EEG-based Feature Extraction And Optimal Channel Selection Related To Emotion Recognition

Emotions,a complex activity which integrates the feelings,thoughts and behaviors of human beings,are the spontaneously psychological and physiological reactions or stimulated by the external environment.At first,people roughly predicted different emotional states through the obvious external voice,intonation,volume of the sound and so on.With the progress of science and technology,objective physiological signals such as pulse,blood pressure,body temperature,Electroencephalogram(EEG)and Functional Magnetic Resonance Imaging(fMRI),et al.are applied the field of emotional recognition.EEG has been widely used in the field of emotion recognition research for its advantages of low camouflage and high time resolution.The EEG signal,a series of spontaneous and rhythmic electrical activities of cerebral neurons,is acquired by multiple channels distributed in different locations on the surface of the human scalp or cerebral cortex through the connection of channel circuit system and amplification.The recorded EEG signals of human different emotional conditions commonly reflect of the very different electrical activity of rational neurons in the cerebral cortex.Studies have shown that high-frequency components of EEG such as ? and ? bands are better at distinguishing emotions than those lower ones.In addition,brain as a complex non-linear system,it has gradually become the researches central issue in EEG data analysis using non-linear features.For example,recognition of emotional arousal and valance can be achieved by the feature of sample entropy.And applying the feature of Differential Entropy(DE)in the field of emotional recognition,the classification accuracy of machine learning is about 84.22%.It has shown that non-linear features as a category of EEG features possess higher classification efficiency for emotional recognition.However,the higher emotional classification rate,typically by the use of an increasing number of channels and the amount of EEG data,which will go against to the rapidly analysis of emotions.To solve this problem,we find the linear and non-linear EEG features particularly related to emotional recognition by a reasonable algorithm.We calculate the features of each channel of EEG and screening out a few channels that are closely related to emotional recognition through SVM algorithm.Firstly,we designed an emotion-induced paradigm based on pictures stimulus form and then used the EEG acquisition device to record EEG signals.Thirty male healthy participants were invited to watch four categories of emotional pictures(HVLA,LVLA,HVHA,LVHA,on behalf of relaxed,depressed,delightful and fearful,respectively).To obtain a more pure EEG signal,the original EEG Artifact signals were removed by some imperative steps.Secondly,the EEG signals were decomposed to ? rhythm(13-30 Hz)and ? rhythm(30-45 Hz)bands by wavelet transform that belongs to linear analysis method.Two Nonlinear features: Differential Entropy and Information Entropy are extracted from each channel.Finally,EEG features of all channels were screened by F-score algorithm,and the four categories of emotions were classified by Support Vector Machine(SVM).In order to ensure the effectiveness of each type of emotional induction during the experimental design process of this study,the SAM was used to analyze the effectiveness of EEG data when the number of effective scores of each type of emotional pictures reached more than half of the number of emotional pictures of this type.Thus,the 24 subjects who could induce the corresponding emotional categories successfully were screened out,and eventually available brain-electric data was obtained.According to the data of 24 subjects,the average of the F-score of ?-wave,?-wave,Information Entropy and the Differential Entropy were used as evaluation indexes of emotion validity for each channel.The classification accuracy rate of the filtered five channels was 81.1523%,which were FT7,T7,FC4,TP10,O1.And six channels with a classification rate of 83.69% are FT7,T7,FC4,TP10,O1,FP1.The EEG topographic maps of each subject were plotted using Different Entropy and Information Entropy in four emotional states,for illustrating the effectiveness of the features and a small number of channels we selected.Summarize the research results: the transformed F-score algorithm is used to select the combination of features and the small number of channels closely related to emotion,which can greatly shorten calculation time and have great value in realizing real-time on-line recognition of emotion and human-computer interaction techniques.