pith. sign in

arxiv: 2606.08858 · v1 · pith:KHGVNMJHnew · submitted 2026-06-07 · 💻 cs.CV · cs.AI

Intelligent Character Recognition of Handwritten Forms with Deep Neural Networks

Hartwig Grabowski This is my paper

Pith reviewed 2026-06-27 18:27 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords handwritten formscharacter recognitiondeep neural networksEMNIST datasetsingle-task approachdetection and classificationartificial training data
0
0 comments X

The pith

A single deep neural network executes both detection and classification of handwritten characters on forms and reaches 88.28 percent accuracy on real exam data.

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

The paper sets out to show that detection and classification of handwritten Latin letters can be performed together inside one deep neural network rather than as two separate tasks. Training data is created artificially by overlaying characters from the EMNIST set onto the blank forms instead of relying on hand-labeled examples. This unified model is reported to outperform the conventional two-task pipeline and to deliver an overall recognition rate of 88.28 percent when applied to genuine handwritten exam sheets. A reader would care because the method removes the need for separate detection stages and for manual annotation of large training sets, which are common bottlenecks in form-processing systems.

Core claim

The authors demonstrate that a deep neural network trained to carry out both detection and classification in a single task, using training data manufactured by overlaying EMNIST letters onto the underlying forms, is superior to the state-of-the-art two-task approach and attains an overall recognition rate of 88.28 percent on real handwritten exam data.

What carries the argument

A unified deep neural network that integrates character detection and classification into one task, trained on artificially generated data.

If this is right

  • The single-task network outperforms the standard two-task method on the same forms.
  • An overall recognition rate of 88.28 percent is obtained on real handwritten exam data.
  • The approach is applied to handwritten Latin letters using the EMNIST dataset.
  • Limitations observed in the EMNIST dataset require further customization of the training data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The artificial-data technique could reduce the labor required to prepare training sets for other handwritten-document tasks.
  • If the unified network scales, it may allow end-to-end processing pipelines for forms without intermediate detection modules.
  • The same overlay method might be tested on non-Latin scripts once suitable base datasets become available.

Load-bearing premise

Artificially manufactured training data created by overlaying EMNIST letters onto the forms accurately captures the distribution and variability of real handwritten input without introducing systematic biases.

What would settle it

Running the trained model on a fresh collection of real handwritten exam forms from different writers and institutions and checking whether the recognition rate drops substantially below 88.28 percent.

Figures

Figures reproduced from arXiv: 2606.08858 by Hartwig Grabowski.

Figure 1
Figure 1. Figure 1: A table is printed on the lower right side of the paper. Each column represents one question. Each question must be answered by a capital Latin letter [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: The first 40 characters from the EMNIST Balanced Letterset. Each sample has a shape of 28x28 pixels. 3 Datasets The approach presented here uses ANN for the classification task. However, for training the ANN much training data is required and even specialized data aug￾mentation methods which use trained decoder networks to generate variations of the sample characters require 200 and more samples for each c… view at source ↗
Figure 4
Figure 4. Figure 4: Segmentation of the table is based on text marker detected with tesseract [41] [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: The architecture of the CNN: Every two convolution layers are followed by BatchNormalization and MaxPooling, Softmax is used for Output Layer. Detailed de￾scription in [38] [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The first 8 letters of the EMNIST dataset and their predicted character by the CNN. All predictions are correct, letter ‘f’ and ‘F’ are accumulated in one class. drawn dots and dashes this approach is error-prone. However, it turned out that this effect can easily be compensated with data augmentation by zooming out the images of the training set with a factor up to 3, which improved the accuracy up to 81.… view at source ↗
Figure 8
Figure 8. Figure 8: Eight letters cropped from the table and their predicted classes. First line shows the prediction without data augmentation, second line with data augmentation [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Misplaced letter: ‘I’ is written below the table. handles the placement of the letters above or below the table. A YOLOv5 model was used as target detection model and trained to detect the letters in, above or below the table. In order become independent from hardcoded cell positions and sizes, the model was trained to detect the printed digits above the letters, too ( [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 10
Figure 10. Figure 10: Segmentation of digits and the letters with the YOLOv5s model. Letter outside the table are detected, too (right) [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: 18 letters and one digit are projected into the table cells with random spatial deviation. The bonding boxes are calculated during the projection [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Segmentation and classification of the letters with the trained YOLOv5 model. file [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: First 20 samples of class ‘O’ (first line), of class ‘0’ (digit 0) (the second line), of class ‘I’ (third line) and of class ‘L’ (fourth line) from the EMNSIT Balanced Dataset [PITH_FULL_IMAGE:figures/full_fig_p010_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Crossed out letters are falsely classified (left). The letter ‘F’ comes in two shapes (middle, right), but only one shape (right) is part of EMNIST data set. 2. Letter ‘I’ and ‘L’: Frequently, letter ‘I’ was classified as ‘L’. In the “EM￾NIST Balanced Letter” dataset letter ‘l’ (lowercase ‘L’) was merged with the class of letter ‘L’ (uppercase ‘L’) and letter ‘i’ (lowercase ‘I’) was merged with ‘I’ (upper… view at source ↗
Figure 16
Figure 16. Figure 16: New added classes: ‘F’ in alternative style (first line) and crossed out letters (second line) [PITH_FULL_IMAGE:figures/full_fig_p011_16.png] view at source ↗
read the original abstract

