Research Of EEG Emotion Recognition Algorithm Based On Independent Component Analysis

Human-Computer Interaction(HCI),a method of information exchange between humans and computers,involves researches on image processing,computer vision,and biomedical signal processing.As an effective communication bridge,HCI equipments have been rapidly developed and improved in recent years,and have been widely applied to our daily lives.However,traditional HCI devices generally cannot perform emotional interactions between humans and computers,which is hard to adjust the interaction mode according to the user's emotional state.As a result,their applications will be limited to a large extent.The development of HCI devices with emotional perception has been obtained increasing attention.Nowadays,in order to perceive the emotional state of the user,the sentiment analysis method can be summarized into two major categories,i.e.,biometrics-based method(e.g.,facial expressions,speech signals and gestures)and physiological-based method(e.g.,EEG signals,skin electrical signals,and ECG(electrocardiogram)signals).Among them,EEG-based emotion recognition has become a new research hotspot because of the close relationship between EEG and emotion,and its non-invasive and non-disguised charactistics,as well as easy-to-collect signal.This thesis focuses on EEG emotion recognition based on independent component analysis(ICA)and channel selection methods.The specific works are as follows:(1)Proposed an ICA-based method for extracting emotional EEG spatial features.In the spatial domain filters design phase,the filters were designed using three types of EEG data(positive,neutral,passive).Specifically,we performed ICA analysis on the full-channel data of a single trial,and designed valid spatial domain filters based on the mapping mode and projection position between the independent components and the channels.In the feature extraction stage,the obtained valid spatial domain filter was used to linearly project the emotional EEG trials,then we used singular value decomposition to process the result so as to obtain the final spatial domain features.Furthermore,we fed the extracted spatial domain features into SVM for emotion recognition,and selected the filter with the highest recognition rate as the optimal ICA spatial filter.In the public database(MAHNOB-HCI-TAGGING DATABASE)and the self-established database(including 9 subjects' emotional EEG data),we applied the proposed method to perform emotion recognition on the three types of full-channel emotional EEG data.The average recognition rates obtained were 81.12%and 74.48%,respectively.Compared with the traditional frequency domain features(The logarithm of the power spectrum and the logarithmic Power spectrum of brain asymmetry)were increased by 11.32%and 11.19%,respectively.(2)Proposed an emotional channel selection method.Since the collection of full-channel EEG signals would spend more time and effort,it is not conducive to the development and application of wearable devices.To address this problem,this thesis used the optimal filter designed by the result of full-channel emotion recognition to extract the independent components around these channels,then the independent components were used to calculate the spatial features for emotion recognition.Based on the results,we further calculated the emotional correlation coefficients of the channel and obtained the candidate channel sets according to its values.In this way,the optimal channel set was determined by testing the candidate channel set.In the public database and the self-established database,we carried out two independent channel selection experiments,that is,the optimal channel set for each subject and the common channel set for all subjects.The average recognition rates of optimal channel set on the two databases were 86.91%and 76.52%,respectively,which obtained a releative increasement of 3.91%compared to the full-channel set.Similarly,the average recognition rates of the common channel set were 83.42%and 75.39%,respectively,which was 1.61%higher than the full-channel set.The experimental results reveal that the use of less channel data for EEG emotional recognition can not only ensure the recognition accuracy rate to a certain extent,but also improved it.(3)Developed an EEG emotion recognition experimental platform based on the proposed method under the Matlab platform.The platform mainly consists of three modules:data loading and preprocessing,full-channel emotion recognition,and optimal channel emotion recognition.Among them,the full-channel emotion recognition module was used to analyze and present the relevant results of the full-channel emotion recognition.The optimal channel emotion recognition module was used to select an optimal channel set for subsequent emotion recognition tasks.By testing the data from the self-established database on this experimental platform,the validity and practicability for the EEG-based emotion recognition have been verified.