Exploiting Script Similarities to Compensate for the Large Amount of Data in Training Tesseract LSTM: Towards Kurdish OCR
Authors
Applied Sciences
Vol. 11(20)
pp. 9752
October 19, 2021
Abstract
This study adapts an Arabic Tesseract LSTM model to printed Sorani Kurdish using limited Kurdish ground truth and synthetic font-based training material. The authors report 95.45% average character accuracy on an evaluation set of 11 court-document images. A separate Tesseract evaluation reports 2.93% CER and 13.52% WER for the kur_ara model; these results come from different evaluation protocols.
Keywords
OCR
Sorani Kurdish
Tesseract
LSTM
Printed Text
Related Datasets
KurdishOCR Ground Truth
2021
522 line-image/transcription pairs reported by the associated paper
PNG, TXT
Ground-truth material released for printed Sorani Kurdish OCR research by Saman Idrees and Hossein Hassani. The associated study reports 522 line-image/transcription pairs. The repository distributes …
Citation
Idrees, S., & Hassani, H. (2021). Exploiting Script Similarities to Compensate for the Large Amount of Data in Training Tesseract LSTM: Towards Kurdish OCR. Applied Sciences, 11(20), 9752.