Hugging Face's logo

Datasets:
mrrtmob
/
km_en_image_line

Tasks:
Image-to-Text
Modalities:
Image
Text
Formats:
parquet
Languages:
Khmer
English
Size:
1M - 10M
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
Community
1
km_en_image_line / README.md
mrrtmob's picture
mrrtmob
Update README.md
a883d42 verified
preview code
|
raw
history blame contribute delete
2.1 kB
metadata 
dataset_info:
  features:
    - name: image
      dtype: image
    - name: text
      dtype: string
  splits:
    - name: train
      num_bytes: 26163114415
      num_examples: 6497555
    - name: val
      num_bytes: 2166258930
      num_examples: 781722
  download_size: 28216796784
  dataset_size: 28329373345
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
license: cc-by-4.0
task_categories:
  - image-to-text
language:
  - km
  - en
size_categories:
  - 1M<n<10M
pretty_name: KM-EN-OCR-Lines

Khmer-English Image Line Dataset 📄

A large-scale synthetic dataset for training OCR models on Khmer and English text. This dataset contains 5 million high-quality synthetic images of text lines.

🎯 Dataset Overview

  • Languages: Khmer, English, and mixed
  • Format: Image-text pairs
  • Use Case: OCR model training

📋 Data Fields

  • image: PIL Image of the text line
  • text: Ground truth text string

💾 Usage

Load with Hugging Face 

from datasets import load_dataset

dataset = load_dataset("mrrtmob/km_en_image_line")

# Access an example
example = dataset['train'][0]
image = example['image']  # PIL Image
text = example['text']    # str

Train with Kiri OCR 

kiri-ocr train \
    --hf-dataset mrrtmob/km_en_image_line \
    --epochs 50 \
    --batch-size 32

🎨 Dataset Features

  • Multiple Khmer and English fonts
  • Light and dark backgrounds
  • Realistic augmentations (noise, blur, rotation)
  • Variable text lengths (5-100 characters)

📚 Citation 

@dataset{km_en_image_line,
  author = {mrrtmob},
  title = {Khmer-English Image Line Dataset},
  year = {2026},
  publisher = {Blizzer},
  howpublished = {\url{https://huggingface.co/datasets/mrrtmob/km_en_image_line}}
}

⚖️ License

CC BY 4.0

🔗 Related

☕ Support