Study On Facial Expression Recognition Based On Manifold Lerning

Facial expression is a kind of important means that we communicate and expressthe emotions. Without special training, People often difficult to conceal their facialexpression. So, by the analyzing facial expression, we can obtain important information.Facial expression recognition is an important component that the implementation ofhuman-computer interaction, and an important research topic in the field of patternrecognition, image processing, in recent years, more and more researchers focus on thistopic. In this paper, we select the static facial expression image as the research object,studies how to preprocessing the human face image and how to locating the face organsand extracting local feature, and study of the facial expression recognition based onmanifold learning in depth.In this paper, the main research work including the following four aspects:1) After reading a large number of references of facial expression recognition athome and abroad, we summarized the development of the current situation and theexisting problems in the filed of facial expression recognition, we summarized the keysteps of facial expression recognition system such as local segmentation, featureextraction, feature dimension reduction, we research the development of background andcurrent situation of manifold learning algorithm in depth.2) We studied the human face localization and local facial expression featureextraction. Apply the AdaBoost algorithm to positioning face. Apply integral projectionmethods to dividing the face to mouth, nose,eyes,eyebrows, and in this section, a newmethod is presented to calculating the crest and the trough in integral projection methods,then apply the method of LBP and Gabor wavelet algorithm to extract local features. Byexperiment to verifying the effect of local facial expression feature extraction and facepositioning.3) The two algorithms of linear dimension reduction: principal component analysis(PCA) and linear analysis (LDA) and three nonlinear dimension reduction algorithm oftypical algorithms: the equidistance mapping (ISOMAP), locally linear embedding (LLE)and the local tangent space alignment (LTSA) has carried on the theoretical analysis andexperimental comparison.The experiment contents is using artificial data to verifying theeffect of low dimensional embedding and using facial expression library data tocomparing recognition rate of facial expression. The experimental results show thatLTSA algorithm has certain advantage than other methods. 4) To solve the problem that the ILTSA could not efficiently handle ever-increasing data set,this paper studied generalization method for the ILTSA and its application in facial expressionrecognition emphatically. First introduces two generalized local tangent space alignmentalgorithm, then to apply the k neighbor calculation methods of supervision to thegeneralization of the LTSA algorithm. Further to improve the shortcomings of the LTSAalgorithm, the supervised GILTSA algorithm is proposed. Finally, the variousgeneralization manifold learning algorithm be applied to experiments of face recognitionand facial expression recognition. The experimental results showed that supervisedGILTSA algorithm results better.