Emotion Recognition Based On Multi-domain Feature Fusion Of EEG In Key Brain Regions

Brain-Computer Interface(BCI)technology is widely used in the fields of psychiatry,psychology and driving status detection,among which the accurate recognition of emotional states is the key part of the research process.It has been demonstrated that fusing temporalfrequenctial-spatial dimensions of emotion features can produce abundent EEG information,and improving the accuracy of emotion recognition;moreover,exploring brain regions with high correlation with emotion is vital to improve the classification performance.The influence of neuronal activity on emotion generating processes in different brain regions has not been fully considered.Therefore,in order to fuse multidimensional EEG features and explore the influence of different brain regions on the effect of emotion classification,three novel methods are proposed in this paper.which is designed for feature extraction and classification research The specific research work is as follows:(1)First,the connectivity patterns between different brain regions are explored by functionally connected brain networks which is constructed phase-locked value(PLV)of EEG signals in different frequency bands during feature extraction,and the complexity on functionally connected brain networks fusing time-domain and space-domain features were calculated,and the four network feature parameters were fused as features for the dichotomous classification study of emotions,and it was concluded that under this feature extraction approach,the highest correlated brain regions and frequency bands with emotion:the anterior and right temporal lobes of the brain are inextricably linked to the emotional activation process;the highest accuracy of 83.44% was achieved in the dichotomous classification study of positive and negative emotions on the Gamma frequency band.(2)In order to further improve the accuracy of emotion recognition,a new emotion classification feature extraction method based on the construction of PLV was proposed to obtain three distance matrices d F,d S and d LE by three distance measures(Frobenius paradigm,Spectral paradigm and log-Euclidean paradigm,respectively),and to calculate these distance matrices on eight network feature parameters on these distance matrices.The distance matrices and network feature parameters are fed as two features to multiple machine learning classifiers to validate the performance of the proposed method.On the SEED dataset containing three types of emotions,binary classification tasks are performed between every two emotions.d F matrices obtained an average classification accuracy of 83.96% in the Alpha band between positive and neutral emotions,84.12% in the Beta band between positive and negative emotions,and 84.12% in the Delta band between neutral and negative emotions.d F matrices obtained an average classification accuracy of 83.96% in the Alpha band between positive and neutral emotions.The average classification accuracy obtained for the Delta band between neutral and negative emotions was 83.56%.Therefore,the feature extraction method based on the distance matrix can effectively improve the accuracy of emotion classification.(3)To explore the influence of different brain regions on the classification accuracy of emotion recognition,the brain was divided into four regions: frontal,parietal,occipital,and temporal lobes.Then the differential entropy and differential asymmetric features of each brain region on the five frequency bands are calculated separately,and they are coupled as fourdimensional structure of features,then are fed into 4D-CRNN for learning frequency-space domain and time-domain features,respectively,and the tri-classification of emotions is carried out in the classifier.To validate the performance of the proposed model,four model evaluation metrics of accuracy,precision,recall,and specificity were used for metrics.We found that the frontal lobe was the highest correlated brain region with emotions,and the highest accuracy rates were achieved on all five frequency bands(Delta,Theta,Alpha,Beta,Gamma)in this brain region: 84.09%,86.31%,89.43%,92.35%,and 95.52%.In the third part of the study,the key brain regions with the highest correlation with emotions were identified.The proposed classification methods improve the accuracy of emotion recognition,and provides a new perspective to further improve the performance of emotion recognition models.In the third part of the study,the key brain regions with the highest correlation with emotions were identified.The proposed classification method improves the accuracy of emotion recognition to a certain extent and provides a new perspective to further improve the performance of emotion recognition models.