A New Perspective on Principal Component Analysis using Inverse Covariance

In this research paper we have proposed a new method of the orthogonal projection approximation for the extraction of the principal eigenvectors of a data while using Σ-1 inverse covariance regularization, thus naming the new method as Inverse Covariance Principal Component Analysis (ICPCA). The basic idea lies in the mapping of an input space into a feature space via inverse covariance Σ-1 factorization and then computing the principal components in the extracted feature space. The performance of the proposed method has been shown quantitatively and qualitatively on a well known the Essex University's image database. The comparison shows that the proposed method outperforms competing Eigenvalue Decomposition (EVD) method (classical Principal Component Analysis) in variance coverage as well as in the execution time.