Comparing subspace methods for face recognition

Methods of representing images by projection onto a linear subspace, such as Principal Component Analysis (PCA), 2-Dimensional PCA (2D PCA), and two variants of robust Linear Discriminant Analysis (RLDA) using PCA to refine the chosen subspace, are compared by the accuracy of linear and nearest neighbour classification of the reduced and original image vectors. Experiments with the Olivetti-ATT Database of Faces and the Face Expression Database have shown that the RLDA with an extended subspace outperforms the other methods.