The Research Of Emotion Recognition Based On Features Of Brain Networks

Understanding the emotions of others is a precondition for human learning,social cognition and social interaction.How to enable computers to sense human emotions and make appropriate decisions accordingly,also known as affective computing,is becoming an important research direction in the field of artificial intelligence.affective computing has become a global research hotspot in recent years,due in part to the rapid growth of wearable computing devices and the urgent need for a more immersive Human Interface(HCI)environment.There are many ways of affective computing.With the rapid development of current brain imaging technology,more and more attention has been paid to emotion analysis and recognition directly based on brain activity status.However,the characteristic modes of brain under different emotions are explored to improve the accuracy of emotion recognition still a challenging task.Based on DEAP database,this thesis uses the multi-channel EEG(EEG)to construct brain connection network to study the mechanism of emotion processing and emotion classification.The main work includes:1)Firstly,Granger causality is used to construct the EEG brain effective network under different emotional states,and the differences of brain effective network and information interaction under different emotional states are analyzed.The experimental results show that(1)the prefrontal lobe plays an important role in the emotional activity of the brain;(2)the brain information interaction is more active in the negative emotion state;(3)the parietal lobe is more active in the negative emotion,which may has acertain relationship with the involvement of the parietal lobe affording alert mechanism.2)Building emotional brain function network based on EEG-Pearson correlation coefficient.Analysis of functional networks also shows that functional networks under negative emotions have greater degrees and clustering coefficients,which means network connections under negative emotions are more active.Then,the network features were extracted and the machine learning algorithm was used to classify the different emotional states.The classification accuracy was 70.22% on Valence and 62.26% on Arousal.The results show that brain functional networks based on Pearson correlation coefficients have a good effect on characterization and differentiation of different emotional tasks.3)Construction of minimum spanning tree(MST)based on EEG brain function network,and use MST features for emotion analysis and recognition.The use of MST not only solves the threshold selection problem caused by individual differences in brain network analysis,but also greatly reduces the complexity of brain network analysis.By using the graph theory to extract and analyze the related features in MST,we can get:(1)Comparing the MST features of different genders,women are more sensitive to emotional stimuli than men;(2)MST comparison results under different emotions,Negative emotions have a more active structure than positive emotions,which is consistent with previously found findings and experience.(3)The machine learning algorithm based on the characteristics of MST emotion classification,classification accuracy of Valence 73.67%,Arousal 69.84%.MST not only simplifies the complexity of brain network analysis,but also improves the accuracy of emotion recognition to a certain extent.4)Emotion recognition based on EEG nonlinear approximation entropy is studied.A fast approximate entropy algorithm based on matrix operation is proposed.Experimental results show that the improved approximate entropy operation can improve 90.26 times faster than the classical algorithm.Subsequently,in the sentiment recognition using the approximate entropy feature,compared with 65.12% of the classification accuracy obtained in previous studies,the speed of the algorithm is obviously improved without losing the accuracy,which proves the validity of the algorithm.