Datasets:

SoyVitou
/

KHOB-LEVEL2

dataset_info:
  features:
    - name: id
      dtype: string
    - name: image
      dtype: image
    - name: text
      dtype: string
    - name: file_name
      dtype: string
  splits:
    - name: train
      num_bytes: 1460609496
      num_examples: 2213
  download_size: 1611806392
  dataset_size: 1460609496
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

KHOB‑LEVEL2 (Scene Text OCR Subset)

Important Attribution (Full Credit)

This dataset is a cropped/simplified subset of Level 2: Scene Text Images from the original open‑source project:

Khmer OCR Benchmark Dataset — created and maintained by EKYC Solutions and collaborators.

Original GitHub repository: https://github.com/EKYCSolutions/khmer-ocr-benchmark-dataset?tab=readme-ov-file
Official full dataset download (Google Drive): https://drive.google.com/drive/folders/1YczKEIha64Qpjk9Wsbp0-r7PFoYrgvga

All credit goes to EKYC Solutions and the original authors. This Hugging Face version is only shared as a convenient subset for quick OCR testing.

Background

In Cambodia, Khmer Optical Character Recognition (OCR) is an active area of AI development. However, evaluating Khmer OCR engines is difficult because:

Many OCR systems benchmark only on internal datasets (biased results).
Some benchmarks rely on synthetic computer‑generated datasets, which do not reflect real‑world images.

The Khmer OCR Benchmark Dataset addresses this by providing a standardized benchmark with increasing difficulty levels.

Dataset Description

This subset is derived from:

Level 2: Scene Text Images

Level 2 focuses on Khmer printed text appearing in real‑world scenes. Images were captured using mobile devices (e.g., iPhone 13) and include:

Curved Khmer printed text
Occlusions and noise
Reflections
Varied lighting (bright and low light)
Real‑world backgrounds

What does Level 2 evaluate?

OCR models that perform well here can recognize Khmer printed text in uncontrolled real‑world environments.

Dataset Format

This subset is provided in a simple structure that maps each cropped image to a single text label.

Column	Description
`image`	Scene text image (Level 2 subset)
`text`	Ground‑truth transcription
`file_name`	Original image filename
`id`	Unique identifier

Example entry:

{
  "image": "LS_10_img_01.png",
  "text": "ហាមចោលសំរាម"
}

Usage Example

from datasets import load_dataset

# Replace with your dataset repo ID on Hugging Face
ds = load_dataset("SoyVitou/KHOB-LEVEL2")

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

Original Source & License

This subset is released in the same open‑source spirit as the original benchmark. For the complete dataset, license details, and full acknowledgments, please refer to the official repository:

https://github.com/EKYCSolutions/khmer-ocr-benchmark-dataset

Acknowledgment

This benchmark dataset was initiated in collaboration with institutions in Cambodia, including:

EKYC Solutions
Prudential Life Assurance PLC
Paragon International University

We are grateful for their contribution to improving Khmer OCR research.

Note: This Hugging Face dataset is only a small cropped subset for convenience. For serious benchmarking, please download and use the complete dataset from the original authors.

Next Steps

If helpful, we can extend this dataset card with:

A BibTeX citation
Additional tags (e.g., ocr, scene-text, khmer)
Train/test split description and data curation notes