The automatic processing of handwritten forms remains a challenging task, wherein detection and subsequent classification of handwritten characters are essential steps. We describe a novel approach, in which both steps -- detection and classification -- are executed in one task through a deep neural network. Therefore, training data is not annotated by hand, but manufactured artificially from the underlying forms and yet existing datasets. It can be demonstrated that this single-task approach is superior in comparison to the state-of-the-art two-task approach. The current study focuses on hand-written Latin letters and employs the EMNIST data set. However, limitations were identified with this data set, necessitating further customization. Finally, an overall recognition rate of 88.28 percent was attained on real data obtained from a written exam.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a single deep neural network that jointly executes detection and classification of handwritten Latin letters on forms, trained on synthetically manufactured data created by overlaying EMNIST characters onto form templates rather than using manual annotations. It claims this unified approach is superior to the conventional state-of-the-art two-task pipeline and reports an overall recognition rate of 88.28% when evaluated on real handwritten exam data, while noting limitations in the EMNIST dataset that required customization.

Significance. If the central claims are substantiated with full methodological details and validation, the work could offer a practical simplification for handwritten form processing by collapsing separate detection and classification stages into one model, with the synthetic data generation method providing a scalable alternative to manual labeling. This would be relevant for applications such as automated exam grading, though the result's impact depends on demonstrating that the performance gain is not an artifact of the training distribution.

major comments (2)
  1. [Abstract] Abstract and results description: the superiority claim over the two-task baseline and the specific 88.28% recognition rate are presented without any architecture details, training protocol, baseline implementations, error analysis, or statistical tests, rendering the central performance and superiority assertions unverifiable from the manuscript.
  2. [Data generation / methods] Data generation section: the claim that artificially overlaid EMNIST characters suffice for training a model that generalizes to real exam forms rests on the untested assumption that this synthetic distribution matches real handwriting variability (slant, pressure, connected strokes, form noise); no quantitative checks such as feature histograms, domain-adaptation metrics, or ablation on real vs. synthetic test sets are reported, which directly undermines the generalization and superiority conclusions.
minor comments (1)
  1. [Abstract] The abstract mentions EMNIST limitations requiring customization but provides no description of the specific modifications made or their impact on the final model.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback. We address each major comment below and indicate planned revisions to improve verifiability and support for the generalization claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results description: the superiority claim over the two-task baseline and the specific 88.28% recognition rate are presented without any architecture details, training protocol, baseline implementations, error analysis, or statistical tests, rendering the central performance and superiority assertions unverifiable from the manuscript.

    Authors: The full manuscript provides architecture details (Section 3), training protocol (Section 4), baseline comparisons (Section 5.2), and error analysis (Section 5.3). The abstract is intentionally concise per journal norms but we agree it should be expanded for standalone clarity. We will revise the abstract to summarize key architecture elements, training protocol, and include the 88.28% result with context. Statistical significance tests comparing the single-task and two-task approaches will be added to the results section. revision: partial

  2. Referee: [Data generation / methods] Data generation section: the claim that artificially overlaid EMNIST characters suffice for training a model that generalizes to real exam forms rests on the untested assumption that this synthetic distribution matches real handwriting variability (slant, pressure, connected strokes, form noise); no quantitative checks such as feature histograms, domain-adaptation metrics, or ablation on real vs. synthetic test sets are reported, which directly undermines the generalization and superiority conclusions.

