Research On Gesture Recognition Of EMG Signals Based On Deep Learning

With the development of the tide of artificial intelligence,the application of human-computer interaction technology is gradually widespread,among which the new type of human-computer interaction based on gesture recognition shows superior application value.Surface electromyography has become the most widely used signal source in the field of gesture recognition due to its advantages of being easy to obtain and rich in action information.However,most of the existing gesture recognition methods based on surface electromyographic signals face the problems of complex feature extraction process of surface electromyographic signals and few types of gestures that can be recognized.To address these issues,this paper carried out the following work:(1)We design an ensemble classifier based on the support vector machine model.The basic idea of combining classifiers is to convert a large number of gesture classification problems into a multi-level small number of gesture classification problems.Based on this model,we have increased the number of recognizable actions from about 10 to 40.Furthermore,we demonstrate that the classification accuracy of the combined SVM-based classifier(classification accuracy 72%)is higher than that of a single SVM-based classifier(classification accuracy 63%).(2)To address the laborious and time-consuming feature extraction process of the support vector machine model,we have proposed two deep learning network models capable of automatically extracting surface electromyography signal features.The first model,based on a single network architecture of residual network,enables automatic extraction of the time-domain features of surface electromyography signals.The second model,based on residual network and bidirectional gated recurrent unit,enables automatic extraction of both time-domain and frequency-domain features of surface electromyography signals.We validated the effectiveness of these models using the publicly available Nina Pro DB2 dataset.The results demonstrate superior classification performance of the proposed models compared to many other classification models reported in the literature,with classification accuracy of 80.85% and 84.55% for the residual-network-based gesture recognition model and the dual-stream recognition model,respectively.