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

Technical Summary: Preserving Myanmar's Ancient Heritage via Pyu Character Recognition

Problem Statement
The ancient Pyu script, a foundational element of Myanmar's cultural heritage and a UNESCO-recognized historical writing system, faces a critical bottleneck in digital preservation. Despite the existence of invaluable epigraphic treasures on stone, clay, and metal from the 1st century BCE to the 9th century CE, the Pyu language lacks standardized computational resources. Specifically, the script is not yet formalized under the Unicode consortium, preventing the use of standard Natural Language Processing (NLP) pipelines and Optical Character Recognition (OCR) engines. Furthermore, there is a complete absence of standardized, labeled computational datasets required to train modern deep learning models. While recent advancements in Burmese script recognition (e.g., myMNIST) have utilized deep learning, these methods cannot be applied to Pyu due to the lack of training data and the distinct, extinct nature of the script.

Methodology
To bridge this digital gap, the study proposes a comprehensive framework for digitizing, augmenting, and classifying Pyu characters through a multi-stage workflow:

1. Dataset Curation and Physical Acquisition:

   • Reference: The study adheres strictly to the canonical tables in Pyu Reader: A History of Pyu Alphabet by Thiripyanchi U Tha Myat [1], covering the 33 Pyu consonants.
   • Acquisition: Due to the lack of digital fonts, the author physically hand-wrote 25 unique stylistic variations for each of the 33 consonants on white paper using black ink. This yielded a baseline of 825 raw images (33 classes × 25 variations).
   • Preprocessing: Images were digitized under controlled lighting, resized to 224 × 224 pixels, converted to single-channel grayscale, and binarized to produce high-contrast white strokes on a black background.

2. Data Augmentation Protocol:

   • To mitigate data scarcity and simulate the physical irregularities of historical inscriptions (e.g., weathering, erosion, and camera perspective), a multi-operator augmentation pipeline was applied.
   • The dataset was expanded six-fold (from 825 to 4,950 images) using five transformations:
     • Rotation: Random angles between -10° and 10°.
     • Translation: Random shifts of ±10 pixels on horizontal and vertical axes.
     • Gaussian Blur: Simulating ink diffusion or stone erosion (σ = 0).
     • Affine Transformation: Simulating stroke speed/pressure variations via shearing and stretching.
     • Perspective Transformation: Simulating non-perpendicular camera angles common in archaeological photography.

3. Model Architecture and Training:

   • Architecture: A modified single-channel ResNet-18 was employed. The first convolutional layer was altered to accept 1-channel grayscale input (224×224×1) instead of 3-channel RGB, and the final fully connected layer was reconfigured to output 33 classes.
   • Training Setup: The dataset was stratified into an 80% training set (3,960 images), 10% validation set (495 images), and 10% test set (495 images). Training utilized the AdamW optimizer, Categorical Cross-Entropy loss, a learning rate of 0.001, and a batch size of 64 over 10 epochs on dual NVIDIA T4 GPUs.

Key Contributions
The paper makes three primary contributions to the field of computational paleography:

1. The First Public Pyu Dataset: The introduction of the first systematically curated, publicly available handwritten dataset of the 33 Pyu consonants, hosted on GitHub, Hugging Face, and Kaggle.
2. Specialized Augmentation Protocol: The definition of a robust data expansion pipeline tailored for low-resource historical languages, specifically designed to simulate the physical degradation and geometric distortions found in epigraphy.
3. Baseline Benchmark: The establishment of a performance benchmark using a modified ResNet-18, validating the visual distinctiveness of the curated script and providing a reference point for future classification studies.

Experimental Results
The modified ResNet-18 model demonstrated rapid and stable convergence:

• Convergence: The model achieved 100% validation accuracy by the third epoch, with validation loss dropping to 0.0081.
• Test Performance: On the completely unseen test set (495 images), the model achieved 100.00% classification accuracy.
• Metrics: Precision, recall, and F1-scores reached 1.00 for all 33 consonant classes. The confusion matrix confirmed that every test sample was correctly classified, indicating that the augmented dataset successfully preserved the structural distinctiveness of the characters.

Significance and Limitations
The paper positions this work as a foundational step toward the digital preservation of Myanmar's ancient heritage. By providing the first standardized dataset, it enables the application of modern computer vision techniques to a previously inaccessible script.

However, the authors maintain a modest and realistic perspective regarding the results. They explicitly state that the 100% accuracy was achieved under controlled conditions where characters were isolated, centered, and high-contrast. The authors caution that this baseline does not yet solve the problem of recognizing degraded, noisy, real-world stone inscriptions found in archaeological sites, which suffer from erosion, cracking, and complex backgrounds. Consequently, this study serves as a necessary "stepping stone" rather than a final solution. Future work is identified as focusing on expanding the dataset to include vowels and diacritics, and applying domain adaptation and transfer learning to bridge the gap between these clean synthetic images and actual field epigraphy.