The SND Headline OCR Dataset is a synthetic Optical Character Recognition (OCR) dataset designed for training and evaluating OCR systems on Sindhi television news headlines.

It consists of Sindhi headline text paired with corresponding RGB images rendered in a style that resembles television news broadcasts. The images imitate common TV news graphics, including breaking news banners, ticker-style headlines, lower-third captions, and headline overlays commonly seen on news channels.

The dataset aims to provide realistic OCR training data for recognizing Sindhi text in broadcast media environments, where text often appears with colored backgrounds, varying font styles, and visually rich layouts.

text_source:
  - "Sindhi news headlines"
  - "Broadcast-style headline text"
  - "Synthetic headline generation"

image_format:
  color_mode: "RGB"
  style: "Television news headline"
  appearance:
    - "Breaking news banners"
    - "Lower-third captions"
    - "News ticker graphics"
    - "TV channel headline overlays"

ground_truth:
  - "Unicode Sindhi headline text"

Limitations

limitations:
  - "The dataset is synthetically generated and may not fully capture the diversity of real television broadcasts."
  - "Graphics and layouts are inspired by common TV news designs but do not replicate any specific broadcaster."
  - "The dataset focuses on headline text rather than complete video frames."
  - "Performance on real broadcast footage may require additional fine-tuning using real-world data."
  - "Visual effects such as motion blur, compression artifacts, and live-video distortions may be underrepresented."

Citation

If you use the SND Headline OCR Dataset in your research, publications, or products, please cite it as:

@dataset{mahdi2026sndheadlineocr,
  author    = {Danish Mahdi},
  title     = {SND Headline OCR Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/DanishMahdi/SND_Headline_OCR}
}

Downloads last month: 13

Total file size:

1.01 GB