Abstract
As Filipino data scientists, the proponents developed a real-time Optical Character Recognition (OCR) model that recognizes baybayin script in response to the call for cultural preservation. The model can recognize Baybayin script in real-time with 98.51% accuracy using a mobile camera or webcam. A dataset of 38,000 photos of 19 different handwritten Baybayin characters was utilized to train the deep neural networks. The dataset was resized and data augmentation was performed to prepare the data. A baseline Convolutional Neural Network (CNN), VGG-16, Resnet50, Efficientnet B0, and state-of-the-art CoAtNet were all evaluated. The VGG-16 model obtained the best results, with an accuracy of 98.51%. This is comparable to various previous studies that have an accuracy in the range of 84% – 96.5%, and the other letter classification models trained on the EMNIST dataset, with the highest accuracy of 95.88% for VGG-5 (Spinal FC). It is notable that in this study, the older VGG-16 model was able to outperform the newer CoAtNet model.
Recommendations for further study include improvements using object detection models so the predictions are independent of video framing, expanding the dataset to have words and paragraphs in Baybayin to allow conversion of entire passages, and also to account for different variations and its diacritics. Translation to other languages such as English and Spanish, or evaluating Baybayin alongside latin text can also be explored.
The findings of this study can be utilized to speed up the translation of Baybayin texts, assist linguists and historians in their research, and make Baybayin more accessible to the general public.
