctuation and ambiguous characters that are hard to read in isolation.
  • The new 251-page line-annotated benchmark enables standardized end-to-end page-level evaluation for future OCR systems.
  • The line-level output is a natural input for language-model-based post-correction, potentially yielding further gains as language models improve.

Reading between the lines

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

  • The claimed 5.4% gain is contingent on judging OCR quality by an order-independent metric; if downstream tasks require reading order, the paper's own data suggests the advantage shrinks or reverses, since the line pipeline's CRR drops from 96.27 with oracle ordering to 85.76 with default ordering.
  • Line-level recognition may transfer to non-document images only if line detection is robust; Table 4 shows detection choice swings CRR from 24.30 to 85.76, so line segmentation, not recognition, may be the true remaining bottleneck.
  • A testable extension is to evaluate line-fine-tuned PARSeq on public scene-text datasets with line-level annotations to see whether the gain generalizes beyond clean document pages.
  • Coupling line-level OCR with a reading-order model (the paper leaves this to future work) could close the CRR gap and make the accuracy claim hold under stricter, order-sensitive evaluation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper argues that the main remaining bottleneck in document OCR is word segmentation, and proposes replacing the conventional line detection -> word detection -> word recognition pipeline with line detection followed directly by line-level recognition. The implementation uses Kraken for line detection and a PARSeq model fine-tuned on 6M synthetic line images generated with TRDG. The authors also introduce a new benchmark of 251 English page images with line-level annotations. On this benchmark they report a Flexible Character Accuracy (FCA) of 97.62 for the proposed Kraken+PARSeq_line pipeline versus 92.15 for the best word-based baseline (CRAFT+PARSeq), a character recognition rate (CRR) of 85.76 versus 83.46, and an inference time of 0.53s/page versus 2.11s/page, leading to claims of a 5.4% accuracy improvement and a 4x speedup.

Significance. If the central claim is correct, line-level recognition is a simple and effective way to sidestep word-detection errors and to exploit sentence context in document OCR, and the paper would make a useful contribution both methodologically and through the release of a public page-level benchmark. The paper also provides a clean experimental setup in which all word recognizers are retrained with a fixed charset, and the comparison covers a wide range of detection-recognition pairs. These are real strengths. However, the main quantitative claim is currently supported by an order-independent metric while the task is end-to-end OCR, and the comparison is not controlled for training-data domain. Both issues are load-bearing and need to be addressed before the 5.4% improvement claim can be accepted.

