Bangla Sign Language Recognition: Comparing Deep Learning Approaches using a Custom Dataset

Abstract

The rate of disability in Bangladesh is increasing every year and every year these people cry out for recognition. The impaired have been demanding more awareness and education in the field of sign language. Translating signs using motion-based deep learning is an evolving field. Our focus was to assimilate the largest word level dataset and apply that to create a system that could be used in the recognition of Bangla Sign Language. To further develop this, we increased the data and class variations through different techniques. We trained three different models with the dataset to measure the degree of accuracy between the models in detecting the signs. Modifications were also made to one of the models to actuate performance improvements. The comparison between these three models led to an analysis that indicates that the graph based model outperformed the others. With further additions to variety and samples for each class, practical implementation could be made to translate signs in real time to help bridge the communication troubles between us and the hearing/speaking impaired.