Elamur Research⌕
☰
home / papers / 2020 / A Hybrid Deep Learning Model for Arabic Text Recognition

A Hybrid Deep Learning Model for Arabic Text Recognition

Summary

Printed Arabic is hard for OCR. The script is cursive, letters join within words, ligatures are common, letter shapes change with position, and many letters differ only by their dots. Labelled data is also scarce. The authors target an especially difficult case: printed text set in fonts that imitate handwriting. They built a segmentation-free, end-to-end recognizer that feeds a CNN feature extractor into a bidirectional LSTM. It is small enough to train on a single CPU. For training and testing they compiled the Arabic Multi-Fonts Dataset (AMFDS): more than two million word images, rendered in 18 highly cursive fonts from words in an Arabic Wikipedia dump.

One model covering all fonts reached 98.76% character recognition rate (CRR) and 90.22% word recognition rate (WRR). Performance dropped on an unseen dataset (85.15% CRR) and on noise-corrupted images (77.29% CRR).

The model, the datasets and the generation toolkit are public, so others can extend the work to new fonts. The authors suggest adding a language model to improve word-level accuracy.

Key points

  • Architecture: CNN layers followed by a BDLSTM. The CNN output is a sequence of 32 time steps × 256 features, decoded into an alphabet of about 39 symbols, with no explicit character segmentation.
  • AMFDS: over 2 million synthetic word images in 18 Arabic fonts that mimic handwriting, with words sampled from an Arabic Wikipedia dump.
  • On seen data, one model for all fonts scored 98.76% CRR and 90.22% WRR.
  • Generalization was weaker: 85.15% CRR on a disjoint unseen set and 77.29% CRR on images with salt-and-pepper and speckle noise.
  • The released model, datasets and generation toolkit allow retraining for other fonts.

Interesting ideas

  • I1. Large labelled data can be generated by rendering real corpus text under many surface styles, which removes the need for manual annotation and keeps vocabulary realistic. — For Mova (learn): Render known-good sentences through controlled surface variations (contractions, casing, punctuation styles) to mass-produce labelled training pairs whose gold analysis is preserved by construction.
  • I2. Skipping a brittle pre-segmentation step and letting a sequence model align input to output avoids errors where unit boundaries are ambiguous. — For Mova (tokens): For fused forms like clitics, contractions or hyphenated compounds, a character-sequence model could propose splits jointly with tags instead of relying on a fixed rule-based tokenizer first.
  • I3. Reporting results separately on seen data, a disjoint unseen set and artificially noised inputs exposes how much a high headline score depends on familiar conditions. — For Mova (learn): Add a held-out-domain split and a noise-injected split (typos, dropped punctuation) to every parser evaluation as standing stress controls next to the in-domain score.
  • I4. Character-level and word-level rates are reported side by side, and the large gap between them shows how small unit errors compound into whole-item failures. — For Mova (tree): Report per-arc LAS alongside whole-sentence exact match, so it is visible when rare errors still corrupt most sentences fed to frames.
  • I5. The appendix inspects which symbol wins at each time step, giving an interpretable trace of what the network "reads" at each position. — For Mova (tokens): Expose per-position top candidates and their scores from neural taggers, so the explainable pipeline can show why a token got its tag and where a confusion arose.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cited by

2 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 1 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.14569/issn.2156-5570 · arXiv 2009.01987

@misc{fasha2020hybrid,
  title = {{A Hybrid Deep Learning Model for Arabic Text Recognition}},
  author = {Fasha, Mohammad and Hammo, Bassam and Obeid, Nadim and Widian, Jabir},
  year = {2020},
  doi = {10.14569/issn.2156-5570},
  eprint = {2009.01987},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.14569/issn.2156-5570}
}

