openpecha/uchen-ume-classification-benchmark

Uchen-Ume Classification Training Dataset

This repository contains the primary training and evaluation data for the Uchen-Ume Classifier. This dataset is used to train models to distinguish between formal "headed" block scripts (Uchen) and cursive "headless" script families (Ume).

Repository Structure

The dataset is organized into three main directories to support different training and validation stages:

• source/: Contains the raw, original manuscript images as provided by BDRC.
• preprocessed/: Contains three variants of the data used for different experiments:
  • whole_page: Images resized (short-edge 224px) and center-cropped to 224x224.
  • patches_color: Images split into 5 to 6 patches per page with a 10% overlap.
  • patches_clahe: Same 10% overlap patch layout, but with Contrast Limited Adaptive Histogram Equalization (CLAHE) applied to improve stroke visibility in low-contrast scans.
• benchmark/: A "Gold Standard" evaluation set containing exactly 5 images per class across 18 script types. These images were selected after a strict taxonomical cleanup and were never seen by the model during the training, validation, or testing phases.

Project Context

• Project: The BDRC Etext Corpus
• Developed by: Dharmaduta in collaboration with BDRC
• Source of Images: Buddhist Digital Resource Center (BDRC)
• Funding: Khyentse Foundation
• License: CC0 (Public Domain)

Dataset Stats & Splits

The following numbers represent the data used for the main Uchen-Ume binary classification task:

• Train: 3,048 images
• Validation: 762 images
• Test: 762 images
• Total Samples: 4,572

Balanced Class Distribution

• Uchen (Class 0): 2,286 images (includes uchen_sugdring, uchen_sugthung)
• Ume (Class 1): 2,286 images (includes petsuk, peri, tsegdrig, khyuyig, etc.)

Selection & Tagging Methodology

Selection Process

Images were selected from the BDRC digital library to provide a comprehensive representation of Tibetan script geometry. To maintain a clean binary signal, we manually excluded pages that were "Difficult," "Multi-script," or "Non-Tibetan" from this specific training set.

Annotation

Tagging was performed using a custom-developed script classification tool developed by Dharmaduta and BDRC.

• Expert Review: A team of expert Tibetan paleographers reviewed every scan to assign ground-truth labels.
• Taxonomical Cleanup: The benchmark folder specifically reflects a cleaned taxonomy, ensuring that the 18 sub-classes are correctly represented for the most rigorous model testing possible.

Acknowledgments

We thank the Buddhist Digital Resource Center for their archive and the Khyentse Foundation for funding The BDRC Etext Corpus project.