| In recent years, the facial expression recognition technology is widely applied to various areas such as human computer interaction (HCI), emotion analysis, psychiatry and so on. The majority of researches on expression recognition are working for static expression recognition, but a very important limitation to the static information is the fact that still images usually capture the apex of the expression while the expression is actually dynamic and people seldom show apex of their facial expression during normal communication with their counterparts. Therefore, the study of dynamic facial expression recognition is of great theoretical significance and practical value.Many researchers on the dynamic expression recognition have made some achievements, but most of the recognition is not accurate enough to be practical. In the order to improve the recognition rate of dynamic expression recognition, this paper studies around the three steps of dynamic facial expression recognition which are face detection, expression feature extraction and the classification of facial expressions. For the face detection and the localization of key regions, our work adopt the AdaBoost algorithm using Haar-like features, and then the extended ASM model is applied to locate the positions of feature points based on the detected face area and the located key region. For the extraction of expression features, our work firstly extracts the geometric features and the texture features of the expression sequence based on the located feature points respectively, then the geometric features and the texture features are combined together to generate the hybrid features which we’ve taken as the final expression sequence features. For the classification of facial expressions, our work uses auto-regressive models to train expression classifiers, and a novel but effective method is proposed to measure the similarity which assists the classifiers to make decisions. The characteristics and innovations of this article are mainly presented as follows:(1) In the process of locating feature points, the Random Forest algorithm is used to set the initial positions of these points, which lead the result of location through Active Shape Model more accurate and efficient;(2) In the process of expression sequence features extraction, the hybrid features based on sequence are finally extracted. This kind of feature contains rich expression information when facial expression changes. The hybrid features consist of the geometric features and the texture features. The geometric features of one sequence are facial feature point displacements between the normalized neutral and expressive face images in the sequence, while the texture features of one sequence are normalized cross correlation coefficients of the corresponding texture regions which are centered at feature points in the normalized neutral and expressive face image in the sequence.(3) In the process of classification, the second-order auto-regressive models are used to train expression classifiers. To provide better distinctive capability for recognition, a line segment based method is proposed to compute the similarity between the predicted and actual expression sequences.Our experiments including training and testing are performed on CK+ database. Firstly, based on one same classification algorithm, we proved that the recognition rate was higher by using the hybrid features rather than only using geometric features or the texture features. And then through the contrast experiment, we verified the effectiveness of the proposed measurement of similarity based on the line segment method. In the final, the recognition algorithm proposed in this paper was compared with other recognition algorithms, results shown the superiority of our proposed method. |
