Character recognition system for pegon typed manuscript.

Area of Science:

• Computational Linguistics
• Digital Humanities
• Document Image Analysis

Background:

• The Pegon script, an Arabic-based writing system for Javanese and other Indonesian languages, is endangered, primarily existing in private collections and Islamic boarding schools (pesantren).
• Digitization through Optical Character Recognition (OCR) is crucial for preserving the Pegon script, yet no prior research exists on OCR systems for this specific script.
• Existing Arabic OCR systems are not directly applicable, necessitating tailored solutions.

Purpose of the Study:

• To develop and evaluate an OCR system specifically for typed Pegon manuscripts.
• To introduce novel synthesized and real annotated datasets for training and testing Pegon OCR models.
• To compare the performance of deep learning techniques against conventional methods for Pegon text recognition.

Main Methods:

• Development of synthesized and real annotated datasets for Pegon script.
• Implementation of a deep learning-based OCR pipeline.
• Utilized YOLOv5 for text line segmentation and a CTC-CRNN (Connectionist Temporal Classification-Convolutional Neural Network Recurrent Neural Network) architecture for text recognition.

Main Results:

• Deep learning techniques significantly outperformed conventional methods, which failed to detect Pegon text.
• The proposed system achieved an F1-score of 0.94 for text line segmentation.
• The system demonstrated a Character Error Rate (CER) of 0.03 for text recognition.

Conclusions:

• The developed OCR system, leveraging deep learning, is effective for digitizing Pegon typed manuscripts.
• The novel datasets and proposed methodology provide a foundation for future research in Pegon script preservation.
• This work highlights the potential of adapted deep learning architectures for recognizing under-resourced scripts.