EEG-based Emotion Recognition Using Multi-Dimensional Convolutional Neural LSTM via Attention Mechanism

Abstract

For emotion recognition, firstly, we transform the differential entropy and power spectral density features extracted from different channels of EEG signals into a spatial representation with a multi-dimensional structure. Secondly, a convolutional neural network (CNN) and a neural network with bidirectional long short-term memory (Bi-LSTM) are combined together to form a deep learning model. Among them, CNN extracts the effective frequency and spatial information in each time segment of the input EEG signal, and the Bi-LSTM strengthens the temporal dependence of the output information from CNN. Further, the attention enhancement mechanism is fused into the Bi-LSTM module to extract more discriminative spatial-temporal features. The proposed model is extensively trained and tested on DEAP dataset to verify its advantages in different aspects. The experimental findings show that the accuracy of emotion recognition is also enhanced to some degree.