Lightweight Sign Language Recognition Through Tiny MobileNet and 3D Multi-Head Soft Attention Mechanism

Sign language plays a very vital role in communication for people having impaired hearing and speech capability. The correct interpretation of the gestures can help bridge the gap in communication between people with hearing impairment and the world. This research describes an efficient tiny MobileNet with 3D multi head soft-attention (TinyNet-3DSA) model to preserve the performance and to reduce the computational resources. The robustness of the proposed approach is assessed on two distinct sign languages: the Massey university dataset which is American Sign Language (MU-ASL), and ArSL2018 dataset which is Arabic Sign Language (ArSL). The proposed approach has achieved standing accuracies up to 100% and 99.4% using only around 96,742 and 94,690 parameters on the MU-ASL and ArSL2018 datasets respectively. This study demonstrated that it is possible to achieve state-of-the-art accuracy with minimal 96,742 and 94,690 parameters on the MU-ASL and ArSL2018 datasets respectively computational overhead, thereby making real-time ASL and ArSL recognition smooth deployment on different edge-based platforms.