Eeg Emotion Recognition Based on Efficient-Capsule Network with Convolutional Attention

Abstract

EEG-based emotion recognition, as a pivotal component in human-computer interaction, has garnered considerable scholarly interest. And finding EEG features with stronger time-space-frequency correlation as well as reducing the computational overhead while enhancing the model structure have been the focus of research in this field. Therefore, in this paper, a deep learning model, ECNCA, is optimized. Firstly, by mining the temporal, frequency and spatial features in the EEG data, four frequency bands, namely theta, alpha, beta and gamma, are spliced and fused to make full use of the information in the EEG data for emotion classification. Secondly, the input data was strengthened by CNN and attention mechanism, and Efficient-Capsule was used to complete the classification of emotions as a way to achieve the goal of high accuracy with low overhead. Finally, we conducted various experiments on the SEED dataset and DEAP dataset, and the highest accuracy of the emotion triple classification task and quadruple classification were 94.96% and 91.34%, respectively. In addition, the experiments demonstrate that ECNCA also has an advantage over CNN and CapsNet in terms of computational overhead. This study can provide some reference for emotional experience and emotional computing.

Keywords: Deep learning, Electroencephalogram (EEG), Emotion recognition, Efficient-Capsule network, Attention mechanisms.