Research On Feature Representation Of EEG Signals In Emotion Recognition

Emotion recognition is the core technology of Artificial Intelligence and Human-Computer Interaction. In Information Age, it’s more necessary for computer to understand and express human emotions. In reality, emotion recognition has been applied to various fields, such as medical treatment, education and business. Because emotion is a fairly complex cognitive process, so for better recognition result, continuous and deep studies should be conducted.EEG signal is a kind of electro-physiological signals, directly reflecting brain’s activities, which is widely applied to emotion recognition due to its precision and objectivity. This paper studied emotion recognition based on EEG signal, focus on emotion related EEG features extraction and representation. In this paper, three jobs were mainly conducted as followed.1) Emotion arousal. Emotion arousal is the key premise of emotion recognition research, affecting the availability and accuracy of the EEG data. This paper used widely accepted CAPS (Chinese Affective Picture System) and IAPS (International Affective Picture System) as stimuli to design arousal file. Three kinds of emotions namely the positive, the negative and the neutral were induced in the way that same kind pictures were continuously presented to subjects. Continuous stimuli enhanced subjects’ emotion feeling, so as to acquire higher recognition result.2) EEG time-frequency feature extraction. This paper extracted three time-frequency features using wavelet. Three features are subbands’ energy, energy rate and root mean square of wavelet coefficient and reflected emotion related activity degree of EEG. SVM acquired average recognition accuracy 82.87% which showed that these features were efficient in emotion recognition and. Compared with IAPS, EEG signals collected with CAPS acquired higher recognition accuracy. This showed that emotions exist differences in different cultural background. Besides, considering EEG signals’ time varying, nonlinear feature and the process of emotion, we clipped signal samples with overlap. The recognition result of SVM proved the advantage of this method.3) Emotion related EEG regions partition. More and more studies focus on the importance of specific brain regions’ emotion feature in recognition. But brain regions partition in most studies was based on distance and symmetry, which ignored the correlation and diversity between electrodes. This paper proposed an EEG region partition method based on EEG electrode similarity cluster and the cluster center represented all the electrodes’features in cluster. In this method, electrodes’correlation coefficients were computed with time-frequency features to represented similarity. Two electrodes were labeled as connected when their similarity were highest and higher than threshold. Connected electrodes were clustered into a class, as so as a partition. Cluster center represented all electrodes in this cluster, which eliminated the redundant and reduced the dimensionality of EEG data. Compared with general EEG partition methods based distance, this acquired higher emotion accuracy.