    Authors: The manuscript evaluates the model directly on real exam forms and explicitly notes EMNIST limitations plus required customizations. We agree that explicit domain-gap quantification would strengthen the generalization argument. In revision we will add feature histogram comparisons between synthetic and real data plus an ablation study reporting performance on held-out real versus synthetic test sets. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claim with no derivation chain or self-referential fitting

full rationale

The paper describes an empirical deep-learning pipeline that trains a single network on synthetically overlaid EMNIST data and reports 88.28 % recognition on held-out real exam forms. No equations, fitted parameters, or mathematical derivations are present that could reduce to self-definition, fitted-input-as-prediction, or self-citation load-bearing. The superiority claim is an experimental comparison against a two-task baseline; it does not rely on any internal construction that forces the outcome. The distribution-match assumption between synthetic and real handwriting is a standard generalization risk, not a circularity in any derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are identifiable from the abstract; the central claim rests on an unstated assumption that synthetic data faithfully represents real handwriting distributions.

pith-pipeline@v0.9.1-grok · 5643 in / 1039 out tokens · 17669 ms · 2026-06-27T18:27:30.791570+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 27 canonical work pages

  1. [1]

    Image classification on emnist-letters.https://paperswithcode.com/sota/ image-classification-on-emnist-letters, last accessed 2023/04/05

    2023

  2. [2]

    IOP Conf

    Adriano, J.E.M., Calma, K.A.S., Lopez, N.T., Parado, J.A., Rabago, L.W., Cabardo, J.M.: Digital conversion model for hand-filled forms using optical char- acter recognition (ocr). IOP Conf. Ser.: Mater. Sci. Eng.482, 012049 (2019). https://doi.org/10.1088/1757-899X/482/1/012049

    work page doi:10.1088/1757-899x/482/1/012049 2019

  3. [3]

    In: Couprie, M., Cousty, J., Kenmochi, Y., Mustafa, N

    Alh´ eriti` ere, H., Ama¨ ıeur, W., Cloppet, F., Kurtz, C., Ogier, J.M., Vincent, N.: Straight line reconstruction for fully materialized table extraction in degraded document images. In: Couprie, M., Cousty, J., Kenmochi, Y., Mustafa, N. (eds.) Discrete Geometry for Computer Imagery, pp. 317–329. Springer International Publishing, Cham (2019).https://doi...

    work page doi:10.1007/978-3-030-14085-4_25 2019

  4. [4]

    Computers in Human Behavior 27, 1834–1839 (2011).https://doi.org/10.1016/j.chb.2011.04.004

    Barchard, K.A., Pace, L.A.: Preventing human error: The impact of data entry methods on data accuracy and statistical results. Computers in Human Behavior 27, 1834–1839 (2011).https://doi.org/10.1016/j.chb.2011.04.004

    work page doi:10.1016/j.chb.2011.04.004 2011

  5. [5]

    In: 2011 Int

    Ciresan, D.C., Meier, U., Gambardella, L.M., Schmidhuber, J.: Convolutional neu- ral network committees for handwritten character classification. In: 2011 Int. Conf. on Document Analysis and Recognition. pp. 1135–1139. IEEE, Beijing, China (2011).https://doi.org/10.1109/ICDAR.2011.229

    work page doi:10.1109/icdar.2011.229 2011

  6. [6]

    Cohen, G., Afshar, S., Tapson, J., van Schaik, A.: Emnist: an extension of mnist to handwritten letters.http://arxiv.org/abs/1702.05373(2017)

    Pith/arXiv arXiv 2017

  7. [7]

    In: 2017 Int

    Cohen, G., Afshar, S., Tapson, J., van Schaik, A.: Emnist: Extending mnist to handwritten letters. In: 2017 Int. Joint Conf. on Neural Networks (IJCNN). pp. 2921–2926. IEEE, Anchorage, AK, USA (2017).https://doi.org/10.1109/ IJCNN.2017.7966217

    arXiv 2017

  8. [8]

    Computer Science (2005)

