Effects of Data Augmentation on a CNN Model for Baybayin Character Recognition

Abstract

In 2018, the Philippine government approved House Bill 1022 which declared the Baybayin script as the official national writing system of the Philippines. In accordance with the bill’s mandate to promote, protect, preserve, and conserve the Baybayin script, this study takes an in depth look at the effects of data augmentation (DA) techniques on a Convolutional Neural Network (CNN) model for Baybayin character recognition. The dataset was preprocessed to balance the class distribution then we explored geometric and photometric DA methods to observe its effects on model performance using the YOLOv8 algorithm. The DA techniques were set to small uniform intervals when performing the experiments and then key metrics: precision, recall, F1-Score, and mAP are evaluated. The results demonstrate varying impacts of different DA techniques on model performance, with detailed analyses of rotations, shearing, and noise injections. The study contributes to understanding, promoting, and preserving the Baybayin script through machine learning advancements in learning and using the script.
CCS CONCEPTS
• Computing methodologies ➝ Artificial intelligence ➝ Natural language processing ➝ Phonology/morphology.