web/README.md

metadata

title: OCR Dataset Generator
emoji: 📝
colorFrom: purple
colorTo: blue
sdk: docker
app_file: Dockerfile
app_port: 7860
pinned: false

Synthetic Text Recognition Dataset Generator

Universal OCR Dataset Builder for Low-Resource Languages

A high-performance, client-side OCR dataset generator with a beautiful web interface. Built with Next.js + TypeScript, designed for low-resource languages like Kashmiri, and compatible with all major OCR frameworks.

✨ Features

• Multi-format output: CRNN, TrOCR, PaddleOCR, CSV/JSON, HuggingFace
• Multi-font support: Upload custom fonts with percentage-based distribution
• 25+ augmentation types: Rotation, blur, noise, brightness, skew, and more
• Custom backgrounds: Use preset paper textures or upload your own images
• Unicode-safe: Full RTL support for Arabic, Hebrew, Kashmiri, etc.
• Reproducible: Seeded random generation
• Client-side: All generation happens in your browser - no data uploaded

🚀 Quick Start

Web Interface (Recommended)

Visit the app and:

1. Configure - Set dataset size, image dimensions, upload fonts
2. Preview - See live preview of your configuration
3. Generate - Click start and download your ZIP file
4. Statistics - View generation statistics

Local Development

cd web
npm install
npm run dev

📦 Output Formats

Format	Files Generated
CRNN / PaddleOCR	labels.txt
TrOCR / JSONL	data.jsonl
CSV	data.csv
JSON	data.json
HuggingFace	dataset_dict.json

🎨 Background Styles

• Preset styles: Aged paper, notebook, newspaper, parchment, and more
• Custom images: Upload your own background textures
• Mix mode: Randomly combine multiple backgrounds by percentage

📜 License

MIT