ABSTRACT
Motor imagery (MI) classification using electroencephalography (EEG) is crucial to a brain-computer interface (BCI)-based neuro-rehabilitation system. However, due to the complicated non-linear and non-stationary characteristics of MI-EEG signals, discriminating the motor imagery from EEG signals remains a difficult task. In this paper, a new framework combining time-frequency analysis and convolutional neural networks is proposed for MI-EEG signal feature extraction and classification. Firstly, the time-varying autoregressive model-based (TVAR) power spectral density (PSD) estimation approach is utilized to generate a high-resolution time-frequency spectrogram. Then informative MI-EEG representations were obtained through extracting and refining the time-frequency spectrogram in the specific Mu and Beta frequency bands. Finally, an elaborated squeeze-and-excitation convolutional neural network (SECNN) is exploited to learn high-level features and perform motor imagery task classification. The proposed approach achieves competitive results on the publicly available BCI competition II dataset III, indicating that incorporating the novel time-frequency analysis methods and deep learning contributes to improving the recognition accuracy of motor imagery based BCI systems.
