ducto489/ocr_datasets

Combined OCR Dataset for Text Recognition

Dataset Description

This is a large-scale dataset (~11M training, ~0.9M validation images) for Optical Character Recognition (OCR), aggregated from several common benchmarks and sources (see Sources below). It includes scene text, handwritten text, and synthetic images with corresponding text labels.

Training Code: https://github.com/ducto489/lib_ocr

Dataset Structure

./data/
├── train/
│   ├── images/*.jpg  # Training images
│   └── tgt.csv       # Training filenames and labels
└── val/
    ├── images/*.jpg  # Validation images
    └── tgt.csv       # Validation filenames and labels

• tgt.csv: Comma-separated image_name,label. Labels are pre-normalized (see below).

Data Details

• Sources: Aggregated from vietocr, Deep Text Rec. Benchmark, Synth90k, and Cinnamon AI Handwritten.
• Normalization: Text labels have undergone character normalization (e.g., smart quotes to " , various dashes to -, accented chars simplified like Ö->O). This simplifies the vocabulary. See the training repo README for the full list.
• Language: Predominantly English. Vietnamese samples (1.76%) are included but limited.
• Licensing: Combines multiple sources. Users must verify compliance with original dataset licenses before use.

Citation

Please cite the original source datasets if you use this data. Citing this aggregation is also appreciated:

@misc{ducto489_ocr_datasets,
  author       = {Nguyễn Minh Đức},
  title        = {Combined OCR Dataset},
  year         = {2024}, # Update year as needed
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/ducto489/ocr_datasets}
}

Dataset Curators

Aggregated by ducto489 / Nguyễn Minh Đức.