Efficient Models Based on Deep Learning Technique for Indian Sign Language

Deep and transfer learning opened up new avenues of research and applications in many fields. One of the most significant applications of deep learning algorithms is sign language recognition and it is getting prevalent day by day. Correct sign language interpretation is helpful in reducing the communication barrier between the deaf-mute community and the rest of the society. Although speech is the most common form used to convey or express feelings. However, people with difficulty in speaking or listening use sign language. Communicating with them without knowing sign language is a big challenge. To overcome this problem and to have effective interpretation of sign language to text, the suggested methodology would be helpful. Our work is on hand pose interpretation used in Indian Sign Language(ISL). This research developed four enhanced deep-learning models built on the concept of transfer learning along with landmark points on hands and suggested methodology. It recognises the numeric and alphabetic hand gestures of ISL. We have worked on different models namely Vgg16, Vgg19, Resnet50 and MobileNetV2. To achieve the desired result, we have demonstrated the superior performance of the above-mentioned models with our recommended modified architecture. The performance of the models on each evaluation metrics demonstrates the competent results produced by our proposed methodology. The validation accuracy for Vgg16 is 99.67, Vgg19 is 99.16, ResNet50 is 98.72 and MobileNetV2 is 99.69. Of these, ResNet50 has the lowest accuracy. On the other hand, MobileNetV2 is the model which achieves the highest validation accuracy, lowest validation loss, smallest model size and moderately lesser epochs to get trained among others.