Bengali Sign Language Characters Recognition by Utilizing Transfer Learned Deep Convolutional Neural Network

摘要

Being the fifth most-spoken native language of the world, Bengali is spoken by approximately 265 million people worldwide. In Bangladesh alone, among nearly 130 million native speakers, 13 million individuals are undergoing the hearing impairment problems. Hence, developing a recognition system of the Bengali sign alphabets has been an area of interest for decades. Previously, numerous researchers contributed by either developing benchmark datasets or providing ideas regarding classifiers for accurate recognition. In most of the cases, either the dataset in consideration was not well-constructed or the recognition accuracy was not satisfactory. In this research, we began with a benchmark dataset of Bengali sign alphabets with 38 signs. We applied augmentation and a modified InceptionV3 architecture. The experimental analysis showed that our trained model achieved an overall accuracy of 94.41% which outperformed the previous best-known outcomes by a noteworthy margin.