Deep Convolutional Neural Network-Based Bangla Sign Language Detection on a Novel Dataset

Abstract

Bangla Sign Language (BdSL) is the communication medium used by the deaf and mute people of Bangladesh. As they have speaking and hearing disabilities, they communicate with each other and with the rest of the world through sign language. It works as a communication medium and decreases the communication gap with normal people. In this research, we have worked with Bangla Sign Language detection for digit and letter signs. We have concentrated on building a complete BdSL dataset and applying various classifiers to our novel dataset. We have mostly focused on building convolutional neural networks to classify image data in this article. In addition, we have applied K-nearest neighbor (KNN), random forest (RF), support vector machine (SVM) and decision tree (DT) models to check the performance of each model on our dataset. On average, we have observed the highest accuracy of 95% for digit recognition and 91.5% for letter recognition on the CNN model. We have used 1500 images for digit and 4320 images for the letter dataset. We believe that our dataset will help researchers to carry out more in-depth studies as we have plans to make it openly accessible. Therefore, it will reduce the scarcity of the Bangla sign language dataset and expand research in this area.