Real-Time Bangla Sign Language Recognition Using Transfer Learning Model

Sign languages are recognizable languages that are expressed by body, face, and hand gestures. The gap that separates us from the world of people with poor linguistic or auditory abilities is filled using sign language. A variety of gestures are done with the hands, fingers, arms, heads, and faces that aid in communication between the deaf and the hearing impaired and vice versa. In Bangladesh and some parts of India, deaf individuals typically interact through Bangla Sign Language (BdSL), usually referred to as Bengali or Bangladeshi Sign Language. The aim of our research is to design architecture by using a transfer learning model to recognize Bangla sign language and form simple words by capturing images consecutively in real-time. For this study, we used a publicly available dataset namely BDSL 49 which has 14475 images that include different images of 2 special characters, 10 numeric characters, and about 37 alphabets to correctly recognize the Bangla sign. In this research, we implemented some transfer learning models such as MobileNet, Xception, ResNet50, and InceptionV3. After 25 epochs Xception performed best out of all the models where the testing accuracy was $\mathbf{9 7. 8 6 \%}$. The testing accuracy of the Xception model proves its adequate ability for real time identification of BdSL sentences.