TibetanLayoutAnalyzer

TibetanLayoutAnalyzer is a tool for analyzing and detecting layouts in Tibetan documents. It includes functionality for generating synthetic training data and training an object detector for Tibetan text blocks.

Training data

Synthetic training data for Tibetan number detection is generated with generate_training_data.py of the TibetanOCR submodule. Make sure to update folders for background images.

python ext/TibetanOCR/generate_training_data.py --corpora_folder "data/tibetan numbers/corpora" --background_train "data/tibetan numbers/bg_train" --background_val "data/tibetan numbers/bg_val" --font_path="ext/Microsoft Himalaya.ttf"

⚠️ Make sure to set the path to Microsoft Himalaya.ttf correctly as this font renders Tibetan correctly.

Command-line Arguments

The script supports various command-line arguments to customize the data generation process:

• --background_train: Folder with background images for training (default: './ext/TibetanOCR/data/background_images_train/')
• --background_val: Folder with background images for validation (default: './ext/TibetanOCR/data/background_images_val/')
• --dataset_folder: Folder for the generated YOLO dataset (default: './data/yolo_tibetan/')
• --corpora_folder: Folder with Tibetan tibetan numbers corpora (default: './data/corpora/UVA Tibetan Spoken Corpus/')
• --train_samples: Number of training samples to generate (default: 2)
• --val_samples: Number of validation samples to generate (default: 1)
• --no_cols: Number of text columns to generate [1....5] (default: 1)
• --font_path: Path to a font file that supports Tibetan characters (default: 'ext/Microsoft Himalaya.ttf')
• --single_label: Use a single label "tibetan" for all files instead of using filenames as labels (flag, no value required)

Training the Object Detector

After generating the synthetic data, you can train the YOLOv11 object detector: Training of YOLOv11n is done by a CLI call to Ultralytics.

yolo detect train data=yolo_tibetan/tibetan_yolo.yml epochs=1000 imgsz=1024

Note: The data path needs to match your system setup. The script generate_training_data.py will output the actual data location after generation.

Model Export and Inference

After training, export the model to TorchScript format:

yolo detect export model=runs/detect/train/weights/best.pt format=torchscript

We can now employ our trained model for inference on new images:

yolo predict task=detect model=runs/detect/train/weights/best.torchscript imgsz=1024 source=path/to/your/images/*.jpg

Results will be saved in the runs/detect/predict directory.

Contributing

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

License

tbd.