Research Of Dynamic Facial Expression Recognition Technology

The design of the interface of human-computer interaction of new generation and affective intelligence are two international research frontiers. In these two frontiers, a key technology is how to obtain people's hidden feeling. The facial expression contains abundant information related to the change of the states of heart as the main behavior way of people's hidden feeling. Therefore the automatic recognition technology of facial expression has received the researchers' extensive concern.Combining the psychological research results, in this thesis, we have set up one basic facial expression database which includes 840 video sequences at first, secondly, we confirm two interfaces used to structure the platform of facial expression recognition considering the dynamic characteristic, then making use of the well-known algorithms about facial regions location, feature extraction and classification, we build one instance of the system of facial expression recognition, finally, the instance has been tested using four kinds of basic facial expression in the completed database and the general recognition rate is 91.7%. The result proves the platform runs normally and can offer the good experimental environment for development and improvement of the follow-up expression recognition system , at the same time, the instance is close to demand of practical application, and lays a foundation for future application.Expressional feature extraction is a more central question in the system, because of the complexity of the applicational environment, how to get the steady information which can reflect the facial expression essentially is a difficult problem urgently to be solved. We carry on deep analysis about the effective normalization and describing rationally of the expression feature, and has applied the analysis result to the instance structured.Generally the geometrical normalization process is important to the extraction of facial expression feature. In commonly used normalization methods, the unstability of the fiducial features always causes normalization error. We analyze quantitatively the serious influence to feature extraction caused by normalization error, and propose the weighted optimal matching algorithm to solve it. The algorithm is founded on the basis of template matching theory, and assigns corresponding weight to the point by its motive degree when carrying on the calculation of correlative coefficient. Experimental results show that the algorithm can restrain the disturbance that normalization error produces, and the facial expression feature can be extracted more accurately.After building the eigenspace using the principle component analysis and reprocessing...