major comments (4)
  1. [Sec. 5.1, Table 1] The headline comparison is confounded by training-data distribution. The word-level recognizers are trained on SynthText and MJSynth (scene-text word crops, Sec. 5.1 and 5.3), whereas the line-level PARSeq is trained on TRDG-generated line images with 3,309 fonts, document-style distortions, and a Wikipedia-derived line corpus (Sec. A.1). The test set consists of 251 document page images. It is therefore possible that most of the 5.47-point FCA gap (97.62 vs 92.15) reflects a domain match between the line-model training distribution and the document test set, not the benefit of bypassing word segmentation. A word-level control trained on TRDG word crops (or a line-level model trained on scene-text-like data) is needed to support the paradigm claim.
  2. [Sec. 5.4, Sec. 6.2, Table 3] The paper's central accuracy claim is based on FCA, which is explicitly order-independent. For end-to-end OCR, reading order matters. The paper's own Table 3 shows that CRR for the proposed pipeline is 85.76 with default ordering and 96.27 with ground-truth ordering, while FCA remains 97.62 in all cases. This demonstrates that FCA masks a substantial line-ordering deficiency. On CRR, the improvement over the best word pipeline is only 2.3 points (85.76 vs 83.46), and Tesseract's CRR is higher (88.16, Table 2). The 5.4% improvement claim is therefore not supported for order-sensitive end-to-end OCR. The authors should report order-aware results, improve line ordering, or explicitly limit the claim to order-independent character accuracy.
  3. [Sec. 1.5, Table 1] The 4x efficiency claim is measured per page excluding Kraken's time, even though Kraken is used in both pipelines. Since Kraken is common to the proposed pipeline and the word-based baselines, the true end-to-end speedup is (2.11 + t_K)/(0.53 + t_K), which is strictly less than 4 and can approach 1 if the line-detection time is large. The authors should report end-to-end inference times including line detection, or at least state the speedup as a recognition-stage-only figure.
  4. [Sec. 1.5, Sec. 5.3] The phrase 'current state of the art of 92.15% for CRAFT based word detection, and PARSeq for word recognition' refers to the authors' own retrained models, not to a published SOTA result. The word-recognition models are retrained on synthetic datasets with a fixed 95-character charset, which is reasonable for a controlled comparison, but the paper should clearly distinguish 'state of the art' from 'best in our controlled set'. Without this clarification, the claim that line-level OCR improves over SOTA is overstated.
minor comments (5)
  1. [Sec. 3 / Abstract] The claim that no public dataset exists for line-level or page-level OCR benchmarking is too broad; there are public document-level datasets (e.g., historical newspaper and born-digital document benchmarks) with line or region annotations. Please qualify the claim to 'English page images with line-level annotations for benchmarking end-to-end word-level vs line-level OCR'.
  2. [Algorithm 1] The similarity function D and threshold tau are not precisely defined. Also, the greedy matching procedure can insert a Kraken line at an arbitrary position without enforcing global reading order; a short explanation of how the output order is recovered after Algorithm 1 would help.
  3. [Table 3] The abbreviation 'B.O.' is defined only in the caption, and the phrase 'blind ordering by line centroid' is ambiguous: does it mean ordering by the y-coordinate first and then x-coordinate? Please state the ordering rule explicitly.
  4. [Figure 1] There is a typo in the figure: 'Piepeline' should be 'Pipeline'.
  5. [Eq. (1)] In the product formula, the conditioning variables are not fully spelled out in every factor. If the notation is meant to show P(t_{i,j} | Z_i, T_{(i,<j)}, pi(T_i)), please write it consistently in the equation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the line-level OCR claim is an empirical comparison using external components; the main risks are metric-selection and training-distribution confounds, not circularity.

full rationale

The proposed pipeline is assembled from externally published components: Kraken line detection [19], PARSeq recognition [3], TRDG synthetic line generation [4], and Clausner et al.'s Flexible Character Accuracy [9]. The 5.4% accuracy improvement is an empirical end-to-end comparison on a new 251-page dataset, not a quantity derived from the model's own definitions. No parameter is fitted to the test set and no prediction is a renamed fit. The only self-citation in the paper, ref [11] co-authored by A. Garai, is used as background on multi-stage line segmentation and is not load-bearing for the central claim. The paper's choice of FCA over CRR is a metric-selection issue: FCA deliberately ignores line reading order, and Table 3 shows CRR swings from 85.76 to 96.27 depending on ordering, so the headline metric may overstate end-to-end quality in the reading-order sense. However, FCA is an externally published, order-independent character accuracy measure; it is not defined in terms of the proposed model, so this is not circularity. The skeptical concern that word-level baselines were trained on scene-text crops (SynthText/MJSynth) while the line-level model was trained on document-style TRDG data is a training-distribution confound affecting the validity of the attribution of the gain, not a circularity. There is no load-bearing self-citation chain, no ansatz smuggled in via citation, no renaming of a known result, and no fitted input called a prediction. The paper is therefore not circular, though its headline accuracy claim carries separate validity risks that should be evaluated under correctness/robustness rather than circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central result is an empirical comparison, not a derivation. The model weights are learned from synthetic data and not enumerated here. Hand-set choices (input size, charset) and background assumptions about the detection model and synthetic-to-real transfer are listed above. No new physical or mathematical entities are introduced.

free parameters (3)
  • Line-level recognition input size = 32x400 pixels
    Chosen to represent full text lines; hand-selected, not fit to the benchmark test set.
  • Recognition charset = 95 characters plus space for line model
    Word models use a 95-char set without space; the line model adds a space. This difference is necessary but affects comparability.
  • Similarity threshold tau in Algorithm 1 = 90%
    Used only in ablation experiments to reorder lines; not part of the deployed pipeline.
assumptions (3)
  • domain assumption Kraken provides consistent and sufficiently accurate line segmentation for both compared pipelines
    Section 4.1 and 5 use Kraken as the fixed line detector; no fine-tuning or robustness study on the 251 pages is reported.
  • domain assumption Models trained only on synthetic images generalize to real document images
    Section 5 states training only on synthetic data is inline with recent SOTA, but no evidence on distribution shift is given.
  • domain assumption Flexible Character Accuracy is a valid measure of end-to-end OCR accuracy
    Section 5.4 adopts FCA from Clausner et al. for headline results; the paper's own CRR analysis shows order-dependent accuracy differs substantially.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Stop at Words? Unveiling the Bigger Picture through Line-Level OCR." pith.science (2026). https://pith.science/paper/7I3VEE42

