Multi-Scale Convolutional Neural Network with Sinc-Layer and Adaptive Feature Weight Distribution for Motor Imagery EEG Recognition

Electroencephalography (EEG) signal decoding is instrumental in the brain-computer interface (BCI) system. Complex characteristics of EEG pose challenges to the decoding work. In this paper, a multi-scale convolutional neural network with sinc-layer and adaptive feature weight distribution is proposed to decode motor imagery (MI) EEG signals. Sinc-layer can transform kernels of one-dimensional convolution into band-pass filters and it designs the cut-off frequency as learning parameters which can effectively decode the temporal features in the EEG signal. Then a multi-scale convolution module is established, which exploits two one-dimensional convolution and a two-dimensional convolution to extract features from EEG signals. An adaptive feature weight distribution layer is designed to assign weights to features after multi-scale convolution module, aiming to reduce the negative impact of redundant features on recognition accuracy. The results reveal that the method proposed in this paper obtains better results compared with the baseline model in the BCI IV dataset 2b and the dataset collected in our laboratory. Finally, the experimental results show that the proposed method is feasible performs well in the BCI system. Therefore, the proposed method can stimulate the development of human-computer interaction.