Facial Expression Feature Extraction And Recognition Based On Dynamic Image Sequence

Facial expression recognition is using computer to analyze facial expressions andits changes, and then to determine the internal emotional and ideological activities, atlast people achieve more intelligent among human-computer interaction. The purposeof typical facial expression recognition is to classify input of static images or videosequences into six basic facial expression. Static expression characteristics obtainedby static images, although simple, direct and fast, can be very good for certain facialexpression recognition, but it contains limited information, it is easy to be affected byother factors (such as illumination, etc.), facial expression recognition of untrainedimage get poor results[1]. Since facial expression is formed by facial muscle movement,caused by the dynamic facial features, such as eyebrows, eyes, mouth and skindeformation. Through dynamic image sequences obtained by dynamic movementcharacteristics can better reflect the essence of expression movement, also can carryout non-specific expression recognition better, and dynamic image sequenceexpression contains more information, such as facial muscle movement and neutralexpression, etc. Therefore, the application of dynamic expression image sequence ismore extensive, and more meaningful. In this paper, the basic research content is asfollows:1.First study of facial expression feature extraction algorithms, mainly themethod based on active appearance model research. This section contains the buildingof point distribution model and the training of shape. Compare the time-consumingand calculation error of several Algorithms based on active appearance, includingmodel multi-scale active shape model and reverse combination algorithm, choose suitable algorithm for dynamic sequence. The last is the improvement on the activeappearance model, the active apparent difference model algorithm is analyzed andproposed.2.Reduce the dimension of the facial expression feature eigenvalues, firstcompared the different of PCA and LDA. Then aiming at the shortcomings of thelinear dimension reduction methods, select a kind of nonlinear dimension reductionbased on manifold learning methods, compared to principal components analysis andmanifold learning dimension reduction. In the second chapter of active appearancemodel pca dimension reduction was improved, choose a manifold learning algorithm.3.Research support vector machine classification method of expression, choosean encoding method based on multi-class support vector machine. Then experimentsanalyzed by the expressions of support vector machine classification results andcauses.4.Dynamic texture features were extracted from image sequence. the firstexpounds the basic methods of local binarization, then basic LBP method is extendedto3d airspace LBP-TOP method, finally,apply LBP-TOP method for facial expressionrecognition.5.Summarize the work, and then studied the shortcomings of the previous workand proposed the prospect for further improvement and further work.