Two Dimensional Convolutional Neural Network Approach for Real-Time Bangla Sign Language Characters Recognition and Translation

This paper presents a convolutional neural network (CNN) architecture to recognize Bangla Sign Language (BdSL) characters. The proposed CNN architecture also translates the recognized signs into respective textual Bangla characters and provides real-time prediction. The automatic recognition and translation of sign language make communication between dumb and deaf and normal people natural. The recognition structure is validated with a dataset of a total of 4600 selected hand sign images collected from volunteers using an HD Webcam. The proposed structure guarantees the functionality of enhancing the directed dataset by capturing new images of variant persons in real-time. The proposed model recognizes all the 36 letters and 10 digits of the BdSL with significant accuracy. A state-of-the-art result with a validation accuracy of 99.57% and a validation loss of 0.56% has been achieved in the recognition and translation of the BdSL characters.