Billet Number Recognition Based on Test-Time Adaptation

Xiuzhuang Zhou; Yuan Wei

REVIEW 5 major objections 6 minor 25 references

Billet Number Recognition Based on Test-Time Adaptation

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding test-time adaptation and encoding-rule priors lifts billet number recognition from 58% to 80% on machine-printed steel billets and from 24% to 70% on handwritten ones.

desk verdict A practical OCR/TTA combo for steel billet numbers with plausible gains, but the blank-restoration rule is under-validated and the citation list has fixable errors. read the letter →

arxiv 2502.09026 v1 pith:ATLKQZ5I submitted 2025-02-13 cs.CV

Yuan Wei , Xiuzhuang Zhou This is my paper · ORCID

classification cs.CV

keywords steelbilletnumberrecognitiontest-timeadaptationentropyminimizationscenetextCTCblanktokenrepairpriorknowledgepost-correctionDBdetectionSVTR

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 aims to make automated reading of ID numbers stamped or written on steel billets accurate enough for real factory use. It argues that off-the-shelf scene text recognition fails because factory images differ from training data and because damaged characters are dropped by CTC post-processing. The proposed fix combines test-time adaptation, which adjusts a few model parameters on unlabeled test images by minimizing prediction entropy, with prior knowledge of billet number encoding rules, which rejects and replaces implausible recognitions. On 500 machine-printed and 500 handwritten test images, the combined method raises accuracy from 0.5841 to 0.7990 and from 0.2371 to 0.7043, respectively.

What carries the argument

The argument rests on four mechanisms. The DB (differentiable binarization) network detects text regions by learning a pixel-wise threshold, producing detection boxes around the billet numbers. The SVTR network recognizes the cropped text as a character sequence, outputting per-position class probabilities, and its batch-normalization layers carry the adaptation. The test-time adaptation module minimizes the entropy of the SVTR predictions and back-propagates only into the batch-normalization affine parameters ($\gamma$, $\beta$), keeping the rest of the model fixed. Finally, a prior-knowledge post-processor applies the billet-number encoding rules to the model's top-ranked candidates, choosing the highest-probability candidate that obeys the rules, and a CTC blank-restoration rule replaces runs of three or more consecutive blank tokens with the most probable non-blank character, on the assumption that such a run marks a damaged rather than a misaligned character.

What would settle it

Build a test set of billet images with known per-character box alignments, and compare recognition with and without the blank-restoration rule on intact characters whose detection boxes have been deliberately shifted by a few pixels; if the rule inserts characters in those misaligned cases or lowers accuracy relative to plain blank removal, the alignment assumption is wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that two lightweight post-hoc mechanisms can make a pretrained scene text recognizer usable for steel billet numbers without retraining on large labeled target sets. First, test-time adaptation minimizes the model's prediction entropy on each arriving batch, updating only the scale and shift parameters of the batch-normalization layers, so the recognizer shifts toward the test distribution. Second, a prior-knowledge layer uses the billet-number encoding rules to reject recognition outputs that violate the allowed letter/digit structure, re-scoring the model's candidate dictionary, and a separate CTC repair step treats a run of three or more consecutive blank tokens as evidence of a damaged character and restores the most probable non-blank character. On real factory images, the combination raises accuracy from 0.5841 to 0.7990 on machine-printed numbers and, after fine-tuning on printed data, from 0.2371 to 0.7043 on handwritten numbers.

Load-bearing premise

The CTC repair rule assumes that at the right image scale almost every detection box lines up with one character, so a run of three or more blank tokens must mean a damaged character rather than a misaligned detection box; if the boxes are often shifted, the rule will insert characters in the wrong places.

Editorial extensions

If this is right

If the full method is used, factory OCR on machine-printed billet numbers reaches 79.90% accuracy, a level materially closer to usable automated tracking than the 58.41% baseline.
Because test-time adaptation needs no labels and no original training data, the same pretrained recognizer can be pointed at a new mill's imaging conditions and improve as batches of arriving billets stream through.
Encoding-rule checks turn an open-vocabulary recognition problem into a constrained one, so any structured identifier with position-specific letter/digit rules can be post-corrected in the same way.
The CTC blank-run repair recovers characters lost to paint peeling or oxidation, a failure mode that retraining the recognizer alone would not fix.

