Mutual Neighborhood Based Discriminant Projection for Face Recognition

Linear Discriminant Analysis is optimal under the assumption that the covariance matrices of the conditional densities are normal and all identical. However, this doesn't hold for many real world applications, such as Facial Image Recognition, in which data are typically under-sampled and non-Gaussian. To address this deficiency the Non-Parametric Discriminant method has been developed, but it requires model selection to be carried out for selecting the free control parameters, making it not easy for use in practice. We proposed a method, Mutual Neighborhood based Discriminant Projection, to overcome this problem. MNDP identifies the samples that contribute most to the Baysesian errors and highlights them for optimization. It is more convenient for use than NIDA and avoids the singularity problem of LDA. On facial image datasets MNDP is shown to outperform Eigenfaces and Fisherfaces under various experimental conditions.