The Research Of Facial Expression Feature Extraction Algorithms Based On Manifold Learning

Facial expression implicates abundant and exquisite emotion. Expression recognitionhas broad prospects in the field of medicine and civilian applications. The main task offacial expression recognition is to reduce the dimension of expression feature space,access to the most discriminant feature, finally be able to perform facial expressionclassification quickly and accurately. Therefore, feature extraction is a key step ofexpression recognition. It can affect the final results of expression recognition directly.This paper mainly study expression recognition feature extraction on the basis oflearning theory and algorithm of manifold. Main work and study are as follows:(1) Although traditional supervised NPE take the class information into consideration,it can’t preserve the manifold structure of neighbor data. For this question, a supervisedfeature extraction algorithm (NSNPE) based on a new distance between samples isproposed. First a certain kind of dissimilarity between data points is defined by ImageEuclidean Distance; and then combined the above dissimilarity with class labels toachieve a new distance formula; finally the new distance is introduced into NPE toperform dimension reduction. The new distance is only take the spatial information ofimage pixels account, which preserve the manifold structure well, but also make the datain different class separate far from each other, the data in the same class close to eachother.(2) PCA can only keep the global structure, while NPE preserves the similaritybetween neighbor data, but ignores the difference between them. Focusing on theproblems mentioned above，an unsupervised feature extraction method is proposed byfusing global and local various feature, and is applied to facial expression recognition. Atfirst, it uses PCA to preserve global structure; then it defines a local diversity scatter and alocal similarity scatter by manifold learning algorithms, combining with local maximumscatter difference criterion, it can efficiently preserve the variety of local manifold; finally,the low dimensional feature is extracted by combining the global feature with localvarious feature for expression classification. This method takes into account the global structure and local similar and different feature in neighbor structure among data, makingthe low-dimensional feature obtained more comprehensive, which is conducive toexpression classification.In this paper, the NPE algorithm is studied in the field of supervised andunsupervised. NSNPE is proposed when the samples category labels are known. TheGLDPE algorithm according to NPE basic principle is proposed when sample categorylabels are unknown. Finally, the experiments with SVM classifier were conducted toverify the effectiveness and superiority of the two algorithms.