Representing Image Matrices: Eigenimages Versus Eigenvectors

We consider the problem of representing image matrices with a set of basis functions. One common solution for that problem is to first transform the 2D image matrices into 1D image vectors and then to represent those 1D image vectors with eigenvectors, as done in classical principal component analysis. In this paper, we adopt a natural representation for the 2D image matrices using eigenimages, which are 2D matrices with the same size of original images and can be directly computed from original 2D image matrices. We discuss how to compute those eigenimages effectively. Experimental result on ORL image database shows the advantages of eigenimages method in representing the 2D images.