@misc{pith2026250821693,
  author       = {Pith},
  title        = {Pith review of: Why Stop at Words? Unveiling the Bigger Picture through Line-Level OCR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7I3VEE42}},
  note         = {Machine review of arXiv:2508.21693}
}
read the original abstract

Conventional optical character recognition (OCR) techniques segmented each character and then recognized. This made them prone to error in character segmentation, and devoid of context to exploit language models. Advances in sequence to sequence translation in last decade led to modern techniques first detecting words and then inputting one word at a time to a model to directly output full words as sequence of characters. This allowed better utilization of language models and bypass error-prone character segmentation step. We observe that the above transition in style has moved the bottleneck in accuracy to word segmentation. Hence, in this paper, we propose a natural and logical progression from word level OCR to line-level OCR. The proposal allows to bypass errors in word detection, and provides larger sentence context for better utilization of language models. We show that the proposed technique not only improves the accuracy but also efficiency of OCR. Despite our thorough literature survey, we did not find any public dataset to train and benchmark such shift from word to line-level OCR. Hence, we also contribute a meticulously curated dataset of 251 English page images with line-level annotations. Our experimentation revealed a notable end-to-end accuracy improvement of 5.4%, underscoring the potential benefits of transitioning towards line-level OCR, especially for document images. We also report a 4 times improvement in efficiency compared to word-based pipelines. With continuous improvements in large language models, our methodology also holds potential to exploit such advances. Project Website: https://nishitanand.github.io/line-level-ocr-website

Figures

Figures reproduced from arXiv: 2508.21693 by the authors.

Figure 1
Figure 1. Comparison between existing and proposed pipelines for OCR. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Sample images from the proposed dataset. Additional images are included in the supplementary material. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Left: Character frequency distribution, and Right: Words per Line distribution in our proposed dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Sample synthetic line-level images generated using TRDG [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Challenging examples for end-to-end word pipelines. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Sample failure cases of word detection models. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Additional sample images from the proposed dataset [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. 2019. Character Region Awareness for Text Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2019

  2. [2]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural Ma- chine Translation by Jointly Learning to Align and Translate. arXiv preprint arXiv:1409.0473 (2014)

    arXiv 2014

  3. [3]

    Darwin Bautista and Rowel Atienza. 2022. Scene Text Recognition with Permuted Autoregressive Sequence Models. In European Conference on Computer Vision . Springer-Verlag, Cham, Switzerland, 178–196

    work page 2022

  4. [4]

    Edouard Belval. 2018. TextRecognitionDataGenerator. GitHub. https://github. com/Belval/TextRecognitionDataGenerator

    work page 2018

  5. [5]

    Maurits Bleeker and Maarten de Rijke. 2019. Bidirectional Scene Text Recognition with a Single Decoder. arXiv preprint arXiv:1912.03656 (2019)

    work page Pith review arXiv 2019

  6. [6]

    Mélodie Boillet, Christopher Kermorvant, and Thierry Paquet. 2021. Multiple Document Datasets Pre-Training Improves Text Line Detection with Deep Neural Networks. In 2020 25th International Conference on Pattern Recognition (ICPR) . IEEE, Los Alamitos, CA, 2134–2141

    work page 2021

  7. [7]

    Qingwen Bu, Sungrae Park, Minsoo Khang, and Yichuan Cheng. 2023. SRFormer: Text Detection Transformer with Incorporated Segmentation and Regression. arXiv:2308.10531 [cs.CV]

    work page Pith review arXiv 2023

  8. [8]

    Huiyang Chen, Jing Liu, and Weimin Zhou. 2022. Natural Scene Text Detection Algorithm Based on Improved DBNet. In 2022 IEEE 5th International Conference on Electronic Information and Communication Technology (ICEICT) . IEEE, Los Alamitos, CA, 186–190. https://doi.org/10.1109/ICEICT55736.2022.9909509

    work page arXiv 2022

Show all 46 references
  1. [9]

