Fine-Tuning an Arabic OCR Model using Tesseract 5

Table of Contents

• Introduction
• Key Areas of Focus
• Methodology
  • Data Collection and Preprocessing
  • Dataset Composition
  • Image Generation
• Data Augmentation
• Training
• Evaluation
• Results

Introduction

Optical Character Recognition (OCR) is essential for digital processing and preservation of textual information. Despite advancements in OCR for various languages, Arabic OCR remains challenging due to its complex script. This project aims to fine-tune an Arabic OCR system using Tesseract 5.0, achieving high accuracy in text recognition.

Key Areas of Focus

• Data Collection and Preprocessing
• Dataset Composition
• Image Generation
• Training and Evaluation
• Challenges and Limitations

Methodology

Data Collection and Preprocessing

We built a comprehensive Arabic dataset by collecting Arabic words from various online sources, ensuring quality through preprocessing steps like removing non-Arabic characters and normalizing text.

Dataset Composition

Our dataset included about 1 million Arabic sentences, split into:

• 60% tashkeel-free sentences
• 40% tashkeel-included sentences

This balance helps the OCR model handle both scenarios effectively.

Image Generation

Using the Text Recognition Data Generator (TRDG) library, we created images from Arabic text, ensuring diversity with three distinct Arabic fonts: Baghdad, Bahij, and Rakkas. Images were generated in .tif format with corresponding ground truth text in .gt.txt format.

Data Augmentation

Future improvements will include advanced data augmentation techniques to enhance the model's robustness, such as noise injection, background variation, and character distortions.

Training

We fine-tuned the pre-trained Arabic model ara.traineddata from the official Tesseract repository. Key parameters included:

Maximum iterations: 1000 Lang type: RTL PSM: 13

Evaluation

We evaluated the model using the ocreval tool to measure its performance. The evaluation was conducted on a benchmarking dataset created by Hegghammer, which provided a comprehensive and rigorous assessment of our model's capabilities. We compared the results with pre-trained models and benchmarks from related research, focusing on metrics such as character error rate (CER) and word error rate (WER).

Results

Our fine-tuned model demonstrated significant improvements in word error rate (WER) on datasets without noise, outperforming both the pre-trained Tesseract model and Hegghammer's Tesseract benchmarking results. However, on datasets that included noise, our model's performance was not as strong, highlighting areas for potential improvement in handling noisy data.

License

This project is licensed under the MIT License.