Reading between the lines

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

Beyond the paper, the three-blank repair rule should transfer to other CTC-based recognizers for fixed-length, tightly spaced codes, but only after verifying the character-alignment assumption per imaging device; a different camera geometry could break it.
The paper's conclusion flags noise overfitting as a risk for test-time adaptation, so a natural stress test would be a confidence gate or rolling reset that prevents a single noisy batch from corrupting the adapted model.
The prior-knowledge gain is bounded by how much of the billet number is actually rule-constrained; the reported gains (about 10 points on printed, 7 points on handwritten, after fine-tuning) suggest one could quantify how much benefit remains if the encoding rules are partially unknown.

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.

Desk Editor's Note

read the letter

Short version: this is a practical OCR paper that staples a known test-time adaptation method (TENT) onto a standard detection/recognition pipeline and adds two prior-knowledge post-processing tricks for steel billet numbers. The combination appears to work on real data, but the paper under-validates the one genuinely unusual component.

What's new: the specific system — DB + SVTR + TENT-style entropy minimization + encoding-rule-based correction + a CTC blank-run restoration heuristic — is not something I've seen in prior work. The industrial setting is real, and the reported gains (0.58→0.80 on machine-printed, 0.24→0.70 on handwritten after fine-tuning) are the kind of numbers a factory would care about. The authors don't oversell a new theory; they present it as an applied engineering contribution, which is fair.

What's solid: the experiments include a fine-tuning baseline, a TTA-only row, and a combined row, giving some sense of each block's contribution. The TTA implementation follows TENT in a straightforward way. The encoding-rule correction is sensible domain knowledge. The conclusion also honestly notes the computational cost and the risk of overfitting to noisy test data.

Where it gets soft: the blank-restoration rule in Sec. 3.3.2 is the most novel piece and the least supported. The paper asserts that at a suitable image size, most detection boxes align, so 3+ consecutive CTC blanks imply a damaged character rather than misalignment. No alignment statistics, blank-run length distributions, or sensitivity analysis on the threshold of 3. And the row labeled 'Inference with Prior Knowledge Constraints' (Table 4, 0.5841→0.6852) bundles the encoding-rule replacement and the blank-run restoration together, so you can't tell what the blank-run rule contributes. There are also no error bars or significance tests, no comparison to other TTA methods or earlier billet-specific recognizers, and no released code/data. The citation pattern has a couple of clear mismatches: Zhao et al. [13] in the text actually points to Long et al. in the references, and Wang et al. [18] 'self-consistency' points to a non-local networks paper. These are fixable but sloppy.

Is the central claim circular? No. The encoding rules are external, the test sets are held out, and the thresholds are empirical. The risk is not circularity; it's incomplete validation of the blank-run heuristic.

Bottom line: for a reader working on industrial OCR or test-time adaptation in niche domains, this paper is worth a look. It deserves a real peer review, but I'd push for major revisions: separate ablation of the two prior-knowledge mechanisms, threshold sensitivity, alignment statistics, error bars, and a cleaned reference list. I wouldn't cite it yet until the artifacts or at least the ablations are available.

Referee Report

5 major / 6 minor

Summary. The paper proposes a billet number recognition pipeline for steel production that combines a DB text detector with an SVTR recognizer, a test-time adaptation (TTA) module based on entropy minimization of batch-normalization parameters, and prior-knowledge post-processing. The post-processing has two parts: automatic correction of recognition results using billet encoding rules, and a CTC blank-token restoration mechanism that replaces runs of three or more consecutive blank tokens with the highest-probability non-blank character. Experiments on 500 machine-printed and 500 handwritten billet-number images report accuracy improvements from 0.5841 to 0.7990 and from 0.2371 to 0.7043, respectively, along with corresponding reductions in edit distance.

