Mobile Phone Decoder of Two-Cell Contracted Braille for Filipino Words Using Convolutional Neural Networks

Abstract

This paper presents the results of using MobileNetV2, EfficientNetV1, and EfficientNetV2 for decoding 25 classes of two-cell contracted Filipino braille words. Test accuracies of 91.20% and 89.20%, and F1-scores of 0.91 and 0.89 for the top two models, the EfficientNetV1B0 and EfficientNetV2B0, respectively, were acquired. Transfer learning was used from these CNN models using the weights from ImageNet for pre-training. 1250 images were used for this research, with 50 images per class. For training the models, 70% of the images were allocated, 10% for validation, and 20% for testing. An equal number of samples were allocated for each class in this arrangement of the datasets. The model was implemented in an Android phone application through TensorFlow lite to allow mobile decoding of the braille codes. This output was aimed at creating a reliable and portable platform that will aid Special Education (SPED) teachers in the Philippines in studying and teaching two-cell contracted Filipino braille words.