THE FIRST OPEN ACCESS DATASET FOR BANGLA SIGN LANGUAGE AND AN ARROWHEAD DETECTION TECHNIQUE WITH CNN MODEL

Abstract

Sign Language is the method of interaction between the hearing-impaired people and the general people. It is the only way to decrease the communication gap of deaf community and the normal people. A machine translator could be potent solution for solving this problem. But collecting hand sign data of sign language from reliable source is too much difficult to researchers. This project is conceived from the above scenario. In this project, we made two open access isolated datasets- Ishara-Bochon and Ishara-Lipi and its recognition model. Ishara-Bochon contains 100 sets of 10 different classes for Bangla Sign Language digits. And Ishara-Lipi contains 50 sets of 36 classes for Bangla Sign Language characters. The image data are collected from different deaf and general volunteers from different institutes. Our datasets could be used to build computer vision based or any other type of system that allows users to search the meaning of BdSL signs. We attempted to represent a BdSL recognizer model which will help hearing impaired people to remove communication gap with generals. In proposed method we used multi-layered Convolutional Neural Network (CNN). CNNs have capability to learn structures automatically from raw data. Our model gained 92% accuracy on our digits datasets and 86% accuracy on our characters dataset. In the future, further AI and data analytics will add values to the services delivered to the end users.