Facial Expression Recognition Research Based On Feature Learning

A significant feature of the intelligent interaction is that people's intentions and emotions are better perceived.Expressions of human can be a appropriate display of people's inner activities,emotional state and their own intentions,so the fast and accurate expression recognition for more intelligent human-computer interaction is very important.The task of expression recognition has been a hot topic of research with a lot of achievements,but the study of more extensive facial expression recognition still has important academic value and application significance because of the complexity itself.In this thesis,the goal is to construct a robust expression recognition system and the focus is that how to extract the more expressive expressions.The features of human facial expression are mainly studied from two aspects: geometric shape and texture statistics.In the aspect of geometry shape,exploring the complete feature regions of each expression,a method based on the complete feature set is proposed.In textural statistics,a method based on expression space is proposed with the probability linear discriminant analysis.Moreover,the geometry shape and texture statistics are different views for learnings of expression features,so the combination of these two constitutes a new expression recognition framework(SS-PLDA),and the proposed method is proved to be effective through the experiments on generalized databases.The main work of this paper includes:1)proposed a method of expression recognition based on the complete feature setIn this thesis,each expression is mapped to different regions of the human face,and the concept of a complete feature set of facial expressions is introduced,in which the complete feature set is formed by all relevant expression regions of human faces.The complete feature set of the expression is essentially an exploration about geometry and shape information on human faces,and the regions that are not related to facial expressions are excluded from the expression feature,which makes it a low computational complexity but high effectiveness.And the performance is verified by experiments.2)proposed a method of the expression analysis and recognition based on the expression spaceIn this thesis,a generative model called probabilistic linear discriminant is used to decompose the information of face into expression-related,expression-independent and random noise.It mainly focuses on the learning of expression space in the generative model.And the similarity measurement of any two samples in the expression space is obtained based on the Bayesian method,thus a kNN expression classifier is designed based on this metric method.The validity of this method is verified through experiments on the popular expression databases.3)proposed a expression recognition framework based on SS-PLDAThe expression recognition framework based on SS-PLDA is composed of the sparse selection and the probability linear discriminant analysis.As previously mentioned,the sparse selection represents a kind of expression feature learning method based on geometric shape while the PLDA based on texture statistics.Thus the features extracted contain both the geometrical and texture statistical information,which is a more expressive expression feature.