STA-Net: Deep Spatial-Temporal Attention Network for Emotion Detection using EEG

EEG signals are captured from the cerebral cortex, and neuropsychological studies have confirmed that EEG signals are related to human emotions. To address the problem of undermining temporal and frequency features of EEG signals, we propose a deep learning framework, Deep Spatial-Temporal Attention Network (STA-Net), for detecting human emotions from electroencephalograms. STA-Net utilizes temporal, spatial, and channel attention mechanisms to simultaneously extract discriminative features in the temporal and spatial domains of EEG signals and improve the network’s feature representation by explicitly modeling the interdependencies between channels. This paper also proposes the use of wavelet denoising and sliding window algorithms to preprocess EEG data. The method was experimented on the DEAP database to evaluate the performance of the proposed deep learning network in the classification of low emotional arousal states and high emotional arousal states, achieving a classification accuracy of 93%, which significantly outperforms methods such as SVM and LSTM (p-0.05).