Comparative Assessment of Independent Component Analysis (ICA) for Face Recognition

This paper addresses the relative usefulness of Independent Component Analysis (ICA) for Face Recognition. Comparative assessments are made regarding (i) ICA sensitivity to the dimension of the space where it is carried out, and (ii) ICA discriminant performance alone or when combined with other discriminant criteria such as Bayesian framework or Fisher’s Linear Discriminant (FLD). Sensitivity analysis suggests that for enhanced performance ICA should be carried out in a compressed and whitened Principal Component Analysis (PCA) space where the small trailing eigenvalues are discarded. The reason for this finding is that during whitening the eigenvalues of the covariance matrix appear in the denominator and that the small trailing eigenvalues mostly encode noise. As a consequence the whitening component, if used in an uncompressed image space, would fit for misleading variations and thus generalize poorly to new data. Discriminant analysis shows that the ICA criterion, when carried out in the properly compressed and whitened space, performs better than the eigenfaces and Fisherfaces methods for face recognition, but its performance deteriorates when augmented by additional criteria such as the Maximum A Posteriori (MAP) rule of the Bayes classifier or the FLD. The reason for the last finding is that the Mahalanobis distance embedded in the MAP classifier duplicates to some extent the whitening component, while using FLD is counter to the independence criterion intrinsic to ICA.