Orthogonal quadratic discriminant functions for face recognition

Abstract

Small sample size (SSS) problem is usually a limit to the robustness of learning methods in face recognition. Especially in the quadratic discriminant functions (QDF), too many parameters need to be estimated and covariance matrix of a class is usually singular. In order to overcome the SSS problems, we proposed a novel approach called orthogonal quadratic discriminant functions (OQDF). The OQDF assumes probability distribution functions of each two classes of face images have a uniform shape. Then, three OQDF models are developed. The Laplacian smoothing transform (LST) and Fisher's linear discriminant (FLD) are employed to preprocess the face images for the OQDF classifier. Finally, we evaluate our proposed algorithms on two face databases, ORL and Yale. © 2009 Springer Berlin Heidelberg.