    Deodhare, D., Suri, N.R., Amit, R.: Preprocessing and image enhancement algo- rithms for a form-based intelligent character recognition system. Computer Science (2005)

    2005

  9. [9]

    In: Singh, S., Singh, M., Apte, C., Perner, P

    Gatos, B., Danatsas, D., Pratikakis, I., Perantonis, S.J.: Automatic table detection in document images. In: Singh, S., Singh, M., Apte, C., Perner, P. (eds.) Pattern Recognition and Data Mining, pp. 609–618. Springer Berlin Heidelberg (2005). https://doi.org/10.1007/11551188_67

    work page doi:10.1007/11551188_67 2005

  10. [10]

    Gesmundo, A.: A continual development methodology for large-scale multitask dynamic ml systems.http://arxiv.org/abs/2209.07326(2022)

    arXiv 2022

  11. [11]

    Girshick, R.: Fast r-cnn.http://arxiv.org/abs/1504.08083(2015)

    Pith/arXiv arXiv 2015

  12. [12]

    In: 2014 IEEE Conf

    Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for ac- curate object detection and semantic segmentation. In: 2014 IEEE Conf. on Com- puter Vision and Pattern Recognition. pp. 580–587. IEEE, Columbus, OH, USA (2014).https://doi.org/10.1109/CVPR.2014.81

    work page doi:10.1109/cvpr.2014.81 2014

  13. [13]

    Goswami, R., Sharma, O.P.: A review on character recognition techniques. Int. J. Computer Applications83, 18–23 (2013).https://doi.org/10.5120/14460-2737

    work page doi:10.5120/14460-2737 2013

  14. [14]

    In: Proc

    Green, E., Krishnamoorthy, M.: Model-based analysis of printed tables. In: Proc. 3rd Int. Conf. on Document Analysis and Recognition. pp. 214–217. IEEE Comput. Soc. Press, Montreal, Canada (1995).https://doi.org/10.1109/ICDAR.1995. 598979

    work page doi:10.1109/icdar.1995 1995

  15. [15]

    Journal of Information10(2016)

    Islam, N., Islam, Z., Noor, N.: A survey on optical character recognition system. Journal of Information10(2016)

    2016

  16. [16]

    Intelligent Character Recognition of Handwritten Forms 13 In: 2019 IEEE Winter Conf

    Jayasundara, V., Jayasekara, S., Jayasekara, H., Rajasegaran, J., Seneviratne, S., Rodrigo, R.: Textcaps: Handwritten character recognition with very small datasets. Intelligent Character Recognition of Handwritten Forms 13 In: 2019 IEEE Winter Conf. on Applications of Computer Vision (WACV). pp. 254–262 (2019).https://doi.org/10.1109/WACV.2019.00033

    work page doi:10.1109/wacv.2019.00033 2019

  17. [17]

    Jeevan, P., Viswanathan, K., Anand, A.S., Sethi, A.: Wavemix: A resource-efficient neural network for image analysis.http://arxiv.org/abs/2205.14375(2023)

    arXiv 2023

  18. [18]

    In: 2019 Int

    Jha, M., Kabra, M., Jobanputra, S., Sawant, R.: Automation of cheque transaction using deep learning and optical character recognition. In: 2019 Int. Conf. on Smart Systems and Inventive Technology (ICSSIT). pp. 309–312. IEEE, Tirunelveli, India (2019).https://doi.org/10.1109/ICSSIT46314.2019.8987925

    work page doi:10.1109/icssit46314.2019.8987925 2019

  19. [19]

    Kabir, H.M.D., Abdar, M., Jalali, S.M.J., Khosravi, A., Atiya, A.F., Nahavandi, S., Srinivasan, D.: Spinalnet: Deep neural network with gradual input.http:// arxiv.org/abs/2007.03347(2022)

    arXiv 2007

  20. [20]

    In: Zhang, Y.D., Mandal, J.K., So-In, C., Thakur, N.V

    Khobragade, R.N., Koli, N.A., Lanjewar, V.T.: Challenges in recognition of on- line and off-line compound handwritten characters: A review. In: Zhang, Y.D., Mandal, J.K., So-In, C., Thakur, N.V. (eds.) Smart Trends in Computing and Communications, pp. 375–383. Springer Singapore (2020).https://doi.org/10. 1007/978-981-15-0077-0_38