Significance. If the reported results are reproducible, the proposed combination is practically valuable for industrial billet tracking, since it reduces recognition errors substantially on real, held-out test data without requiring labeled target data. The paper has several strengths: it evaluates on real collected images, reports both image-level accuracy and edit distance, uses held-out test sets, and describes the baseline configuration in enough detail to be reimplemented from PaddleOCR. The evaluation is not circular: the encoding rules are external domain knowledge, and the thresholds are not fitted to the evaluation metric. However, the significance of the central claim is currently limited by missing ablations and missing baselines. In particular, the novel blank-restoration mechanism is bundled with the encoding-rule correction in the reported 'Inference with Prior Knowledge Constraints' rows, so the contribution of the paper's most distinctive component is not established. No comparison to existing billet recognition methods or to other TTA methods is provided, and no error bars or significance tests are reported.

major comments (5)

[Sec. 3.3.2, Tables 4 and 5] The claim that the CTC blank-restoration mechanism improves accuracy is not supported by an ablation. The rows labeled 'Inference with Prior Knowledge Constraints' combine the encoding-rule replacement of Sec. 3.3.1 with the blank-token restoration of Sec. 3.3.2, so the gains from 0.5841 to 0.6852 (Table 4) and from 0.5634 to 0.6356 (Table 5) cannot be attributed to the blank-restoration rule. Please provide separate ablations for encoding-rule correction only, blank restoration only, and the combination, together with a sensitivity analysis for the threshold of three consecutive blank tokens and a statement of how the number of CTC time steps relates to image width.
[Sec. 4.1 and Tables 4-5] It is not clear which checkpoint is used in each experimental row. Section 4.1 states that the SVTR network is fine-tuned on 4,200 machine-printed images, but Table 4 labels the baseline as 'DB + SVTR' and the text describes it as an 'unrefined SVTR network.' Table 5 includes an explicit 'DB + SVTR + Fine-tuning' row for handwritten numbers. The authors should state for every row whether the starting model is the pretrained PaddleOCR checkpoint or the fine-tuned checkpoint; without this, the TTA and prior-knowledge gains cannot be interpreted.
[Sec. 3.3.1] The encoding rules are described only by example ('the first character ... must be a letter'). A complete, reproducible specification is needed: the exact grammar of a billet number, position-wise allowed character sets, field lengths, and any date, furnace, or checksum constraints, together with the fallback behavior if no candidate in the top-k satisfies the rule at a given position. Without this, the prior-knowledge correction cannot be reimplemented or compared.
[Sec. 4.3] The experiments report single-point accuracy on 500 test images per condition and no variance or significance testing. For example, the difference between 0.7637 and 0.7990 in Table 4 corresponds to roughly 18 images, and the word 'significant' in the abstract and conclusion is not supported by any statistical test. Please add confidence intervals, repeated-run variability, or at least a permutation or bootstrap test on the matched test set, and report per-file error counts.
[Sec. 4.3] No comparison is made to existing billet number recognition methods (e.g., Refs. [12], [14], [15]) or to other test-time adaptation methods. Because Sec. 3.2 says the TTA component 'adopt[s] the approach proposed by Tent et al.,' a direct TENT baseline is needed to confirm that the reported TTA gain is not simply the known TENT behavior on this data. A comparison with a simple confidence-threshold or dictionary-lookup post-processor would also help isolate the value of the prior-knowledge mechanism.

minor comments (6)

[Tables 4 and 5] The header 'Eidt distance' is a typo and should read 'Edit distance'.
[Sec. 2.1 and References] Reference [13] is cited as 'Zhao et al.' but the listed reference is Long et al.'s scene text detection and recognition review; similarly, reference [18] is cited for 'self-consistency' but the listed reference is the Non-local neural networks paper. The citations should be corrected to the actual works.
[Sec. 3.2, Figure 3] Figure 3 is invoked to support the entropy-error-rate relationship, but no axes, dataset, or quantitative protocol are described. Please specify how the figure was generated.
[Sec. 3.3.1] The correction procedure stops when 'a valid result is identified,' but the paper does not state what happens if none of the candidate characters at a position satisfies the encoding rule; this case should be defined explicitly.
[Sec. 3.3.2] The notation for the CTC blank token is inconsistent (empty string, 'Blank', 'blank'); using a single symbol such as '<blank>' throughout would improve readability.
[Table 1] Table 1 lists methods and their losses but does not give the actual loss expressions for test-time training or adaptation; consider either presenting the explicit objectives or removing the table, since the text already describes the entropy objective.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out test sets, external encoding rules, and no self-citation chain.

