Feasibility Analysis of sEMG Recognition via Channel-Wise Transformer

Surface Electromyography (sEMG) features as objective biomarkers have the potential to facilitate a refined diagnosis, appropriate treatment/rehabilitation, and measurements of Human-Computer Interfaces (HCI). Precisely recognizing the informative features in EMG data is a staple challenge due to its massive, noise-involved, and nonstationary nature. In this paper, we build a channel-wise method assembling the recent powerful Transformer model for recognizing muscle activation. The gesture classification benchmark is demonstrated for proof-of-concept by introducing Ninapro sEMG datasets with the 17 gestures recorded from 40 subjects. From experiment results, our proposed method achieves a competitive overall accuracy of 85.6% by leveraging the sole channel-Transformer.