An Initial Study in Historical Javanese Script Recognition Data Synthetic using Neural Style Transfer
Muhammad 'Arif Faizin, Darlis Herumurti · 2024
Recognizing and digitizing Javanese scripts is challenging due to their unique characters and style, which traditional character recognition methods struggle to handle. This study uses neural style transfer with a pretrained VGG-19 model to create synthetic data, for improving character recognition accuracy. High-resolution scans of Javanese documents and printed fonts were used, split into training, validation, and test sets. Following this, a YOLOv9 object detection model was trained to identify and classify Javanese characters within the augmented datasets. The results indicate that the use of style transfer enhances detection performance, achieving a precision of 92.8%, recall of 91.5%, mAP50 of 92.4%, and mAP50-94 of 81.7%, outperforming previous methods. This demonstrates that the inclusion of style transfer and synthetic data generation increase the model's accuracy in recognizing complex historical texts.