Research On Multidimensional Feature Extraction Methods For Emotion Recognition

As a matter of fact,the emotions determine the living status of human.Positive emotions can enhance their well-being,while negative emotions are on the contrary,and even make people suffer from mental diseases and threaten their lives.As the development of brain and cognitive neuroscience,the emotion recognition for electroencephalogram(EEG)signals has gradually become one of the research hotspots in the field of human-computer interaction.How to automatically extract highly differentiated emotional features and improve the performance of emotion recognition are the primary concerns for the researchers.Since the time-frequency analysis can concurrently express the variety of the signal in both time domain and frequency domain,it is a powerful tool for processing the physiological signal.The main contents of this thesis comprise three multidimensional feature extraction methods.(1)A non-linear feature extraction method based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)is proposed.This method combines the time-frequency analysis method CEEMDAN and the non-linear method multi-scale permutation entropy,which further extracts the multi-scale permutation entropy on the basis of CEEMDAN,and takes the multi-scale permutation entropy as the index to evaluate whether the subliminal emotion priming effect appears.The experimental results reveal that the proposed feature extraction method can better judge the subliminal emotion priming effect.(2)A multidimensional feature extraction method based on variational modal decomposition(VMD)and wavelet packet decomposition(WPD)is proposed.This method uses the integrated time-frequency analysis method VMD-WPD to extract the emotional frequency band related to emotions,and further extracts emotional features including modified multi-scale sample entropy(MMSE),wavelet packet entropy,fractal dimension and so on,and feeds them into random forest(RF)classifier for emotion recognition.The experimental results demonstrate that the multidimensional features based on the emotional frequency band can achieve higher emotion recognition accuracy,and this method also has certain advantages compared with other methods.(3)A multidimensional feature extraction method based on Shapelet and VMD-WPD is proposed.Firstly,the shapelet technology is utilized to automatically extract the ERP features,such as N200,P300 and N300,then VMD-WPD is used to extract the emotional frequency band and calculate its non-linear feature MMSE,and finally ERP features and non-linear feature MMSE are fused to form a fusion feature vector and sent it into RF classifier for emotion recognition.The experimental results show that the improvement of the classification accuracy is attributed to the fusion of ERP features and nonlinear features,and the proposed method is also very competitive compared with other methods.