Ancient but Digitized: Developing Handwritten Optical Character Recognition for East Syriac Script Through Creating KHAMIS Dataset

What it is about

The authors collect KHAMIS, 624 handwritten East Syriac sentences (line images with ground-truth text files), written by 31 students and one professor from a 20-verse poem. They fine-tune Tesseract's Syriac LSTM model on it with the tesstrain tool (learning rate 0.0001, 10,000 iterations) and compare three train/evaluation splits (90/10, 80/20, 70/30). The character error rate (CER) on the evaluation sets is about 9 to 10.5 per cent. On a separate test set of 13 samples written by people outside the group, with added noise, CER is 18.9 to 19.7 per cent and word error rate 62.8 to 65.4 per cent, which the authors report as about twice as good as the default Syriac model. Preprocessing is a mean blur and a global threshold. The authors argue that CER should be the main metric, because WER marks a whole word wrong for one bad character.

Ideas

I1. Make a fixed, held-out line-image set with ground truth for tools/ocrfix, scored by CER first and WER second, with the split reported as well. Use it before and after each correction step, so that the gain is measured and not assumed.
I2. Add the paper's preprocessing as the first stage of tools/ocrfix for scanned Hebrew and Aramaic editions: grayscale, mean blur, global threshold. Measure CER on the same pages with and without it.
I3. A robustness test built the paper's way: text written by people outside the training group, with one word swapped for a name and random characters replaced. The same noise recipe applied to Hebrew lines shows how much the correction step can recover.

For Mova

I1, I2 and I3 go into tools/ocrfix. The letter-form variants that the paper describes (final forms, the kaf and nun shapes) are a check for 1. Script and pointing. The dataset itself is NC and is not used; only the protocol is taken.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2408.13631

@misc{majeed2024ancient,
  title = {{Ancient but Digitized: Developing Handwritten Optical Character Recognition for East Syriac Script Through Creating KHAMIS Dataset}},
  author = {Majeed, Ameer and Hassani, Hossein},
  year = {2024},
  eprint = {2408.13631},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.13631}
}

References

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

Ackroyd, P.R., Evans, C.F. (eds.): The Cambridge History of the Bible: Volume 1, From the Beginning to Jerome. Cambridge University Press, Cambridge, England (1970) find
Alvermann, D.: Word Error Rate & Character Error Rate – How to evaluate a model (2019). https://shorturl.at/88GVf ↗
Amrouch, M., Rabi, M., Es-Saady, Y.: Convolutional feature learning and cnn based hmm for arabic handwriting recognition. In: Mansouri, A., El Moataz, A., Nouboud, F., Mammass, D. (eds.) Image and Signal Processing, pp. 265–274. Springer, Cham (2018) find
Bilane, P., Bres, S., Emptoz, H.: Local orientation extraction for wordspotting in syriac manuscripts. In: Elmoataz, A., Lezoray, O., Nouboud, F., Mammass, D. (eds.) Image and Signal Processing, pp. 481–489. Springer, Berlin, Heidelberg (2008) find
Butts, A.M.: Syriac Syriac-Language Language Clocksin, W.F., Fernando, P.J.: Towards Automatic Transciption of Syriac Handwriting. In: 12th International Conference on Image Analysis and Processing. IEEE, Mantova, Italy (2003) find
Chesley, E., Marcantonio, J., Pearson, A.: Towards syriac digital corpora: Evaluation of tesseract 4.0 for syriac ocr. Hugoye: Journal of Syriac Studies (2019) find
Dehghan, M., Faez, K., Ahmadi, M., Shridhar, M.: Handwritten farsi (arabic) word recognition: a holistic approach using discrete hmm. Pattern Recognition 34(5), 1057–1065 (2001) find
Davis, R.H., Lyall, J.: Recognition of handwritten characters - a review. Image and Vision Computing (1986) find
Fermanian, R., Yaacoub, C., Bilane, P.: Deep recognition-based character segmentation in handwritten syriac manuscripts. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA) (2020) find
Gonzalez, R.C., Woods, R.E.: Digital Image Processing. Pearson, N/A (2018) find
Hamid, A., Haraty, R.: A neuro-heuristic approach for segmenting handwritten arabic text. In: Proceedings ACS/IEEE International Conference on Computer Systems and Applications (2001) find
Khorsheed, M.S.: Recognising handwritten arabic manuscripts using a single hidden markov model. Pattern Recognition Letters 24(14), 2235–2242 (2003) find
Lorigo, L., Govindaraju, V.: Segmentation and pre-recognition of arabic handwriting. In: Eighth International Conference on Document Analysis and Recognition (ICDAR’05). IEEE, Seoul, Korea (South) (2005) find
Mengozzi, A.: Khamis bar Qardahe (2011). https://gedsh.bethmardutho.org/Khamis-bar-Qardahe ↗
Maalej, R., Kherallah, M.: Convolutional neural network and blstm for offline arabic handwriting recognition. In: 2018 International Arab Conference on Information Technology (ACIT), pp. 1–6 (2018) find
Maalej, R., Tagougui, N., Kherallah, M.: Recognition of handwritten arabic words with dropout applied in mdlstm. In: Image Analysis and Recognition, pp. 746–752. Springer, Cham (2016) find
Omniglot: Syriac Alphabet (2023). https://www.omniglot.com/writing/syriac.htm ↗
OpenCV: OpenCV: Smoothing Images (2024). https://docs.opencv.org/4.x/d4/d13/ ↗
(2011). https://gedsh.bethmardutho.org/tutorial py filtering.html Sayre, K.M.: Machine recognition of handwritten words: A project report. In: Pattern Recognition, vol. 5, pp. 213–228 (1973). Pergamon Press The Republic of Iraq: Constitution Project (2005). https://www.constituteproject. org/constitution/Iraq 2005 ↗
Tesseract-OCR: Tesseract-OCR (2024). https://github.com/tesseract-ocr/tesseract ↗
UNESCO: World Atlas of Languages (2010). https://en.wal.unesco.org/discover/languages?text=aramaic ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/langstream-2024-majeed-ancient-but-digitized-developing-handwritten/