    Christian Clausner, Stefan Pletschacher, and Apostolos Antonacopoulos. 2020. Flexible Character Accuracy Measure for Reading-Order-Independent Evaluation. Pattern Recognition Letters 131 (2020), 390–397

    2020

  2. [10]

    Senka Drobac and Krister Lindén. 2020. Optical Character Recognition with Neural Networks and Post-Correction with Finite State Methods. International Journal on Document Analysis and Recognition (IJDAR) 23, 4 (2020), 279–295

    2020

  3. [11]

    Arpita Dutta, Arpan Garai, Samit Biswas, and Amit Kumar Das. 2021. Segmenta- tion of Text Lines Using Multi-Scale CNN from Warped Printed and Handwritten Document Images. International Journal on Document Analysis and Recognition (IJDAR) 24, 4 (2021), 299–313

    2021

  4. [12]

    Shancheng Fang, Hongtao Xie, Yuxin Wang, Zhendong Mao, and Yongdong Zhang. 2021. Read Like Humans: Autonomous, Bidirectional and Iterative Lan- guage Modeling for Scene Text Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 7098–7107

    2021

  5. [13]

    Tongkun Guan, Chaochen Gu, Jingzheng Tu, Xue Yang, Qi Feng, Yudi Zhao, and Wei Shen. 2023. Self-Supervised Implicit Glyph Attention for Text Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15285–15294

    2023

  6. [14]

    Tongkun Guan, Wei Shen, Xue Yang, Qi Feng, Zekun Jiang, and Xiaokang Yang

  7. [15]

    Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. 2016. Synthetic Data for Text Localisation in Natural Images. In IEEE Conference on Computer Vision and Pattern Recognition

    2016

  8. [16]

    Max Jaderberg, Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. 2016. Reading Text in the Wild with Convolutional Neural Networks. International Journal of Computer Vision 116, 1 (Jan. 2016), 1–20

    2016

  9. [17]

    Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. 2023. Revisiting Scene Text Recognition: A Data Perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) . 20543–20554

    2023

  10. [18]

    Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, N. C. Gokul, Avik Bhat- tacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. IndicNLPSuite: Mono- lingual Corpora, Evaluation Benchmarks and Pre-Trained Multilingual Language Models for Indian Languages. In Findings of the...

    2020 doi

  11. [19]

    Benjamin Kiessling. 2020. A Modular Region and Text Line Layout Analysis Sys- tem. In 2020 17th International Conference on Frontiers in Handwriting Recognition (ICFHR). IEEE, Los Alamitos, CA, 313–318

    2020

  12. [20]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient- Based Learning Applied to Document Recognition. Proc. IEEE 86, 11 (1998), 2278–2324

    1998

  13. [21]

    Junyeop Lee, Sungrae Park, Jeonghun Baek, Seong Joon Oh, Seonghyeon Kim, and Hwalsuk Lee. 2020. On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 546–547

    2020

  14. [22]

    Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma. 2022. PP-OCRv3: More Attempts for the Improvement of Ultra Lightweight OCR System. arXiv:2206.03001 [cs.CV]

    2022 arXiv

  15. [23]

    Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. 2020. Real-Time Scene Text Detection with Differentiable Binarization. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 11474–11481

    2020

  16. [24]

    Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. 2022. Real- Time Scene Text Detection with Differentiable Binarization and Adaptive Scale Fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 1 (2022), 919–931

    2022

  17. [25]

    Wong, Zhizhong Su, and Junyu Han

    Wei Liu, Chaofeng Chen, Kwan-Yee K. Wong, Zhizhong Su, and Junyu Han. 2016. Star-Net: A Spatial Attention Residue Network for Scene Text Recognition. In BMVC, Vol. 2. 7

    2016

  18. [26]

    Elman Mansimov, Alex Wang, Sean Welleck, and Kyunghyun Cho. 2019. A Generalized Framework of Sequence Generation with Application to Undirected Sequence Models. arXiv preprint arXiv:1905.12790 (2019)

    2019 arXiv

  19. [27]

