ISL recognition system using integrated mobile-net and transfer learning method.

A sign language recognition system (SLRS) facilitates the flow of information between the signer and non-signer communities. In recent years, deep learning has become ubiquitous and has also been found in the literature of SLRS, but the system designed for real-time recognition is still lacking. This paper presents a small and efficient architecture of Mobilenetv2 integrated with transfer learning for the recognition of Indian sign language (ISL). For this, firstly pre-trained weight file of mobilenetV2 has been used for the feature extraction, and then this model has a global average pooling layer (GAP), followed by a dense layer with a softmax classifier for the classification of ISL gestures. A self-collected ISL dataset consisting of 28 isolated gestures and a publicly available ASL dataset has been used to evaluate the performance. This model???s compact architecture yielded excellent results, with a high classification accuracy of 99.3% and 99.9% for the two datasets, respectively, and low computational complexity. A real-time application prototype using GUI interface has also been designed using the proposed work to translate the gestures of ISL into English. The experimental findings show that the proposed model outperforms the state-of-art method with minimum computational and memory requirements.