Datasets:

craneset
/

arabic-ocr

Modalities:

Image

Text

Formats:

Languages:

Size:

Tags:

Libraries:

License:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 8 rows

Search is not available for this dataset

image imagewidth (px) 4.96k 4.96k

ocr-data

Alshams, The largest Arabic OCR dataset at the word level.

This dataset is specifically designed for fine-grained word-level OCR tasks, providing precise word-level bounding box annotations for each image.

Each word is annotated with pixel-accurate localization, enabling tasks such as text detection, text recognition, and end-to-end OCR.

This repository currently provides an Arabic OCR dataset.

📦 Available Datasets

The technical specifications of each dataset are listed in the table below.

Language	Version	Link	Pages Count	Unique Words	Fonts Count	Full Dataset Link
Arabic	v1.0	Gdrive (only 25k pages)	~521K	3012869	1	To obtain the full dataset, please contact us at: craneset[at]outlook.com [not free]
Arabic	v2.0	Gdrive (only 13k pages)	~534K	2502545	5	To obtain the full dataset, please contact us at: craneset[at]outlook.com [not free]

*The total dataset in Arabic is over 1M pages and 6 separate fonts.
*See the sample folder for examples of each font.
*The count of unique words after removing numbers and punctuation has been calculated.

📁 Dataset Structure

The dataset is organized into three main directories at the root level:

ocr-data/
├── images/
├── labels/
├── texts/

📷 images/

Contains OCR images in PNG format.
Each image has a corresponding JSON annotation file with the same base filename.
The JSON file precisely defines the location of each word in the image.

🏷 labels/

Contains JSON annotation files corresponding to the images.
Each JSON file shares the same base filename as its related image.
These files define word-level annotations with exact bounding box coordinates.
The annotation structure is identical to the JSON format described in the images/ section.

JSON Annotation Format (per image)

import json
with open(path_image_label, 'r', encoding='utf-8') as f:
    data = json.load(f)

{
  "0": {
    "word": "كلمة",
    "location": {
      "x": 3927,
      "y": 481,
      "w": 397,
      "h": 170
    }
  },
  "1": {
    "word": "عربية",
    "location": {
      "x": 3544,
      "y": 481,
      "w": 355,
      "h": 170
    }
  }
}

word: The recognized word in the image.
location: Bounding box of the word:
- x, y: Top-left corner coordinates
- w, h: Width and height of the bounding box