A Comparative Analysis Between Single & Dual-Handed Bangladeshi Sign Language Detection Using CNN Based Approach

Abstract

In order to help the vast majority of the population learn sign language, specialists today frequently apply machine learning techniques. This project seeks to create a model that can recognize the letters in Bangladeshi Sign Language (BdSL) using a deep learning approach. This study is a comparison between the recognition of single-handed and dual-handed Bangladeshi Sign Language. Dataset, KU-BdSL is selected to train the single-handed BdSL and we selected dataset - BDSL 49 for dual-handed BdSL. We suggested pre-trained models of CNN based approach for the purpose of detection and classification. 30 different alphabets from single-handed BdSL and 36 different alphabets from dual-handed BdSL could be classified and recognized by our CNN models. Three pre-trained CNN models were employed. VGG16 outperformed the others by a wide margin. It correctly identified single-handed gestures with 98% accuracy. Regarding the ability to recognize dual-handed Bangladeshi Sign Language, VGG16 once again outperformed competitors with an accuracy of 90%.