Preserving Myanmar’s Ancient Heritage: A Novel Dataset and ResNet Architecture for Pyu Character Recognition

Abstract The ancient Pyu script is a vital cornerstone of Myanmar's cultural heritage, recognized globally by UNESCO. However, preserving and digitizing this historic writing system remains a critical bottleneck due to the complete lack of standardized computational datasets. This paper addresses this digital gap by introducing the first publicly available, systematically curated handwritten dataset of the 33 Pyu consonants, developed in strict accordance with the canonical reference Pyu Reader by Thiripyanchi U Tha Myat. We physically acquired 825 original handwritten character samples (33 consonants x 25 stylistic variations) and digitized them into standardized 224x224 grayscale images with white strokes on a black background. To mitigate data scarcity and simulate historical writing irregularities, we implemented a multi-operator augmentation protocol, incorporating rotation, translation, Gaussian blur, affine, and perspective transforms, expanding the dataset to 4,950 images. The curated dataset has been published openly across GitHub, Hugging Face, and Kaggle to support open-science paleography. To validate the structural consistency of the dataset and establish a benchmark, we evaluated a modified single-channel ResNet-18 architecture. Under controlled testing conditions, the model demonstrated rapid convergence and achieved high classification accuracy on unseen test data. While these results validate the distinct visual characteristics of the dataset, we realistically discuss the limitations of this controlled baseline when transitioning to degraded, noisy physical stone inscriptions (epigraphy). This work establishes a foundational benchmark for Myanmar's computational paleography, paving the way for future extensions into Pyu vowels and real-world transfer learning.