Dimensionality Reduction for Face Recognition

In Advances in Visual Form Analysis -- C.Arcelli, L.P.Cordella and G.Sanniti (Eds.) -- World Scientific, 1997

Abstract:

An image of a face depends not only on its shape, but also on the illumination conditions, facial expresions and view-point possition. In this paper the representation capabilities of the derivatives of the Gaussian filter are studied for these cases where faces vary in facial epresion, small orientations and are viewed in different illumination conditions.
Since these representations are defined in a high-dimensional space, a dimensionality reduction metod is necessary if further studies are to be conducted. For that reason, the MultiDimensional Scaling (MDS) method is used; results are copared with principal components analysis.
Finaly a 3-layer full-conected neural network is used to learn the resulting spaces. The resulting classification will specify which Gaussian filter is better to use in each facial change.
Results show that while faces vary in facial expresions and (small) orientations all Gaussian filters can obtain good classification results. But when illumination changes are also included into the database, the first and fourth derivative of the Gaussian filter best classify the data (the fourth though, it is know to be much more sensitive to noise). Iconic representations (composed of up to the fourth derivative of the Gaussian filters) are also studied; showing, generally, even higer performense than the first and fourth derivative.