Abstract
Since An optimal discriminant vector of linear discriminant (LDA) is similar to normal vector of classification hyperplane of support vector machine (SVM), and a optimal set of uncorrelated discriminant vectors is superior to optimal set of orthogonal discriminant vectors, inspired from the idea of SVM, an optimal set of uncorrelated margin discriminant vectors is presented. A modified SVM is first proposed by adding a constrained condition; then the optimal set of uncorrelated discriminant vectors can be recursively extracted from samples through a quadratic optimal problem. The proposed method inherits the merits of the SVM, and can deal with small sample size problem and be expanded into problem of nonlinear feature extraction through kernel method, The simulations demonstrate the efficiencies of the proposed algorithm.