Research On Emotion Recognition Based On Deep Learning

The emotional state of a person can be reflected through facial expressions,gestures,intonation,etc.,but it is usually difficult to be observed or misled due to human control,thus it is difficult to distinguish its real state.EEG signals directly reflect changes in brain activity and are closely related to changes in emotional state.Based on the emotional EEG data set DEAP and Russell emotional dimension model,this paper carries out signal preprocessing,feature selection and feature dimensionality reduction for emotional EEG features,and then carries out feature classification research based on different dimensions.Through this series of perfect work,the emotional state of EEG is classified and identified.The main research work of this paper is as follows:(1)In this paper,a fractal dimension(FD)method based on Higuchi algorithm and a common spatial pattern(CSP)method are proposed to screen the features.The preprocessed emotional EEG features are analyzed by Higuchi fractal dimension method,that is,the fractal dimension is calculated,and the fractal features are extracted,and then the fractal features are further processed by CSP algorithm to extract different types of features with the greatest difference.(2)A dimension reduction method for kernel principal component analysis((KPCA))based on Gaussian kernel function is proposed.Based on the research of linear discriminant analysis(LDA)and principal component analysis(PCA),the KPCA based on Gaussian kernel function has better dimensionality reduction effect according to the nonlinear characteristics of EEG signals.By comparing the classification accuracy of emotional states under different feature dimensions,it is proved that this method is more efficient in dealing with nonlinear and inter-connected information.(3)The convolution neural network(CNN)is used to train and test the samples.Under the scale of pleasure degree(Valence)and awakening degree(Arousal)based on Russell emotion dimension model,the accuracy of two classification under high and low pleasure degree and awakening emotion is obtained.compared with the traditional machine learning classifier,the classification accuracy of CNN is obviously improved.The third dimension familiarity(Dominance)is mentioned,the experiment proves that under the same conditions,the classification effect based on familiarity is better than that of awakening and pleasure,and corresponds to a higher amplitude of EEG signal in the forehead and parietal occipital region,the classification effect based on familiarity is better than that of Arousal and Valence.So Dominance is a new discriminant dimension for emotional EEG analysis.