Feature extraction by structured stepwise nonparametric maximum margin criterion

In this paper, a new feature extraction method named structured stepwise nonparametric maximum margin criterion (SSNMMC) is proposed. Previous nonparametric discriminant analysis methods only use the point-to-point distance to measure class difference. In the proposed method, point-to-line distance with nearest neighbor line (NNL) theory is adopted and more intrinsic structure information of training samples is preserved in the feature space. Furthermore, the proposed method does not assume that the class densities belong to any particular parametric family and does not depend on the nonsigularity of the within-class scatter matrix, which are shortcomings of conventional linear discrimiant analysis based algorithms. Besides, limitation of feature number is overcome with the proposed method. Experiments on the ORL face database demonstrate the effectiveness of our proposed method.