Optical braille recognition of uncontracted unified English and Filipino braille
Miguel Baliog, Angel Lopez, Maria Franchesca Lopez, Kate Victorino, Joanna Pauline Rivera · 2025
The importance of educational guidance at home has been emphasized since the pandemic. However, for visually impaired individuals, parents and guardians face challenges due to their unfamiliarity with the braille system. As a result, braille modules provided by the teachers are neglected. Leveraging machine learning object detection and classification models, optical braille recognition (OBR) systems could aid in alleviating this communication gap. This study focuses on training an object detection model for an end-to-end OBR. The dataset used is composed of some publicly available datasets, and the braille modules from the Philippine National School for the Blind (PNSB). The PNSB dataset is composed of images taken using a regular mobile phone to simulate the real-world scenario. These are annotated manually due to the absence of transliterated pages. The OBR model is composed of two models: cell segmentation, and cell recognition. These models are trained separately using YOLOv8. The cell segmentation model is trained on the TapVision dataset and the PNSB dataset, while the cell recognition is trained on the AI4SocialGood dataset and PNSB dataset. The discussion delves into the nuances of the performance of cell segmentation, cell recognition, and end-to-end model, highlighting considerations for improvements. Despite challenges, the study underscores the potential of machine learning object segmentation and classification to enhance Braille transliteration, contributing to the broader goal of fostering inclusive education for visually impaired individuals.