SARD: Synthetic Arabic Recognition Dataset

Overview

SARD (Synthetic Arabic Recognition Dataset) is a large-scale, synthetically generated dataset designed for training and evaluating Optical Character Recognition (OCR) models for Arabic text. This dataset addresses the critical need for comprehensive Arabic text recognition resources by providing controlled, diverse, and scalable training data that simulates real-world book layouts.

Key Features

• Massive Scale: 2,621,075 document images containing 794.6 million words

• Typographic Diversity: Five distinct Arabic fonts (Amiri, Sakkal Majalla, Arial, Calibri, Scheherazade New and Traditional Arabic)

• Structured Formatting: Designed to mimic real-world book layouts with consistent typography

• Clean Data: Synthetically generated with no scanning artifacts, blur, or distortions

• Content Diversity: Text spans multiple domains including culture, literature, Shariah, social topics, and more

• Storage Size: 1.54 TB

Dataset Structure

The dataset is divided into five splits based on font name:

• Amiri: ~488585 document images
• Sakkal Majalla: ~488585 document images
• Arial: ~321683 document images
• Calibri: ~321683 document images
• Scheherazade New: ~572863 document images
• Traditional Arabic: ~428676 document images document images

📋 Sample Images

Each split contains data specific to a single font with the following attributes:

• image_name: Unique identifier for each image
• chunk: The text content associated with the image
• font_name: The font used in text rendering
• image_base64: Base64-encoded image representation
• sample_id: Unique ID
• article_link: Link of the source article

Content Distribution

Font Specifications

Page Layout

Usage Example

from datasets import load_dataset
import base64
from io import BytesIO
from PIL import Image
import matplotlib.pyplot as plt

# Load dataset with streaming enabled
ds = load_dataset("riotu-lab/SARD", streaming=True)
print(ds)

# Iterate over a specific font dataset (e.g., Amiri)
for sample in ds["Amiri"]:
    image_name = sample["image_name"]
    chunk = sample["chunk"]  # Arabic text transcription
    font_name = sample["font_name"]
    
    # Decode Base64 image
    image_data = base64.b64decode(sample["image_base64"])
    image = Image.open(BytesIO(image_data))

    # Display the image
    plt.figure(figsize=(10, 10))
    plt.imshow(image)
    plt.axis('off')
    plt.title(f"Font: {font_name}")
    plt.show()

    # Print the details
    print(f"Image Name: {image_name}")
    print(f"Font Name: {font_name}")
    print(f"Text Chunk: {chunk}")
    
    # Break after one sample for testing
    break

Applications

SAND is designed to support various Arabic text recognition tasks:

• Training and evaluating OCR models for Arabic text
• Developing vision-language models for document understanding
• Fine-tuning existing OCR models for better Arabic script recognition
• Benchmarking OCR performance across different fonts and layouts
• Research in Arabic natural language processing and computer vision

Acknowledgments

The authors thank Prince Sultan University for their support in developing this dataset.

Citation

If you use SARD in your work, please cite the following paper:

APA:

Nacar, O., Al-Habashi, Y., Sibaee, S., Ammar, A., & Boulila, W. (2025). SARD: A Large-Scale Synthetic Arabic OCR Dataset for Book-Style Text Recognition. arXiv preprint arXiv:2505.24600.
https://arxiv.org/abs/2505.24600

BibTeX:

@misc{nacar2025sard,
      title={SARD: A Large-Scale Synthetic Arabic OCR Dataset for Book-Style Text Recognition}, 
      author={Omer Nacar and Yasser Al-Habashi and Serry Sibaee and Adel Ammar and Wadii Boulila},
      year={2025},
      eprint={2505.24600},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.24600}, 
}

Usage Note:

This dataset includes text sourced from Alukah. Use is limited to non-commercial purposes in accordance with the source terms. Redistribution, republication, or commercial use is not permitted without permission and proper attribution to the original source and author.