Based On Manifold Learning Facial Expression Recognition Research

Facial expression is an important body language. In our daily life, only 7 percent of information is expressed by oral language and 55 percent of information is expressed by facial expression. The major work of facial expression recognition is using computer to extract facial' features of all expression images. Then acrroding to the difference of features, each image is classified to one class of seven different expressions. It makes computer know the expression states from the classify result and achieve Human-Computer Interaction. Much progress has been got, but, in real life, on the influence of _illumination_, posture, noise, masking, and so on, facial expression recognition technology still needs to do more researches to achieve praticail applications.In this paper,we analyzed the process of domestic and international facial expression recognition technology in recent years and discussed several problems about facial expression recognition by computer. The application of Manifold learning method in facial expression recognition is detailedly introduced and a series of experiments are carried out. The research work in this paper mainly includes the following several respects:1. A nonlinear dimensional reduction method named Manifold Learning was particularly introduced. Some of classical Manifold learning algorithms, such as Isomap, locally linear embedding, Laplacian eigenmaps, Hessian-based locally linear embedding and local tangent space alignment, were recommended in detail. The advantages and disadvantages of above Manifold Learning algorithms was analysised.2. A facial expression recognition method based on Contourlet Transform and Locally Linear Embedding was prresented. It explicitly introduced the theory of Contourlet Transform. Using Contourlet Transform for facial expression feature extraction, it generated the multiresolution and multiscale feature of original image. LLE algorithm was used for feature dimensional reduction. Experiments were carried out on JAFFE database and Cohn-Kanada database. Compared with Wavelet+LLE+SVM and PCA+SVM, the maximum recognition rates for facial expression recognition of non-given person of CT+LLE on JAFFE database and Cohn-Kanada database respectively are 63.81 percent and 69.1 percent, which is higher than that of two methods.3. Original Local Binary Pattern (LBP) operator, multiresolution LBP operator, rotation invariance LBP operator and uniform LBP operator were introduced and the advantages and disadvantages of them was analyzed. We mainly introduced the application of uniform LBP operator for facial expression feature extraction.4. A facial expression recognition method with Local Binary Pattern and Laplacian Eigenmaps was presented. It introduced a framework algorithm for data dimensional reduction which named graph embedding. The Laplacian Eigenmaps algorithm was recomposed by designing the neighboring weight matrix. Plenty of experiments for non-given person facial expression recognition were carried out on JAFFE database and Cohn-Kanada database. The influences of LBP parameters (P,R) and block dividing on experimental result was analyzed. Compared LE with PCA and LDA, the maximum recognition rates of LBP+LE on JAFFE database and Cohn-Kanada database respectively are 70.48 percent and 70.95 percent, which are both higher than LBP+PCA and LBP+LDA. It proves the method presented in this paper is effective and feasible.