Tibetan OCR Tools

Overview

This Python project focuses on generating training data for detecting columns or text blocks of Tibetan texts by embedding Tibetan text into images. It provides a flexible and customizable way to create datasets for machine learning models, particularly for YOLO-based object detection tasks.

The project includes functionality for generating synthetic Tibetan text, reading from existing Tibetan corpora, applying various augmentations, and creating YOLO-compatible datasets.

Features

• Automated Dataset Generation: Simplifies the process of creating training and validation datasets for Tibetan text detection.
• Customizable Input: Allows users to specify various parameters like background images, corpora, font, image size, and more.
• Text Generation: Supports both synthetic Tibetan text generation and reading from existing corpora.
• Image Processing: Utilizes PIL for image manipulation and text rendering.
• Augmentation: Includes rotation and noise augmentation strategies.
• Multiprocessing Support: Leverages parallel processing for efficient dataset generation.
• YOLO Compatibility: Generates datasets in a format compatible with YOLO training.

Getting Started

Prerequisites

• Python 3.x
• PIL (Python Imaging Library)
• PyYAML
• Ultralytics YOLO
• Additional Python libraries: numpy, tqdm, etc.

Installation

Clone the repository to your local machine:

git clone https://github.com/nih23/Tibetan-NLP.git
cd Tibetan-NLP

Install required packages:

pip install -r requirements.txt

Usage

Run the main script with desired arguments:

python main.py --train_samples 1000 --val_samples 200 --augmentation rotate

Command-line Arguments

• --background_train: Folder with background images for training (default: './data/background_images_train/')
• --background_val: Folder with background images for validation (default: './data/background_images_val/')
• --dataset_name: Folder for the generated YOLO dataset (default: 'yolo_tibetan/')
• --corpora_folder: Folder with Tibetan corpora (default: './data/corpora/Tibetan Number Words/')
• --train_samples: Number of training samples to generate (default: 100)
• --val_samples: Number of validation samples to generate (default: 100)
• --no_cols: Number of text columns to generate [1-5] (default: 1)
• --font_path: Path to a Tibetan font file (default: 'ext/Microsoft Himalaya.ttf')
• --single_label: Use a single label "tibetan" for all files (flag)
• --debug: Enable debug mode for verbose output (flag)
• --image_size: Size of generated images in pixels (default: 1024)
• --augmentation: Type of augmentation to apply ['rotate', 'noise'] (default: 'noise')

Training with YOLO

After generating the dataset, you can train a YOLO model with Ultralytics using:

yolo detect train data=yolo_tibetan/data.yml epochs=100 imgsz=1024 model=yolov8n.pt

The model is then converted into a torchscript for inference:

yolo detect export model=runs/detect/train9/weights/best.pt 

Inference

We can now employ our trained model for recognition and classification of tibetan text blocks as follows:

yolo predict task=detect model=runs/detect/train9/weights/best.torchscript imgsz=1024 source=data/my_inference_data/*.jpg

The results are then saved to folder runs/detect/predict

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.