Embedded Hidden Markov Model For Facial Expression Recognition

Facial expression recognition is a quite challenging new domain in pattern recognition. It has been more and more applied to the fields of artificial intelligence, man-machine interaction and the research of psychology. In those applications, exactly recognizing the facial expression is very important to the further exploration of affective computing and psychology analysis. Therefore, facial expression recognition plays a very important role and has become a main part of computer vision now.This thesis focuses on exact facial expression recognition and deals with the design of a recognition system for facial expression. The system uses 2D-DCT to extract the expression features and combines 2D-DCT with vector quantization to produce the observation sequences for HMM. Then it extends the HMM structure to EHMM for building the facial expression models and achieves achievement in theoretical studying and practical application.The main contributions are as follows:Firstly, this system makes use of the two dimension discrete cosine transform (2D-DCT). It divides the single picture of human face into pieces basing on the concrete facial expression feature distribution, and extracts the expression features in detail. The system chooses the number of features to be extracted in every piece according to the structure of EHMM behind, which makes the extracted feature effective and typical.Secondly, because of the redundancy of 2D-DCT features and the EHMM's discrete characteristic, the feature extraction part combines the vector quantization (VQ) algorithm with 2D-DCT to produce the observation sequences. This method reduces the dimensions of feature vector a lot and makes the compute simpler to a great extent.Finally, this thesis builds a facial recognition system based on HMM and gets the results through experiments. Those results are not so good. After analyzing the fact of that, the idea of extending HMM to EHMM comes up. The structure of EHMM is designed according as the fact that facial expression features are usually distributed around the eyes and the mouth. Then a new system is established. Comparing with the original system, the new one increases the recognition rates and improves the recognition speed at the same time.