On The Facial Expression Recognition System Based On Dynamic Sequence Feature

Expression is the external manifestation of inner psychological activities, the methods toextract Facial expression can divided into two major categories, one category is based onstatic images, and the other is based on dynamic image sequences. Extracting a feature basedon static images is simple intuitive and fast, can achieve good recognition during therecognition of a specific person, when recognition of non-specific expression the result ispoor, because the information of expression feature contained by the static image is limited,and different face shape, color, factors such as lighting conditions will affect the result.Theimage sequence can be more effectively reflect the essence of the process of facial expression,it is easier to eliminate the interfering factors, during the non-specific expression recognitioncan also get good results. The dynamic sequence of expression an image contains moreinformation, is broader applications and has more practical significance. This subject is basedon dynamic image sequences. This paper is mainly carried out the study of the followingaspects:1. We studied the face detection algorithm based on skin color, and then discuss the facedetection algorithm based on Adaboost algorithm.Combined with the ELM algorithm we canimprove the detection rate of Adaboost algorithm.2. We studied the tracking theory based on Candide3facial model and according to theresearch we further developed the tracking process. Based on the research of facialtracking,we put forward a dynamic feature extraction method based on six parameters offacial model.We introduced the active appearance model algorithm to locate and track featurepoints of facial expression, then studied the tracking theory. Based on Candide3facial modeland according to the research we further developed the tracking process. We use the DynamicTime Warping (DTW) technique to align the image sequence, and then extract the featurevector. The results show that the method based on the model parameters has a goodclassification results when extract dynamic characteristics. 3. We discuss the key problem of manifold learning and linearization manifold learningin facial expression recognition applications.At the same time; we reveal their differences andrelations. We did some research on linear subspace-based facial expression recognitionmethod and classification.We improve the structure of the expression recognitionsystem.According to the expression category the training samples is naturally divided into sixfeature subset,we study the facial expression classification under the projection of the originalmotion characteristics in the feature subspace. We proposed a recognition method that basedon improved nearest neighbor classifier. The experimental results show that Nearest neighborrecognition method can search neighbors very fast,This method has the advantage of smallamount of calculation，improve the accuracy of the classification.4. We set up a face tracking and expression recognition system that based on the activemachine vision platform. Based on this platform, we did a lot experimental study and relatedtests. We described the architecture of the system, and described the key functional modulesof the facial expression recognition system. The system detect human face in the attentionselected area，track the human face by controlling with the PTZ, and recognize facialexpression in real-time. We did a lot of experiments on face tracking and facial expressionrecognition, including face tracking in complex and simple background, and the accuracy ofthe facial expression recognition, and the platform was verified by experimental results.Finally, the main content of this dissertation is summarized, and the further researchesare discussed.