full rationale

No significant circularity found. The paper's contributions are an engineering pipeline (DB detection, SVTR recognition, TENT-style test-time adaptation, encoding-rule post-processing, and a CTC blank-restoration heuristic) evaluated on held-out test sets of 500 machine-printed and 500 handwritten billet images. The encoding rules (Section 3.3.1) are external domain knowledge, not quantities fitted to the test set; the TTA objective is unsupervised entropy minimization over test inputs without labels; and the accuracy/edit-distance metrics are computed against ground-truth labels never used during adaptation. The 3-consecutive-blank heuristic (Section 3.3.2) rests on an empirical alignment assumption, and the paper lacks an ablation isolating it, but that is a correctness/validation risk, not circularity: the heuristic is not constructed from the evaluation metric and no result is equivalent to its input by definition. There are no self-citations by the authors (the cited works are PaddleOCR, DB, SVTR, and prior TTA methods), so no self-citation chain is load-bearing. Score 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method is a composition of known components and external domain knowledge. The main hand-tuned quantity is the blank-run threshold, and the encoding rules are taken from the factory rather than derived. No new physical or algorithmic entities are introduced.

free parameters (1)

min_consecutive_blanks_for_damaged_character = 3
Chosen heuristically in Section 3.3.2; no sensitivity analysis or validation reported.

assumptions (4)

domain assumption Billet numbers follow encoding rules where each part is exclusively letters or digits and the first character is a letter.
Used for prior-knowledge correction in Section 3.3.1; if rules are incomplete or different, corrections may force errors.
domain assumption Setting an appropriate image size ensures most detection boxes align with target characters.
Required for the three-blank heuristic in Section 3.3.2.
domain assumption Lower prediction entropy indicates lower error rate (Figure 3).
Basis for TTA objective in Section 3.2; no detailed derivation, only an empirical plot.
domain assumption Standard TENT entropy minimization on BN parameters transfers to SVTR without harming recognition.
Section 3.2; relies on TENT results but performance on this architecture is not independently shown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Billet Number Recognition Based on Test-Time Adaptation." pith.science (2026). https://pith.science/paper/ATLKQZ5I

@misc{pith2026250209026,
  author       = {Pith},
  title        = {Pith review of: Billet Number Recognition Based on Test-Time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ATLKQZ5I}},
  note         = {Machine review of arXiv:2502.09026}
}

read the original abstract

During the steel billet production process, it is essential to recognize machine-printed or manually written billet numbers on moving billets in real-time. To address the issue of low recognition accuracy for existing scene text recognition methods, caused by factors such as image distortions and distribution differences between training and test data, we propose a billet number recognition method that integrates test-time adaptation with prior knowledge. First, we introduce a test-time adaptation method into a model that uses the DB network for text detection and the SVTR network for text recognition. By minimizing the model's entropy during the testing phase, the model can adapt to the distribution of test data without the need for supervised fine-tuning. Second, we leverage the billet number encoding rules as prior knowledge to assess the validity of each recognition result. Invalid results, which do not comply with the encoding rules, are replaced. Finally, we introduce a validation mechanism into the CTC algorithm using prior knowledge to address its limitations in recognizing damaged characters. Experimental results on real datasets, including both machine-printed billet numbers and handwritten billet numbers, show significant improvements in evaluation metrics, validating the effectiveness of the proposed method.

Figures

Figures reproduced from arXiv: 2502.09026 by the authors.

**Figure 1.** Overview of our method. 3.1. Text Detection and Recognition Networks The steel billet number recognition task can be regarded as a specific case of scene text recognition, which is typically divided into two stages: text detection and text recognition. In the text detection phase, we utilize the DB (Differentiable Binarization) network [2], a segmentation-based text detection algorithm. The core of this network l… view at source ↗

**Figure 2.** SVTR Network Architecture [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The relationship between entropy and error rate. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Misrecognition of the CTC algorithm in the case of dam [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages

[13]

Scene text detec- tion and recognition: The deep learning era

Shangbang Long, Xin He, and Cong Yao. Scene text detec- tion and recognition: The deep learning era. International Journal of Computer Vision, 129(1):161–184, 2021. 2

work page 2021
[18]

Non-local neural networks

Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaim- ing He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 7794–7803, 2018. 2

work page 2018
[12]

