ASTNet: Kernel Attention-based Spatiotemporal Convolution Network for EEG Emotion Recognition

Affective computing is a hot topic today. In making a machine more intelligent, the first step is to make it recognize and understand human emotions. In this paper, we propose ASTNet, a spatial-temporal convolution network for analyzing human emotions from electroencephalography (EEG). ASTNet consists of a kernel attention block, a spatial convolutional block, and a temporal dependent block, which sequentially learns the representations of EEG. To learn strong features with temporal and frequency attributes, the first block consists of different convolution kernels whose length is related to the sampling frequency of EEG. Although the spatial resolution of EEG is low, the analysis of spatial representation is indispensable. The second block is used to analyze the emotional representation of the global brain signals in the spatial domain. EEG signals are typical digital time series data. To utilize the brain signals’ temporal representation, the third block is used to improve recognition accuracy. Finally, to evaluate the proposed method, we conduct comparative experiments on the DEAP dataset. The experimental results show that the proposed method outperforms these methods on the valence label.