New Algorithm And Its Application To Face Recognition

Face recognition is one of the hottest topics in the field of pattern recognition. It consists preprocessing, feature extraction and classification of face design. Face feature extraction which is also known as the face representation, is the process of feature modeling of the face. The purpose is to extract the low-dimensional feature from the original high-dimensional face model to be used for subsequent classification tasks. Face extraction is the kernel step in face recognition, directly affect the recognition rate, which also the main difficulties in recognition.In this paper, more detailed of facial feature extraction technique is carried out, as follows:(1)The traditional LDA method ignored the boundaries of class structure while considered the center of class in the calculation of between-class scatter matrix, the class boundary structure is very useful in the classification, and therefore the recognition performance of LDA is not stable. We introduce the NSA and the NFA to overcome these shortcomings. NSA redefined the between-class scatter matrix, which not only considered the center of the class, but also the class structure boundaries, make up the defect of the LDA, but the within-class scatter matrix Sw of the NSA is the same like the LDA, which may affect the recognition results. NFA algorithm using a single local average instead of all the selected KNN sample to calculate the between-class scatter vector, while neglect the different KNN points help to build different between-class scattering matrix. We introduce NFA in order to solve these problems, the experimental results in the face database show that, the recognition effect of the NSA and the NFA is better than the LDA.(2)However, face image space dimension is too high, which make it difficult or impossible to find enough training samples to ensure the reversibility of the within-class scatter matrix. The author proposed two-dimensional non-parametric linear discriminant analysis, just the two-dimensional NFA and two-dimensional NSA. The experimental results in the face database show that 2D linear discriminant analysis and 2D non-parametric linear discriminant analysis is not as good as 1D, but more stable than the one-dimensional algorithm, and the running time is shorter than the 1D algorithm.(3)We can solve nonlinear problems as the emergence of the nuclear method, which applied to face recognition can get very good results. to determine the merits of kernel method often depends on what kind of kernel function we chose, so the kernel function is the key to solve the linear problem. A modular KDA which is applied to face recognition is proposed based on the KDA in this paper. The proposed method is a direct method based on linear feature extraction of sub-image matrix. Compares with the previous image vector-based nonlinear feature extraction methods (such as KPCA), you can easily carry out feature extraction method in the smaller image after blocking the original image. The process is simple, modular KDA can avoid using the singular value decomposition of matrix. The experimental results show that the recognition rate of the proposed method is better than the KDA. (4) The basic principle of SVM is to map the data to high dimensional feature space,minimize structural risk minimization (SRM) as the principle of induction,take the constructed low VC dimension optimal hyperplane as a decision surface in high dimensional space.It is based on the principle of structural risk minimization method ,which minimize the risk of upper bound. This paper based on the SVM_LDA and proposed the SVM_NSA method which applied to face recognition.The experimental results shows that the new method is superior to SVM_LDA method.