Unified deep neural networks for end-to-end recog- nition of multi-oriented billet identification number

Gyogwon Koo, Jong Pil Yun, Hyeyeon Choi, and Sang Woo Kim. Unified deep neural networks for end-to-end recog- nition of multi-oriented billet identification number. Expert Systems with Applications, 168:114377, 2021. 2

work page 2021
[14]

Recogni- tion method for handwritten steel billet identification number based on yolo deep convolutional neural network

Qiaojie Sun, Dali Chen, Sen Wang, and Shixin Liu. Recogni- tion method for handwritten steel billet identification number based on yolo deep convolutional neural network. In 2020 Chinese Control And Decision Conference (CCDC) , pages 5642–5646. IEEE, 2020. 2

work page 2020
[15]

Billet number recognition based on esrgan and improved yolov5

Zijia Wang, Yichao Dong, Dan Niu, Minghao Liu, Qi Li, and Xisong Chen. Billet number recognition based on esrgan and improved yolov5. In2022 37th Youth Academic Annual Con- ference of Chinese Association of Automation (YAC) , pages 1384–1389. IEEE, 2022. 2

work page 2022
[1]

His- torical review of ocr research and development.Proceedings of the IEEE, 80(7):1029–1058, 1992

Shunji Mori, Ching Y Suen, and Kazuhiko Yamamoto. His- torical review of ocr research and development.Proceedings of the IEEE, 80(7):1029–1058, 1992. 1

work page 1992
[2]

Real-time scene text detection with differentiable bina- rization

Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable bina- rization. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11474–11481, 2020. 1, 3

work page 2020
[3]

Svtr: Scene text recognition with a single visual model

Yongkun Du, Zhineng Chen, Caiyan Jia, Xiaoting Yin, Tian- lun Zheng, Chenxia Li, Yuning Du, and Yu-Gang Jiang. Svtr: Scene text recognition with a single visual model. arXiv preprint arXiv:2205.00159, 2022. 1, 3

arXiv 2022

Show all 25 references

[4]

Odm: A text-image further alignment pre-training approach for scene text detection and spotting

Chen Duan, Pei Fu, Shan Guo, Qianyi Jiang, and Xiaom- ing Wei. Odm: A text-image further alignment pre-training approach for scene text detection and spotting. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15587–15597, 2024. 1

2024
[5]

Multi-modal in-context learning makes an ego-evolving scene text recognizer

Zhen Zhao, Jingqun Tang, Chunhui Lin, Binghong Wu, Can Huang, Hao Liu, Xin Tan, Zhizhong Zhang, and Yuan Xie. Multi-modal in-context learning makes an ego-evolving scene text recognizer. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pa...

2024
[6]

Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks

Alex Graves, Santiago Fern ´andez, Faustino Gomez, and J¨urgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international confer- ence on Machine learning, pages 369–376, 2006. 1

2006
[7]

Batch number of billets recognition sys- tem design

Xiaojun Zhang. Batch number of billets recognition sys- tem design. Master’s thesis, Shanghai Jiao Tong University, 2009, 2009. 2

2009
[8]

Study on steel plate character recognition in iron and steel logistics

Xiao Zhou. Study on steel plate character recognition in iron and steel logistics. Industrial Control Computer, (2):93–94,
[9]

Study of rolling mill production line of heavy rail steel billet recognition system

Bin Dong. Study of rolling mill production line of heavy rail steel billet recognition system. Master’s thesis, Wuhan Institute of Technology, 2015. 2

2015
[10]

In- spection method for steel billet characters based on svm.Mi- crocomputer Applications, (10):49–51, 2011

Di Wu, Dongsheng Jiao, Xiao Zhang, and Zhanguo Shi. In- spection method for steel billet characters based on svm.Mi- crocomputer Applications, (10):49–51, 2011. 2

2011
[11]

Content based image retrieval system using k-means and knn approach by fea- ture extraction

MJ Sadiq, A Prof, A Kaleem, et al. Content based image retrieval system using k-means and knn approach by fea- ture extraction. Journal of Peritherapeutic Neuroradiology, 24(6):643–649, 2018. 2

2018
[16]

Automatic recog- nition of hot spray marking dot-matrix characters for steel- slab industry

