| Emotion is essential for people’s daily life. Accurately identifying and analyzinghuman emotional states are not only desirable in solving some mental illnesses, butalso are widely used in human-computer interaction, distance education, medicalfields, etc. Compared to other information sources that could be used for emotionrecognition, EEG has advantages of sensitive to emotion, and not prone to beconcealed. In addition, EEG-based emotion recognition is more objective. Therefore,EEG-based emotion recognition has attracted considerable attention and become verypopular. In traditional researches, emotional states are estimated by analyzing theEEG responding to specific emotional stimuli. Meanwhile, the recognition scope andcorrect rate are not optimal and need to be further improved. This method also has alimitation: the recognition model lacks of strong generalization ability and cannotensure successful application to distinguish the emotional state with other stimulimodel. The above-mentioned have become a difficult technique problem to beurgently solved.In order to solve the above problems, a TUNERL afective material database thatcontains six categories of emotions (40videos,100pictures) was established. Thentwo emotion elicitation protocols consisting of video sequences and picture stimuliwere designed to arouse5emotions (joy, calm, sadness, disgust and tension), andcollect the EEG signals of12subjects. Firstly, power spectral density was calculatedto analyze the typical sub-band energy and asymmetry indexes. The Fisherdiscriminant ratio and one-way ANOVA analysis showed that the discriminativefeatures were mainly concentrated in the beta and gamma band, discriminativefeatures were primarily derived from electrodes placed near the prefrontal, occipitaland temporal lobes. Then, the sample entropy was analyzed. The notable result is thatthe EEG evoked by emotional video is more complex compared to the calm state,especially in the temporal lobes, which indicates a strengthening in active neuronalprocess during emotion. At last, the user dependent classification performances wereinvestigated by using SVM based on recursive feature elimination, with differentclasses corresponding to the basic and the valence dimensions model. The impact ofthe source material comparatively on the recognition results was analyzed through theadjustment of the training strategy, in the program of the EEG evoked by video. Results indicated that the power spectrum energy features performance was superiorto asymmetry indexes, and the correct rate was higher with valence than with thebasic emotion labels. After the feature selection, results have been improvedsignificantly. When classifying the EEG induced by video and pictures, the topaverage classification accuracy rate on valence was as high as80.42%(for EEGevoked by video) and85.55%(for EEG evoked by pictures). The result in turnrevealed the effectiveness of the proposed feature extraction methods. Furthermore,this thesis went further into the cross-stimulation model emotion recognition bySVM-RFE, the classification accuracy rate up to68.2%, confirming the feasibility ofcross-stimulation model emotion recognition.The achievements of this thesis will help to provide technical support andtheoretical guidance to establish an independent of induced mode, close to the actualapplications, robust and reliable emotion recognition model, which are expected tofurther applied in the field of emotion recognition application. |
PDF Full Text Request