Simultaneously exploring multi-scale and asymmetric EEG features for emotion recognition

Abstract

In recent years, emotion recognition based on electroencephalography (EEG) has received growing interest in the brain-computer interaction (BCI) field. The neuroscience researches indicate that the left and right brain hemispheres demonstrate differences under different emotional activities, which is an important principle for designing deep learning (DL) model for emotion recognition. Besides, the neural activities of emotions may occur at different time scales and duration, it is beneficial to use multiple kernels with different sizes to learn the temporal multi-scale features. Based on these two angles, we propose Multi-Scales Bi-hemispheric Asymmetric Model (MSBAM) based on a two-branch convolutional Neural Networks (CNN). We first use different kernel in two separate branch to extract distinct temporal representations, and then bi-hemispheric asymmetric feature extractor to learn asymmetric spatial features in each branch. Then, the features from the two branches are combined together for emotion state recognition. Evaluated on the public DEAP dataset, MSBAM yields accuracies over 98% on two-class classification for both valence and arousal states.