Research On The Application Of EEG Brain Network And Lateralization In Emotion Recognition

Emotion is a kind of physiological and psychological reaction of human beings to things based on their subjective experience in a certain environment,which is jointly processed and coordinated by various brain regions.At present,emotion recognition,emotion computing and emotion cognition have become the hotspots of interdisciplinary research in information science,brain science,and artificial intelligence.Electroencephalogram(EEG)signal has become an important research method in emotion recognition due to its advantages of noninvasiveness,real-time difference and not easy to camouflage.Research on emotions based on EEG signals still needs further exploration.For example,there is lateralization in the brain during emotion processing;most studies are often based on the time-frequency domain information of single-channel EEG signals,ignoring the inherent spatial information of EEG signals.This dissertation constructs left and right hemisphere networks in different emotions and analyzes the lateralization phenomenon that exists when the brain processes emotions based on graph theory.Innovatively combines traditional EEG features with functional brain network properties.An emotion recognition model is proposed and a good emotion recognition effect has been achieved.The main research work of this dissertation is as follows:(1)The functional brain networks of left and right hemispheres in different emotions were constructed to explore the mechanism of brain lateralization in emotion processing from the perspective of brain network.Firstly,the EEG signal is preprocessed,including removing the electrooculography and dividing the frequency band.Then,based on the EEG signals of each frequency band,functional hemisphere networks were constructed in high arousal high valence(HAHV),high arousal low pleasure(HALV),low arousal high valence(LAHV),and low arousal low valence(LALV)four emotions.Based on graph theory,we calculated the network topological properties: clustering coefficient,characteristic path length,global efficiency,local efficiency and small-world properties to quantitatively study the topological differences between left and right hemispheric functional brain networks under different emotions.The experimental results showed that the left hemisphere activation was stronger in the HAHV emotional state and the right hemisphere activation was stronger in the HALV emotional state in the gamma band.From the perspective of brain network,this dissertation again confirms that the left and right hemispheres are responsible for processing different types of emotions.(2)Different features are complementary to each other in terms of their ability to characterize emotion.In this dissertation,we propose a method to fuse traditional EEG features with brain network properties and transform them into three-channel EEG images.Traditional EEG analysis methods tend to ignore the inherent spatial information of EEG signals.This dissertation firstly projected the 3D electrode coordinate information to 2D plane using azimuthal equidistant projection,and then transformed the traditional EEG features into threechannel images using cubic interpolation method after obtaining the 2D electrode information,which effectively retains the spatial information of EEG signals.Drawing on the above method,three brain network attributes,namely clustering coefficient,local efficiency and node efficiency,were used to construct three-channel images.The two images with complementary features are stitched and input into the emotion recognition model to effectively extract the spatial information of EEG signals and improve the emotion recognition effect.(3)A CNN-BLSTM fusion of a convolutional neural network(CNN)and a bidirectional long term and short term memory network(BLSTM)is proposed.The CNN-BLSTM model can effectively extract the time domain,frequency domain and spatial information of EEG signals for emotion recognition.The recognition accuracy of arousal degree reached 91% and that of valence degree reached 87% on the DEAP dataset,which is a significant improvement compared with similar studies and validates the effectiveness of the proposed model.