Dimensionality Reduction by Sparsiication in a Local-features Representation of Human Faces

Low-dimensional representations are key to solving problems in high-level vision, such as face compression and recognition. On the basis of a multi-dimensional Gaussian model, Principal Component Analysis (PCA) has been applied to establish that a global representation with dimensionality 400 suuciently encloses the space of human faces. Local Feature Analysis (LFA) has been proposed to achieve the same low dimensionality, but with a topographic set of lters that resemble the local features of human faces. Here we study a sparsiication strategy in the context of LFA and show that it gives a substantially reduced estimate of the dimension-ality of face space, 200. We nd that LFA is a more accurate, non-Gaussian, parameterization of the true probability density of images of human faces, because it is able to account additionally for the partial local symmetries of this ensemble of natural objects.