    2020

  21. [21]

    Khobragade, R.N., Koli, N.A., Makesar, M.S.: A survey on recognition of devnagari script. Int. J. Computer Applications (2013)

    2013

  22. [22]

    Pattern Analysis and Applications5, 31–45 (2002).https://doi.org/10.1007/s100440200004

    Khorsheed, M.S.: Off-line arabic character recognition – a review. Pattern Analysis and Applications5, 31–45 (2002).https://doi.org/10.1007/s100440200004

    work page doi:10.1007/s100440200004 2002

  23. [23]

    Kumar Shrivastava, S., Chaurasia, P.: Handwritten devanagari lipi using support vector machine. Int. J. Computer Applications43, 20–25 (2012).https://doi. org/10.5120/6220-8785

    work page doi:10.5120/6220-8785 2012

  24. [24]

    Li, M., Cui, L., Huang, S., Wei, F., Zhou, M., Li, Z.: Tablebank: A benchmark dataset for table detection and recognition (2019).https://doi.org/10.48550/ ARXIV.1903.01949

    arXiv 2019

  25. [25]

    Li, W., Feng, X.S., Zha, K., Li, S., Zhu, H.S.: Summary of target detection al- gorithms. J. Phys.: Conf. Ser.1757, 012003 (2021).https://doi.org/10.1088/ 1742-6596/1757/1/012003

    2021

  26. [26]

    IEEE Access8, 142642–142668 (2020).https://doi.org/10.1109/ACCESS.2020.3012542

    Memon, J., Sami, M., Khan, R.A., Uddin, M.: Handwritten optical character recog- nition (ocr): A comprehensive systematic literature review (slr). IEEE Access8, 142642–142668 (2020).https://doi.org/10.1109/ACCESS.2020.3012542

    work page doi:10.1109/access.2020.3012542 2020

  27. [27]

    In: 8th Int

    Nath, G.: Isolated ocr for handwritten forms: An application in the education domain. In: 8th Int. Conf. of Business Analytics (2022)

    2022

  28. [28]

    In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F

    Ngo, P.: Digital line segment detection for table reconstruction in document im- ages. In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F. (eds.) Image Analysis and Processing – ICIAP 2022, pp. 211–224. Springer International Publishing, Cham (2022).https://doi.org/10.1007/978-3-031-06430-2_18

    work page doi:10.1007/978-3-031-06430-2_18 2022

  29. [29]

    In: 2020 IEEE/CVF Conf

    Pad, P., Narduzzi, S., Kundig, C., Turetken, E., Bigdeli, S.A., Dunbar, L.A.: Efficient neural vision systems based on convolutional image acquisition. In: 2020 IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). pp. 12282–12291. IEEE, Seattle, WA, USA (2020).https://doi.org/10.1109/ CVPR42600.2020.01230

    arXiv 2020

  30. [30]

    Pal, A., Singh, D.: Handwritten english character recognition using neural network. Int. J. Computer Science and Communication (2010)

    2010

  31. [31]

    Pattern Recognition40, 2110–2117 (2007).https://doi.org/10.1016/j.patcog

    Patil, P.M., Sontakke, T.R.: Rotation, scale and translation invariant handwrit- ten devanagari numeral character recognition using general fuzzy neural network. Pattern Recognition40, 2110–2117 (2007).https://doi.org/10.1016/j.patcog. 2006.12.018 14 H. Grabowski

    work page doi:10.1016/j.patcog 2007

  32. [32]

    Prasad, D., Gadpal, A., Kapadni, K., Visave, M., Sultanpure, K.: Cascadetabnet: An approach for end to end table detection and structure recognition from image- based documents (2020).https://doi.org/10.48550/ARXIV.2004.12629

    work page doi:10.48550/arxiv.2004.12629 2020

  33. [33]

    In: 2016 Int

    Priya, A., Mishra, S., Raj, S., Mandal, S., Datta, S.: Online and offline character recognition: A survey. In: 2016 Int. Conf. on Communication and Signal Processing (ICCSP). pp. 0967–0970. IEEE (2016).https://doi.org/10.1109/ICCSP.2016. 7754291

    work page doi:10.1109/iccsp.2016 2016

  34. [34]

