Surface EMG Decoding for Hand Gestures Based on Spectrogram and CNN-LSTM

Forearm surface electromyography (sEMG) classification for hand movements is a trending research field in several real-life scenarios. The classification, also known as the decoding, is helpful to building dexterous prostheses and dexterous exoskeleton for soldiers. According to previous studies, the classification accuracy is subject to the features extracted from the source signals. The traditional features are usually very carefully designed physical signals. In this paper, we demonstrate that features with specific physical meaning like spectrogram are less effective than a combination of such features and neural networks. Tests are performed on Ninapro database which contains 40 subjects' sEMG data sampled by 12-channel surface electrodes for 50 different movements including finger/wrist gestures and force exertion. Our methods combines the spectrogram, the CNN and the LSTM to fully use the spacial local physical information and sequence's time information. The results show improved classification accuracy (from 75.740% to 80.929% for the basic hand gestures and an overall improvement from 77.167% to 79.329%).