Home
Kiri OCR is a lightweight, high-performance Optical Character Recognition (OCR) library designed specifically for English and Khmer documents. It combines modern Deep Learning techniques to provide accurate text detection and recognition in a compact package.
- Hybrid Transformer Architecture: Utilizes a CNN backbone with a Transformer encoder and a dual CTC + Attention decoder for high accuracy.
- Bi-lingual Support: Native support for Khmer and English (including mixed text), handling complex Khmer scripts correctly.
- Full Pipeline: Includes both text detection (finding text in images) and text recognition (reading the text).
- Lightweight & Fast: Optimized for efficient inference, making it suitable for various deployment scenarios from cloud to edge.
- Easy-to-Use: Simple Python API and a powerful Command Line Interface (CLI).
- Customizable: Train your own models or fine-tune existing ones on your data.
- Hugging Face Integration: Seamlessly downloads models and datasets from Hugging Face.
If you are new to Kiri OCR, we recommend starting with the following pages:
- Installation: Learn how to install Kiri OCR on your system (Linux, Mac, Windows).
- Quick Start: Run your first OCR prediction in minutes.
- Python API: Integrate Kiri OCR into your Python applications.
- CLI Reference: Explore the command-line tools.
For those looking to train their own models or dive deeper:
- Training Guide: Learn how to train the recognition model on your own data.
- Detector API: Use the text detection module independently.
- Detector Training: Train a custom text detector (CRAFT/DB).
- Data Generation: Generate synthetic training datasets for Khmer and English.
- Architecture: Understand the inner workings of the Kiri OCR model.
- GitHub Repository: mrrtmob/kiri-ocr
- Hugging Face Space: Live Demo
- PyPI Package: kiri-ocr
- Datasets: mrrtmob/khmer_english_ocr_image_line
We welcome contributions! Please see our GitHub repository for guidelines on how to contribute code, report issues, or suggest features.