    In: 2022 8th Int

    Raj, S., Gupta, Y., Malhotra, R.: License plate recognition system using yolov5 and cnn. In: 2022 8th Int. Conf. on Advanced Computing and Communication Systems (ICACCS). pp. 372–377. IEEE, Coimbatore, India (2022).https://doi. org/10.1109/ICACCS54159.2022.9784966

    work page doi:10.1109/icaccs54159.2022.9784966 2022

  35. [35]

    Rao, N.V., Sastry, A.S.C.S., Chakravarthy, A.S.N., Kalyanchakravarthi, P.: Optical character recognition technique algorithms. J. Theoretical and Applied Information Technology83(2016)

    2016

  36. [36]

    Rasmussen, L.V., Peissig, P.L., McCarty, C.A., Starren, J.: Development of an optical character recognition pipeline for handwritten form fields from an electronic health record. J. Am. Med. Inform. Assoc.19, e90–e95 (2012).https://doi.org/ 10.1136/amiajnl-2011-000182

    work page doi:10.1136/amiajnl-2011-000182 2012

  37. [37]

    IEEE Trans

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39, 1137–1149 (2017).https://doi.org/10.1109/TPAMI.2016.2577031

    work page doi:10.1109/tpami.2016.2577031 2017

  38. [38]

    In: 2018 Int

    Shawon, A., Rahman, M.J.U., Mahmud, F., Zaman, M.M.A.: Bangla handwritten digit recognition using deep cnn for large and unbiased dataset. In: 2018 Int. Conf. on Bangla Speech and Language Processing (ICBSLP). pp. 1–6. IEEE, Sylhet (2018).https://doi.org/10.1109/ICBSLP.2018.8554900

    work page doi:10.1109/icbslp.2018.8554900 2018

  39. [39]

    Imbalanced data problem in machine learning: A review,

    Shi, H., Zhao, D.: License plate recognition system based on improved yolov5 and gru. IEEE Access11, 10429–10439 (2023).https://doi.org/10.1109/ACCESS. 2023.3240439

    work page doi:10.1109/access 2023

  40. [40]

    Singh, S., Tiwari, S.: Application of image processing and convolution networks in intelligent character recognition for digitized forms processing. Int. J. Computer Applications179, 7–13 (2018).https://doi.org/10.5120/ijca2018915460

    work page doi:10.5120/ijca2018915460 2018

  41. [41]

    In: Ninth Int

    Smith, R.: An overview of the tesseract ocr engine. In: Ninth Int. Conf. on Docu- ment Analysis and Recognition (ICDAR 2007). vol. 2, pp. 629–633. IEEE, Curitiba, Parana, Brazil (2007).https://doi.org/10.1109/ICDAR.2007.4376991

    work page doi:10.1109/icdar.2007.4376991 2007

  42. [42]

    Somashekar, T.: A survey on handwritten character recognition using machine learning technique. J. Univ. Shanghai Sci. Technol.23, 1019–1024 (2021).https: //doi.org/10.51201/JUSST/21/05304

    work page doi:10.51201/jusst/21/05304 2021

  43. [43]

    EAI Endorsed Trans

    Suriya, S., Dhivya, S., Balaji, M.: Intelligent character recognition system us- ing convolutional neural network. EAI Endorsed Trans. Cloud Systems6, 166659 (2020).https://doi.org/10.4108/eai.16-10-2020.166659

    work page doi:10.4108/eai.16-10-2020.166659 2020

  44. [44]

    Applied Sciences 12, 5361 (2022).https://doi.org/10.3390/app12115361

    Tang, M., Xie, S., He, M., Liu, X.: Character recognition in endangered archives: Shui manuscripts dataset, detection and application realization. Applied Sciences 12, 5361 (2022).https://doi.org/10.3390/app12115361

    work page doi:10.3390/app12115361 2022

  45. [45]

    02696(2022)

    Wang, C.Y., Bochkovskiy, A., Liao, H.Y.M.: Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors.http://arxiv.org/abs/2207. 02696(2022)

    2022

  46. [46]

    Wu, Y., Kirillov, A., Massa, F., Lo, W.Y., Girshick, R.: Detectron2.https:// github.com/facebookresearch/detectron2, last accessed 2023/04/05

    2023