    Stephen Merity et al. 2016. Wikitext-103. https://arxiv.org/abs/1609.07843

    2016 arXiv

  20. [28]

    Mindee. 2021. docTR: Document Text Recognition. https://github.com/mindee/doctr

    2021

  21. [29]

    Bastien Moysset, Christopher Kermorvant, Christian Wolf, and Jérôme Louradour

  22. [30]

    Byeonghu Na, Yoonsik Kim, and Sungrae Park. 2022. Multi-Modal Text Recogni- tion Networks: Interactive Enhancements Between Visual and Semantic Features. In European Conference on Computer Vision . Springer-Verlag, Cham, Switzerland, 446–463

    2022

  23. [31]

    Yash Patel and Jiří Matas. 2021. Feds: Filtered Edit Distance Surrogate. In Interna- tional Conference on Document Analysis and Recognition . Springer-Verlag, Cham, Switzerland, 171–186

    2021

  24. [32]

    Weizhen Qi, Yeyun Gong, Jian Jiao, Yu Yan, Weizhu Chen, Dayiheng Liu, Kewen Tang, Houqiang Li, Jiusheng Chen, Ruofei Zhang, et al . 2021. Bang: Bridging Autoregressive and Non-Autoregressive Generation with Large Scale Pretrain- ing. In International Conference on Machine Lear...

    2021

  25. [33]

    Fenfen Sheng, Zhineng Chen, and Bo Xu. 2019. NRTR: A No-Recurrence Sequence- to-Sequence Model for Scene Text Recognition. In 2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, Los Alamitos, CA, 781–786

    2019

  26. [34]

    Baoguang Shi, Xiang Bai, and Cong Yao. 2017. An End-to-End Trainable Neural Network for Image-Based Sequence Recognition and Its Application to Scene Text Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 11 (2017), 2298–2304. https://doi.org/10.1...

    2017

  27. [35]

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

    2007

  28. [36]

    Seth Stewart and Bill Barrett. 2017. Document Image Page Segmentation and Character Recognition as Semantic Segmentation. In Proceedings of the 4th In- ternational Workshop on Historical Document Imaging and Processing (HIP ’17) . ACM, New York, NY, USA, 101–106

    2017

  29. [37]

    Chao Tian, Yifei Wang, Hao Cheng, Yijiang Lian, and Zhihua Zhang. 2020. Train Once, and Decode as You Like. In Proceedings of the 28th International Conference on Computational Linguistics. 280–293

    2020

  30. [38]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. Advances in Neural Information Processing Systems 30 (2017)

    2017

  31. [39]

    Peng Wang, Cheng Da, and Cong Yao. 2022. Multi-Granularity Prediction for Scene Text Recognition. In European Conference on Computer Vision . Springer- Verlag, Cham, Switzerland, 339–355

    2022

  32. [40]

    Salakhutdinov, and Quoc V

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R. Salakhutdinov, and Quoc V. Le. 2019. XLNet: Generalized Autoregressive Pretraining for Lan- guage Understanding. Advances in Neural Information Processing Systems 32 (2019)

    2019

  33. [41]

    Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. 2020. TextFuseNet: Scene Text Detection with Richer Fused Features. In IJCAI, Vol. 20. 516–522

    2020

  34. [42]

    Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao

  35. [43]

    Yu-Xiang Zeng, Jun-Wei Hsieh, Xin Li, and Ming-Ching Chang. 2023. MixNet: Toward Accurate Detection of Challenging Scene Text in the Wild. arXiv:2308.12817 [cs.CV] 9 Line Level OCR, August, 2025 Vempati and Anand A SUPPLEMENTARY MATERIAL A.1 Deep Dive into TRDG Data Generation...

    2023 arXiv

  36. [45]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    DPText-DETR: Towards Better Scene Text Detection with Dynamic Points in Transformer. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 3241–3249

  37. [2015]

    In 2015 13th International Conference on Document Analysis and Recognition (ICDAR)

    Paragraph Text Segmentation into Lines with Recurrent Neural Networks. In 2015 13th International Conference on Document Analysis and Recognition (ICDAR). IEEE, Los Alamitos, CA, 456–460. https://doi.org/10.1109/ICDAR.2015. 7333803

    2015 doi

  38. [2023]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Self-Supervised Character-to-Character Distillation for Text Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 19473– 19484

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.