Emotion Recognition Research Based On Multiscale Entropy

As human-computer interaction and artificial intelligence are developing rapidly in recent years,the research of emotional computing is attracting more and more attention.Emotion recognition is the main content of affective computing,and emotion recognition based on physiological signal is widely accepted.EEG signal is a comprehensive reflection of the electrical activity of brain neurons,and it's able to reflect the brain respond in different cognitive tasks and functional status.Since EEG cannot be deliberately hidden,the EEG based emotion recognition research as a reflection of emotion state is authentic,objective and scientific.The main features of EEG signals are nonlinearity,multi-scale and multi-resolution,and multiscale entropy features can reflect the characteristics of EEG sufficiently.In this paper,permutation entropy(PE),Approximate entropy(ApEn),and sample entropy(SampEn)are applied as entropy features.The four multiscalization methods mentioned are empirical mode decomposition(EMD),discrete wavelet decomposition(DWT),coarse-grained method(CG)and moving-average method(MA).The particle swarm parameter optimized support vector machine is used for emotion recognition.In the identification of three emotional states of positive,neutral and negative,the performances of ApEn,SampEn and PE are firstly compared with those of multiscale ApEn,multiscale SampEn and multiscale PE.It shows that the highest classification accuracy of multiscale entropy feature extraction in the test set is 94.33%,significantly superior to single entropy algorithms.12 multiscale entropy methods are adopted to contrast the emotion classification ability of encephalic regions in order to discuss their sensitivity.O2 proves to possess the highest classification accuracy among all the 16 channels.With CG_PE feature extraction algorithm,its accuracy in the test set reaches as much as 82.86%.The emotion recognition results of different encephalic regions and the two cerebral hemispheres are respectively compared.The results from the encephalic regions show that the test set accuracy of the occipital region based on MA_PE feature extraction has reached 77.05%.As for the two cerebral hemispheres,the result indicates a slightly higher accuracy in the test set of the right hemisphere,and the maximal difference value is 8.61% with MA_PE method.The few exceptions present to both the training set and the test set of DWT_SampEn,as well as the training set of EMD_SampEn.In general,emotional states of the right hemisphere show higher distinction,which reveals that the right hemisphere is more sensitive.In conclusion,multiscale entropy algorithm can extract EEG's multiscale features sufficiently,and is able to implement emotion recognition and emotional sensitive encephalic region analysis.It is an effective algorithm for EEG feature extraction.