| Human-computer interaction(HCI)refers to the process of information exchange between humans and computers through “special dialogue”.In recent years,with the rapid development of information technology and artificial intelligence,HCI equipment has been comprehensively improved and gradually applied to various fields of human life.Traditional devices can complete some basic functions according to people's wishes,however,they cannot adjust the interaction mode according to the user's psychological feeling,and it is difficult to truly achieve “intelligent interaction”.Therefore,it is of great research significance to make computers have strong emotion recognition capabilities for the further realization of artificial intelligence.The researches on emotion recognition mainly uses the following three methods: facial expressions and sounds,peripheral physiological signals(electromyography,skin electricity,and electrocardiogram)and central nervous physiological signals(electroencephalography).Among them,electroencephalography(EEG)signals have obtained an increasing attention in the field of emotion recognition due to their characteristics such as real-time differences and difficult to disguise.At present,most of the research methods on emotion recognition are based on time/frequency features,but the analysis of EEG spatial features was ignored,and this spatial features may bring some research value to emotion recognition.This thesis focused on three-class(positive,neutral,and negative)EEG-based emotion recognition for further analysis using two filtering techniques,i.e.,common spatial pattern(CSP)and independent component analysis(ICA),respectively.The main research works are as follows:(1)A spatial feature extraction method for EEG emotion-related based on CSP was proposed.Firstly,an experiment was designed to obtain three kinds of emotional EEG signals,and data collection was performed.Secondly,the method based on the two-class CSP was extended to a multi-class method by joint approximation diagonalization(JAD)to extract the spatial features of three emotional states.Finally,an improved JAD(I_JAD)method was proposed to solve the shortcomings of the multi-classification algorithm.Experiments were performed on the self-collection database and the public MAHNOB-HCI database.The average recognition rates obtained by the traditional multi-class CSP method were 74.92% and 81.29%,respectively.The average recognition rates reached 83.04% and 92.70% under the improved I_JAD method.Compared with the traditional method,it increased by 8.12% and 11.41%,respectively.(2)An emotion channel selection strategy based on the improved I_JAD was proposed.In order to solve the problems of operational inconvenience and low recognition performance caused by multi-channel data in experimental collection,a channel selection strategy was designed by analyzing the correlation between each emotion and different channels.Specifically,the weight of each channel for the importance of emotion recognition was defined,and the number and position of channels was optimized according to the size of weight value.The self-collection and public MAHNOB-HCI databases were used to test the proposed channel selection strategy,and the average recognition rates were 86.13% and 94.13%,respectively,which were 3.09% and 1.43% higher than the full channel situation.The experiment showed that using fewer channels for emotion recognition could not only simplify operations in practical applications,but also improved the recognition rate to a certain extent.(3)An EEG emotion recognition algorithm based on ICA was proposed.Firstly,ICA analysis was performed on the EEG data of a single sample to extract the independent components of each channel.Secondly,the full-channels filter was used to perform linear projection on EEG data,and the matrix transformation was used to reduce the feature dimension.Finally,the above method was used to extract emotion-related spatial features for each sample,and the filter corresponding to the sample with the highest recognition result was selected as the optimal spatial filter.The average recognition rates obtained on the self-collected and published MAHNOB-HCI databases were 87.21% and 90.69%,respectively.The results showed that the method was feasible for emotion recognition based on EEG.(4)The location information of “source” corresponding to different emotional states was analyzed.Firstly,a large amount of literature has been investigated on the current neural pattern of emotional “source”,and experimental analysis was performed on the public emotion database SEED based on the above two ICA/CSP filter methods.Secondly,the position information of the “source” corresponding to different emotions was obtained,and it was found that the emotional EEG signals collected by the same subject in different time periods had relatively stable neural patterns.Finally,the differences between the two methods in the application of emotion recognition were discussed from multiple aspects. |
PDF Full Text Request