References

23 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. M. Alghamdi and W. Teahan, “Printed Arabic script recognition: A survey,” Int. J. Adv. Comput. Sci. Appl., vol. 9, no. 9, pp. 415–428, 2018, doi: 10.14569/ijacsa.2018.090953. ↗
  2. A. D. Riaz Ahmad, Saeeda Naz, Muhammad Afzal, Sheikh Rashid, Marcus Liwicki, “A Deep Learning based Arabic Script Recognition System,” Int. Arab J. Inf. Technol., no. October, 2017. find
  3. A. A. Shahin, “Printed Arabic Text Recognition using Linear and Nonlinear Regression,” Int. J. Adv. Comput. Sci. Appl., vol. 8, no. 1, pp. 227–235, 2017, doi: 10.14569/ijacsa.2017.080129. ↗
  4. A. E. Hassanien, M. F. Tolba, and A. T. Azar, “Isolated Printed Arabic Character Recognition Using KNN and Random Forest Tree Classifier,” Commun. Comput. Inf. Sci., vol. 488, no. November, pp. 10–17, 2014, doi: 10.1007/978-3-319-13461-1. ↗
  5. A. Zoizou, A. Zarghili, and I. Chaker, “A new hybrid method for Arabic multi-font text segmentation, and a reference corpus construction,” J. King Saud Univ. - Comput. Inf. Sci., 2018, doi: 10.1016/j.jksuci.2018.07.003. ↗
  6. A. Graves, “Offline Arabic Handwriting Recognition with Multidimensional Recurrent Neural with Multidimensional Recurrent Neural Networks,” Adv. Neural Inf. Process. Syst., no. January 2008, 2008, doi: 10.1007/978-1-4471-4072-6. ↗
  7. M. Pechwitz, S. Snoussi, and N. Ellouze, “IFN / ENIT-database of handwritten Arabic words,” no. May 2014, 2002. find
  8. H. A. Bodour Alreha,Najla Alsaedi, “Historical Arabic Manuscripts Text Recognition Using Convolutional Neural Network,” Conf. Data Sci. Mach. Learn. Appl. Hist., 2020. find
  9. A. Mohsin and M. Sadoon, “Developing an Arabic Handwritten Recognition System by Means of Artificial Neural Network,” J. Eng. Appl. Sci., vol. 15, no. 1, pp. 1–3, 2019, doi: 10.36478/jeasci.2020.1.3. ↗
  10. K. Younis and A. Khateeb, “Arabic Hand-Written Character Recognition Based on Deep Convolutional Neural Networks,” Jordanian J. Comput. Inf. Technol., vol. 3, no. 3, p. 186, 2017, doi: 10.5455/jjcit.71-1498142206. ↗
  11. M. Torki, M. E. Hussein, A. Elsallamy, M. Fayyaz, and S. Yaser, Window-Based Descriptors for Arabic Handwritten Alphabet Recognition: A Comparative Study on a Novel Dataset. 2014. find
  12. H. E.-B. Ahmed El-Sawy, Mohamed Loey, “Arabic Handwritten Characters Recognition using Convolutional Neural Network,” WSEAS Trans. Comput. Res., no. 5, pp. 11–19, 2017, doi: 10.1109/IACS.2019.8809122. ↗
  13. S. A. Mahmoud et al., “KHATT: Arabic offline Handwritten Text Database,” Proc. - Int. Work. Front. Handwrit. Recognition, IWFHR, pp. 449–454, 2012, doi: 10.1109/ICFHR.2012.224. ↗
  14. M. Jain, M. Mathew, and C. V. Jawahar, “Unconstrained OCR for Urdu using deep CNN-RNN hybrid networks,” Proc. - 4th Asian Conf. Pattern (IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 11, No. 8, 2020 Recognition, ACPR 2017, pp. 753–758, 2018, doi: 10.1109/ACPR.2017.5. ↗
  15. F. Slimane, S. Kanoun, H. El Abed, A. M. Alimi, R. Ingold, and J. Hennebert, “ICDAR2013 competition on multi-font and multi-size digitally represented arabic text,” Proc. Int. Conf. Doc. Anal. Recognition, ICDAR, no. August, pp. 1433–1437, 2013, doi: 10.1109/ICDAR.2013.289. ↗
  16. F. K. Jaiem, S. Kanoun, M. Khemakhem, H. El Abed, and J. Kardoun, “Database for Arabic printed text recognition research,” Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 8156 LNCS, no. PART 1, pp. 251–259, 2013, doi: 10.1007/978-3-642-41181-6_26. ↗
  17. A. G. Al-Hashim and S. A. Mahmoud, “Benchmark database and GUI environment for printed arabic text recognition research,” WSEAS Trans. Inf. Sci. Appl., vol. 7, no. 4, pp. 587–597, 2010. find
  18. R. Davidson, R., Hopely, “Arabic and Persian Training and Test Data Sets,” 1997. A. CNN and RNN Layers Sample Outputs (a) CNN Layer [1,32]. (b) CNN Layer [2,32]. find
  19. L. Alhomed and K. Jambi , “A Survey on the Existing Arabic Optical Character Recognition and Future Trends ,” Int. J. Adv. Res. Comput. Commun. Eng., vol. 7, no. 3, pp. 78–88, 2018, doi: 10.17148/IJARCCE.2018.7213. ↗
  20. I. Ahmed, S. A. Mahmoud, and M. T. Parvez, “Guide to OCR for Arabic Scripts,” Guid. to OCR Arab. Scripts, no. November 2017, 2012, doi: 10.1007/978-1-4471-4072-6. ↗
  21. H. Scheidl, “Handwritten Text Recognition in Historical Documents,” 2011. find
  22. S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” 32nd Int. Conf. Mach. Learn. ICML 2015, vol. 1, pp. 448–456, 2015. find
  23. G. Hinton, “Neural Networks for Machine Learning - Lecture 6a - Overview of mini-batch gradient descent.,” 2012, doi: 10.1017/9781139051699.031. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/fasha-2020-hybrid-deep-learning-model-arabic/