Expression recognition using semantic information and local texture features

In this paper, the cross-database facial expression recognition problem is studied. First, Gaussian Mixture Model is used to improve the facial landmark detection. Second, the local image features are used to model the facial actions. Deep Neural Network is used to represent the low level data variance. Third, the top level expression recognition rules are used in a Fuzzy Inference System to improve the cross-database performance by 11.6 percent in average. Experimental results show that the expression recognition rate is improved constantly over six emotion types compared with traditional Support Vector Machines and Neural Network classifiers . The averaged improvement against Support Vector Machines is 15.1 percent and the averaged improvement against neural network is 8.1 percent.