An EMG Based Wearable System for Chinese Sign Language Recognition

A sign language recognition system can help individuals who are deaf to communicate with others. Due to its rich information and portable wearability, electromyography (EMG) has emerged as a promising way for such applications. Here, we propose a wearable EMG system designed to recognize 30 Chinese finger-spelling gestures in real time. We introduce a feature extraction and refinement network based on Convolutional Neural Network with Convolutional Block Attention Module (CNN-CBAM) to identify the proper features of EMG. The features extracted by this network demonstrate superior generalization performance compared to traditional EMG feature sets and exhibit notable advantages in multi-participant classification tasks, with a maximum increase in classification accuracy of 12.37%. When integrated with a softmax classifier, it achieves a classification accuracy of 92.32% +/- 1.11%. Additionally, we assessed the real-time performance of our system, yielding a positive predictive value (PPV) of 90.43% during real-time testing.