Creating Multiclass Bangladeshi Sign Word Language for Deaf and Hard of Hearing People and Recognizing using Deep CNN Techniques

Abstract

Sign language provides the solution to the communication obstacle for the community of deaf and hard of hearing. This community often faces problems communicating with normal people and among themselves, and sometimes they need an interpreter. But this is problematic and costly. As they are also vital to society, this issue should be solved automatically by recognizing sign language. Computer vision-based sign language recognition (SLR) can mitigate this issue by automatically recognizing sign language. Much research has been done on Bengali Sign language Recognition. Most of them are limited to recognizing Bengali characters and numbers. However, no significant research has been done on Bangladeshi Sign Word Language. This might have happened because of the scarcity of the Bangladeshi Sign Words dataset. In this paper, we have constructed a primary Bangladeshi Sign Language (BdSL) dataset consisting of Bangladeshi sign word images and proposed a Deep Convolutional Neural Network (DCNN) method for recognizing static Bangladeshi sign words. We recognized 11 Bangladeshi daily useful sign words from static images with high accuracy. The proposed deep convolutional neural network model shows an accuracy of 99.12% for the recognition of 11 classes of images of static signs. The model was trained on a dataset size of 1105 images which is constructed from Bangladeshi Sign language users and volunteers. This research work will help to improve the interaction between the D&D community and the general people.