The Face Recognition Feature Extraction

Face recognition is one of the hot topic in the field of pattern recognition, and it belongs to biometrics. In the field, feature extraction is one of the key steps. In the passed decade years, many correlated algorithms have been proposed to solve the problems. For example, linear discriminant analysis(LDA), principal component analysis(PCA)(or K-L transform), two dimensional PCA, which has been proposed recently, and kernel methods based on support machine (SVM) are proposed to solve nonlinear problem. In this paper, linear and nonlinear methods on feature exaction field are both deeply analyzed, and we do experiments about these classical methods on the face databases of ORL and Yale.Usually we demand strict conditions for the face images that we get for our face recognition. In this paper, we do image processing with the standard faces of the standard ORL face database. We do 8 transformations in this paper. And the transformation is worked out with the C langue. In this way, we can test the robust of the classical methods. And this can make the base of our future work, like the recognition of the cartoon, or the recognition of the face with rich feelings, like happy, sad or making faces.To overcome the limitation of the classical linear discriminant analysis, in this paper, we propose a new kind function of linear discriminant analysis. With the experiments, we observe that the new function is better than the classical one based on the orthogonal discriminant vector set which is obtained by generalized eigenvalue decomposed. And the recognition of the vector set which is obtained by QR decomposition is higher than that is obtained by generalized eigenvalue decomposed based on the new function.Generally, difference of two classes is defined according to poin point distance in the process of feature extraction. In this paper, inspired by nearest feature neighbor line(NFL) theory, a new feature extraction, which are called nearest feature line segment(NFLS) and nearest feature line segment with parameter(NFLSP) based on the distance of the point to the line segment, has been proposed. The NFLS classifier can solve one problem that the NFL can't. With the experiments, we see that the NFLSP classifier can always reach the better result of the two above.