Research On Cross-subject Emotion Recognition Method Based On MI-SFFS Output Probabilit

Due to the diversity of human characteristics,there are significant differences between different human individuals.In addition,due to the poor generalization ability of the features representing human emotions,the accuracy of cross-subject emotion recognition is not high enough.Compared with the methods of emotion recognition using facial,voice or other physiological signals,EEG can avoid the phenomenon that users hide their emotions subjectively.Using EEG can improve the accuracy of cross-subject emotion recognition to a certain extent.In order to truly detect the user’s emotions,improve the generalization ability of the model and improve the classification accuracy,this paper adopts the output probability method based on MI-SFFS to carry out crosssubject EEG emotion recognition research.First,this thesis applies the officially preprocessed DEAP and extracts EEG signals for experiments.In this paper,MNE tool is used to remove electrical ocular artifact and then we compare the processed data with the original data to verify whether the official pretreatment operation is sufficient.Then,the generalization ability of the model is improved through data slicing.According to the physiological mechanism of human brain and the experience of previous researchers,part of the brain electrical channels are selected for the subsequent experimental study.Time domain,frequency domain and nonlinear domain are extracted to characterize the information well contained in EEG signal.Then,a single feature selection method or a combination of two feature selection methods are used for comparison,and then the output of objective evaluation indexes of classifiers are compared and analyzed.Finally,the MI-SFFS(Mutual InformationSequence Float Forward Selection)feature selection algorithm that balance speed and number of features is selected.The retained features after feature selection are sent to SVM,KNN and RF classifiers for classification.The probabilities of the classification results of three classifiers are listed respectively,and the output probabilities of different classifiers are added together to obtain multiple probability features.After combining the probability features and high-dimensional features of different groups,the three classifiers are used for re-classification,and finally the classification accuracy of crosssubject emotion recognition reached 0.7361.Finally,a software platform for cross-subject emotion recognition based on MISFFS output probability was designed and built based on Qt,and the user interface of the system was designed.A variety of functions are realized in the system platform,such as user login,user identity switching,data uploading,performance testing,mood analysis and feature selection and classification.In addition,the user help and feedback function is provided for the convenience of users.