metadata

language:
  - ar
license: cc-by-nc-sa-4.0
task_categories:
  - image-to-text
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: image
      dtype: image
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 1274826565.992
      num_examples: 22092
    - name: validation
      num_bytes: 60515423.616
      num_examples: 1069
    - name: test
      num_bytes: 71481336.008
      num_examples: 1334
  download_size: 1317284681
  dataset_size: 24495
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
tags:
  - ocr
  - htr
  - historic

Muharaf-public line images

This dataset contains 24,495 line images of Arabic handwriting and the corresponding text. The images are Manuscripts from the 19th to 21st century. See the official code, paper, and zenodo archive below. Our work has been accepted to NeurIPS 2024.

How to use

from datasets import load_dataset

ds = load_dataset("aamijar/muharaf-public")

# remember to index into the row first to avoid decoding all images at once (https://huggingface.co/docs/datasets/en/image_load#load-image-data)
print(ds["train"][0]["image"])

# instead of

~~print(ds["train"]["image"][0])~~

Citation

If you are using this dataset in your work, please cite our paper below.

@misc{saeed2024muharafmanuscriptshandwrittenarabic,
      title={Muharaf: Manuscripts of Handwritten Arabic Dataset for Cursive Text Recognition}, 
      author={Mehreen Saeed and Adrian Chan and Anupam Mijar and Joseph Moukarzel and Georges Habchi and Carlos Younes and Amin Elias and Chau-Wai Wong and Akram Khater},
      year={2024},
      eprint={2406.09630},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.09630}, 
}