关键词: 核函数方法, 规范化KDDA, KPCA, 小样本问题

Abstract: Traditional methods, such as PCA(Principle Component Analysis ) and LDA(Linear Discriminant Analysis), not only suffer from the so-called “Small Sample Size”(SSS) problem, but also are insensitive to the high order relations of image pixels. In this paper, we propose kernel machine based regularized discriminant analysis method, which projects the image space to the high dimensional feature subspace through some non-linear transformation, and then performs discriminant analysis using the “kernel-skills” on the new feature subspace. Extensive experiments on ORL database indicate the method proposed outperforms the traditional PCA, KPCA, LDA methods on feature extraction. It can also simplify the classifier design, meanwhile, accomplish high recognition rate.

Key words: kernel methods, Regularized Kernel Direct Discriminant Analysis, Kernel Principle Component Analysis (KPCA), Small Sample Size Problem (SSS)