A Novel DCNN Based MI-EEG Classification Method Using Spatio-Frequency Information

Abstract

Deep convolutional neural network (DCNN) has been successfully applied to improve the classification performance of motor imagery (MI) tasks in the electroencephalogram (EEG) based brain computer interface (BCI). However, there is still a great challenge about how to extract effective features from EEG signals. So, a novel parallel DCNN structure-based MI-EEG classification method is proposed in this paper. Fast Fourier transform is utilized to transform EEG signals from time domain to frequency domain. Then, μ and β rhythms, which are related to MI tasks, are redivided into three sub-bands. The averaged power is calculated for each sub-band. In order to utilize the position information of electrodes, azimuthal equidistant projection is adopted to convert the 3-D location of each electrode into 2-D position. Then, Clough-Tocher interpolation algorithm is employed to generate a spatio-frequency image for each sub-band. Furthermore, a parallel DCNN (pDCNN), which includes three DCNNs correspondings to three sub-bands respectively, is designed to extract and classify spatio-frequency features. The BCI2000 dataset is adopted, and the average accuracy and Kappa value reach 90.25% and 0.81 respectively, which are greater than that of the existing methods. These experimental results show that the effectiveness of the proposed MI-EEG classification method.