EMG hand gesture classification using handcrafted and deep features

Currently, electromyographic (EMG) signal gesture recognition is performed with devices of many channels. Each channel gives a signal that must be filtered and processed, which sometimes can be a slow process that requires high-cost hardware to process all the data quickly enough. This paper presents a combined feature approach method for EMG classification using handcrafted features obtained from time-spectral discrete analysis and deep features extracted from a convolutional neural network (CNN), which classifies signals recorded from a single channel device. The method proposed only requires 100 signals from each gesture for training, thus the time needed to train the system is reduced. The proposed approach combines handcrafted features from a time-spectral analysis, like mean absolute value (MAV), slope sign changes (SSC), peak frequencies, wavelet transform (WT) coefficients, etc, and deep features to create the feature vector. The feature vector is then classified using a multi-layer perceptron classifier (MLPC). Experimental results showed an average classification accuracy of 81.54%, 88.54%, and 94.19% for 8, 6, and 5 gesture-classes, respectively. The results could serve as a basis for a real implementation of EMG signal gesture recognition with a device of only one channel.