EEG Emotion Recognition Based on TQWT-features and Multiscale Convolution Superimposed Transformer

Emotional recognition can help machines to capture the emotional state of human beings, which is an indispensable technology for human-computer interaction and the realization of artificial intelligence. Electroencephalogram (EEG) has been applied to emotion recognition due to its good temporal and spatial resolution. How to capture the correlation of long-distance time-series EEG signal data from both temporal and spatial dimensions and directly affect the effect of emotion recognition is a difficulty in the BCI research. Transformer model can extract long-term correlations of time-series signals. In this paper, a new electroencephalogram emotion recognition method based on spatiotemporal fusion Vision Transformer is proposed. The feature of EEG signals is reconstructed by the tunable Q-factor wavelet transform (TQWT) method. The spatiotemporal correlation between the acquisition channel and the temporal signal is established by position coding for the convolved features. The hybrid model of multiscale convolutional and Vision Transformer (MC-ViT) is utilized for feature extraction and emotion classification. The EEG emotion recognition method based on spatio-temporal fusion vision transformer proposed in this paper has achieved good EEG emotion classification effect on the SEED dataset, which outperforms state-of-the-art deep learning methods.