Due to the randomly generated initial vector (which may converge to local minimum) in the fixed-point algorithm, the existing fast principal component analysis (fast PCA) has unstable performance in the order it generates eigenvectors. In this paper, by modifying the fast PCA algorithm, the deficiency of fixed point algorithm is minimized. To evaluate the merit of the proposed modified algorithm, similarities between standard eigenvectors from eigenvalue decomposition (EVD), eigenvectors from fast PCA, and eigenvectors the proposed modified algorithm are compared. The comparison indicates that the eigenvectors from the modified fast PCA has better similarity to the standard eigenvectors. In addition, the fast PCA and modified fast PCA are compared into the face recognition application to evaluate their performance.
