Research On Emotion Processing And Recognition Technology Based On EEG

Emotions are an important part of the psychological structure of humans.They are indispensable in people's psychological lives.With the development of human-computer interaction(HCI)technology,emotion perception and computing creat a harmonious human-machine environment by making the computer have the ability to perceive,recognize,understand,express and adapt to human emotions,so that the computer has higher and comprehensive intelligence,which is the naturalization and intelligentization of HCI.Electroencephalogram(EEG),as an electrophysiological indicator that can record brain activity,has a millisecond time resolution that matches the speed of brain information processing,and has the characteristics of good portability and non-invasive.It has become an important measurement method for the research of emotion processing and recognition application in the field of brain cognitive science.Exploring the neural mechanism of emotion processing and developing technology for automatic emotion detection and recognition have become very important researches in the fields of artificial intelligence.This thesis focuses on the scientific problem of ‘how to characterize the neural activities of different emotions through EEG signals',and conducts research through external induced emotions and internal induced emotions.For the external induced emotions,three basic issues are explored,including the information mechanism in emotion processing,the key information representation of emotions,and the design of emotion classification models based on neural networks.For internal induced emotions,from the perspective of practical application,this thesis explores whether the EEG signal characteristics have specific relevance and differences with external induced emotions,and constructs an EEG computing model for internal induced emotions.The main research are as follows:1.Research on the processing mechanism of external induced emotions based on steady-state visual evoked potential(SSVEP).How to use EEG signal to explore the information processing of different emotions is the basis and difficulty problem to be solved in emotion perception and computing.In this thesis,based on SSVEP with good phase-locking characteristics,we reconstruct the time series of cerebral cortex with high spatial-temporal resolution by source imaging method,and then study the information processing mechanism of different emotions in the cortex space in depth from the local and global levels.The results reveal that:(1)affective perception is accompanied by the acceleration of information processing speed in the ventral pathway.Unpleasant emotions has the fastest information processing speed in the ventral stream,including the middle occipital and the middle temporal gyruses.The processing speed of the right hemisphere is faster than the left hemisphere,and the direction is from the right hemisphere to the left hemisphere.(2)Positive emotions show a unilateral effect of the left hemisphere,and the left frontal upper gyrus is a key brain area that distinguishes positive emotions from neutral and negative.This study provides an important theoretical basis for the existing emotion research and a theoretical basis for the perception and computing of emotions in HCI.2.Research on external induced emotion recognition method based on empirical mode decomposition(EMD).For the recognition of external induced emotions,one key problem is to find information representation to analyze and model emotion.Among the existing feature extraction methods,the differential entropy feature with the best performance essentially describes the band energy characteristics of EEG signal,but its time resolution is still limited.In response to this problem,an emotion recognition method based on EMD is proposed in this thesis: firstly,EMD is used to adaptively decompose EEG into intrinsic mode functions(IMFs)with different oscillation frequencies,and then multi-dimensional information such as waveform difference,phase difference and normalized energy are extracted in IMF space as representations for emotion recognition.Compared with the differential entropy method,the proposed method has higher time and frequency resolutions.The experimental results on DEAP show that the proposed method improves the accuracy of emotion recognition significantly.The accuracy is 72.10% for binary classification.Compared with fractal dimension method,the accuracy is improved by 13.97%;compared with DE method,the accuracy is improved by 2.90%.This study provides a new information representation of emotions,which is of great significance to improve the recognition accuracy of external induced emotion.3.Research on external induced emotion recognition method based on parametric convolutional neural network(CNN).Due to the high-dimensional characteristics of EEG and the limitation of sample size,it is difficult to improve the recognition performance of external induced emotions based on neural network.In response to this problem,a parametric CNN network architecture-EMONET is proposed in this thesis,which can complete network training through small-scale data sets,and autonomously learn interpretable,spatial and time-frequency domain emotional information from multi-channel EEG signals.The experimental results on the self-built time migration EEG database and public database DEAP reveal that EMONET shows good generalization ability and temporal robustness,and significantly improves the accuracy of emotion recognition.The average accuracy reaches 87.00% under the task of binary classification,which is 12.70% higher than that of traditional SVM method,and 7.31% higher than that of Syncnet neural network.This research provides a new computing model for emotional EEG recognition based on neural networks,which can solve the problem of emotion recognition in cross time domain.4.Research on internal induced emotion recognition pattern based on self-recall.The recognition of internal induced emotions has higher practical application value,but there are many difficulties in the evaluation of emotional induction quality and the calibration of emotional maintenance time.Therefore,the research on the EEG characteristics and recognition patterns of internal induced emotions faces huge challenges.In this thesis,the EEG characteristics and recognition patterns of six types of internal induced emotions,pleasure,neutral,sadness,disgust,anger and fear,are studied for the first time,and two internal induced emotion recognition models are established:(1)The first model uses internal induced emotions for both training and recognition.The recognition accuracy is 87.36% and 54.52% for discriminating positive from negative emotions and classifiying emotions into six categories tasks respectively.(2)The second model uses external induced emotions for training and internal induced emotions for recognition.The accuracy is 78.53% and 49.92% for discriminating positive from negative emotions and classifiying emotions into six categories tasks respectively.In addition,we explores the important features,electrodes,and brain regions of internal induced emotions,and finds that there is a relatively stable neural pattern of internal induced emotions,and it shows significant consistency with external induced emotions.This research provides an important basis and support for the online recognition and application of endogenous emotions based on EEG.