Optimizing Regularized Discriminant Analysis in Virtual Training Samples

In this paper,an improvement is made through selecting a group of normal orthogonal vectors in feature subspace,to generate large amount of virtual training samples.The experimental results on ORL face database show that by the proposed improved method,eigenvalues of the optimized covariance matrices become larger and the matrices inverses become more stable.By using the optimized covariance matrices to optimize the new regularized discriminant analysis (RDA),the correct classification rate is higher than that by the old RDA.