Facial Emotion Recognition Based On Arousal-Valence Emotion Model

The development of artificial intelligence will make the future computer have the ability of not only IQ but also EQ. Affective computing which focus on making computer emotional has attracted more and more attentions. And facial expression recognition has become a hot research topic in the field of affective computing.Research on facial expression recognition is taken in this thesis based on the arousal-valence continuous emotion model and regression methods. Furthermore, the main research issue is how to improve the recognition accuracy by using the correlation between arousal and valence dimensions.Firstly, the correlation between arousal and valence is discussed from the view of psychology and researched based on the statistical method. The experimental results on AVEC2013, NVIE and Recola datasets indicate that the correlation is positive. Then, in order to use the correlation between arousal and valence, MSVR(Multiple Dimensional Output Support Vector Regression) is adopted to train and predict facial emotion, and a new facial emotion recognition method based on MSVR and two-level fusion is proposed which combines feature fusion and decision fusion. The contrast experimental results indicate the proposed method can obtain better recognition result than the traditional support vector regression, relevance vector machine regression and tradictional feature fusion methods.Secondly, in order to extract better facial features and use the correlation between arousal and valence, a feature selection method based on sparse representation is proposed in this thesis. The sparse features are selected on the arousal and valence dimensions respectively; then the SVR(Support Vector Regression) is adopted for training and prediction on arousal and valence dimensions respestively; finally, the output-associative decision fusion method is used to enhance the recognition accuracy. Experimental results show that the proposed method can achieve better recognition performance than the principal component analysis based dimensionality reduction method and traditional decision fusion method.At last, a facial expression recognition prototype system based on arousal-valence emotion model is designed and developed in this thesis. This system includes facial expression extraction, feature extraction, feature selection, feature fusion, regression prediction, decision fusion and other modules. The testing results indicate that this system can effectively recognize a variety of different expressions on the arousal-valence emotion model.