Research On Emotion Recognition Based On EEG

How to get effective emotional signals, interpret individual’s emotion objectively and accurately, moreover, to make the information system have better perception and decision ability, are the focus in the field of information and intelligence. Since emotion is often accompanied by higher cognitive brain activity, objective and accurate recognition of emotional state has been a thorny problem for researchers. In recent years, the development of modern neuroimaging techniques has built a bridge between the subjective world and the objective world. Among them, the Electroencephalography(EEG) has become the principal tool for prospecting brain function and designing brain-computer interface due to its high time resolution, portability and practicality. Research on emotion recognition based on EEG is of great significance in disease treatment, brain-computer interface and information evaluation.In order to meet the requirements of EEG-based emotion recognition technology in terms of stability, accuracy and practicality, this thesis mainly focuses on three aspects of investigation,which are automatic artifact removal methods, feature selection methods and channel selection assisted by functional Magnetic Resonance Imaging(fMRI).1. Research on the online automatic artifact removal methods for EEG. EEG signals are highly time-varying sensitivity. In addition to the noise of the system itself, EEG signals are always accompanied by electrooculography(EOG) and electromyography(EMG), thus the signal noise ratio(SNR) is quite low. Without additional separate reference electrodes, the traditional methods based on machine learning require a large number of artifact samples collected offline. Moreover, the offline trained classifier may not achieve the desired results for mismatching another single data acquisition. Aiming at this problem, this thesis presents an automatic online artifact removal method based on prior artifact information. The combination of discrete wavelet transform and independent component analysis(ICA), namely Wavelet-ICA, is applied to separate artifact components. The artifact components are then automatically identified using prior artifact information, which is acquired online in advance. Subsequently, signal reconstruction without artifact components is performed to obtain artifact-free signals. The results show that using the proposed automatic online artifact removal method, there are statistical significant improvements of the classification accuracies in both two experiments, namely, motor imagery and emotion recognition. This illustrates that this method enhances the EEG data’s ability in cognitive information comprehension.2. Research on the feature selection method in EEG-based emotion recognition. The large amount of the time-frequency features in the EEG signals reflects the different cognitive statuses of our human brain, and is the foundation for emotion recognition. How to select emotion related features among high dimensional features, is the key to realize rapid, accurate and effective emotion recognition. This thesis presents a feature selection method based on sparse learning. Using the regularized orthogonal matching pursuit(ROPM) algorithm, it can quickly and efficiently find the emotion-related features with small dimension by seeking the features that have biggest contribution to reconstructing category information. Experimental results show that, compared with the traditional method, this method can reduce the time consumption and obtain higher classification accuracies in the same dimension of features. In addition, based on the feature selection, the distribution of emotion-related features in different time periods and different frequency bands has been analyzed. This may provide certain reference for the research of the emotion mechanism3. Research on the channel selection assisted by fMRI. Brain activities involving emotion can be interpreted as the specific activity of local area in the brain cortex. The spatial distribution of emotional features in the brain is of great significance for selecting EEG channel and increasing the applicability of online emotion recognition. This thesis proposes an EEG channel selection method based on the EEG forward model. Firstly, according to the structural MRI of the brain, EEG forward model is established to obtain the transfer matrix from the cerebral cortex to the scalp. This makes it possible to map the activation of emotional fMRI experiments to the scalp and obtain an EEG topographic map reflecting the correlation between electrodes and emotion, which can provide the basis for optimizing electrodes used in EEG-based emotion recognition. Experimental data showed that, according to the EEG topographic map reflecting the correlation between electrodes coming from the fMRI activation, selected EEG channels can retain emotion-related EEG signals to the greatest degree. This method also provides a new idea for the research on EEG assisted by fMRI.