Junhui Ge, Licheng Liu, Junxi Sun, Hong Zhao, Langming Zhou, Tianle Cheng, and Changyan Xiao. Automatic recog- nition of hot spray marking dot-matrix characters for steel- slab industry. Journal of Intelligent Manufacturing , pages 1–16, 2023. 2

2023
[17]

Bdr-net: digital recog- nition network for billet surface based on flow alignment and attention mechanism

Jinyu Xu, Yipeng Liu, and Zhitai Liu. Bdr-net: digital recog- nition network for billet surface based on flow alignment and attention mechanism. Signal, Image and Video Processing, 18(11):7939–7952, 2024. 2

2024
[19]

Test-time training with self- supervision for generalization under distribution shifts

Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–
[20]

The norm must go on: Dynamic unsuper- vised domain adaptation by normalization

M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsuper- vised domain adaptation by normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14765–14775, 2022. 2

2022
[21]

Tent: Fully test-time adaptation by entropy minimization

Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Ol- shausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726,

2006 arXiv
[22]

Towards stable test-time adaptation in dynamic wild world

Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. arXiv preprint arXiv:2302.12400, 2023. 2

2023 arXiv
[23]

Pseudo-label: The simple and effi- cient semi-supervised learning method for deep neural net- works

Dong-Hyun Lee et al. Pseudo-label: The simple and effi- cient semi-supervised learning method for deep neural net- works. In Workshop on challenges in representation learn- ing, ICML, volume 3, page 896. Atlanta, 2013. 2

2013
[24]

Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation

Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation. In International conference on machine learning, pages 6028–6039. PMLR, 2020. 2

2020
[25]

Pp-ocrv3: More attempts for the im- provement of ultra lightweight ocr system

Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, et al. Pp-ocrv3: More attempts for the im- provement of ultra lightweight ocr system. arXiv preprint arXiv:2206.03001, 2022. 6

2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [13]

Scene text detec- tion and recognition: The deep learning era

Shangbang Long, Xin He, and Cong Yao. Scene text detec- tion and recognition: The deep learning era. International Journal of Computer Vision, 129(1):161–184, 2021. 2

work page 2021

[2] [18]

Non-local neural networks

Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaim- ing He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 7794–7803, 2018. 2

work page 2018

[3] [12]

Unified deep neural networks for end-to-end recog- nition of multi-oriented billet identification number

Gyogwon Koo, Jong Pil Yun, Hyeyeon Choi, and Sang Woo Kim. Unified deep neural networks for end-to-end recog- nition of multi-oriented billet identification number. Expert Systems with Applications, 168:114377, 2021. 2

work page 2021

[4] [14]

Recogni- tion method for handwritten steel billet identification number based on yolo deep convolutional neural network

Qiaojie Sun, Dali Chen, Sen Wang, and Shixin Liu. Recogni- tion method for handwritten steel billet identification number based on yolo deep convolutional neural network. In 2020 Chinese Control And Decision Conference (CCDC) , pages 5642–5646. IEEE, 2020. 2

work page 2020

[5] [15]

Billet number recognition based on esrgan and improved yolov5

Zijia Wang, Yichao Dong, Dan Niu, Minghao Liu, Qi Li, and Xisong Chen. Billet number recognition based on esrgan and improved yolov5. In2022 37th Youth Academic Annual Con- ference of Chinese Association of Automation (YAC) , pages 1384–1389. IEEE, 2022. 2

work page 2022

[6] [1]

His- torical review of ocr research and development.Proceedings of the IEEE, 80(7):1029–1058, 1992

Shunji Mori, Ching Y Suen, and Kazuhiko Yamamoto. His- torical review of ocr research and development.Proceedings of the IEEE, 80(7):1029–1058, 1992. 1

work page 1992

[7] [2]

Real-time scene text detection with differentiable bina- rization

Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable bina- rization. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11474–11481, 2020. 1, 3

work page 2020

[8] [3]

Svtr: Scene text recognition with a single visual model

Yongkun Du, Zhineng Chen, Caiyan Jia, Xiaoting Yin, Tian- lun Zheng, Chenxia Li, Yuning Du, and Yu-Gang Jiang. Svtr: Scene text recognition with a single visual model. arXiv preprint arXiv:2205.00159, 2022. 1, 3

arXiv 2022