Emotional Music Based EEG Recognition Algarithm

Emotion recognition is a hot topic in human computer interaction, and attractsmore and more researchers from different areas. The development of emotionrecognition technique could make computer communicate with humans in a moreintelligent and human-like way, and improve user experience. Traditional emotionrecognition method is based on human voice and facial expression, and it is not reliabledue to that human can disguise their emotion expression. Emotion recognition based onEEG does not have the shortcoming because EEG is generated from human brain whichdominates generation, perception and expression of emotion. Therefore, EEG canreflect the changes of inner emotion directly.EEG based emotion recognition is a novel field, it involves induction of emotion,feature extraction and classification. Based on the two-dimensional model of emotion,this research selected musical segments with three different emotions—neutral, positiveand negative, and then these stimuli were used for induction of three different kinds ofemotion states.In the research concerning power spectral density, multiple comparison test wasused for selecting some electrodes with discriminative ability. Then, from theseelectrodes, average power was extracted and used to form feature vector of which thenumber of dimensions was reduced by principal component analysis. At last, thedimension reduced feature vectors from5different EEG rhythms were usedrespectively for classification by different classifiers. Results show that features frombeta and gamma rhythms have more power for discriminating different emotion states,and the location distribution of electrodes from which features were extracted hasconsistency among most subjects.In the research regarding brain network, cluster coefficient and path length wereextracted to form feature vector, and used for classification for three emotion states bysupport vector machine. Results show that in9of14subjects classification accuracy isabove70%, and it indicates that brain network attributes could be applied into emotionrecognition. Multiple comparison test shows that brain network attributes from beta and gamma bands have significant difference for different emotional states among mostsubjects, but the subject-averaged network attributes did not show the samephenomenon.