Abstract

Positive definite kernels, such as Gaussian Radial Basis Functions (GRBF), have been widely used in computer vision for design- ing feature extraction and classification algorithms. In many cases non- positive definite (npd) kernels and non metric similarity/dissimilarity measures naturally arise (e.g., Hausdorff distance, Kullback Leibler Di- vergences and Compact Support (CS) Kernels). Hence, there is a prac- tical and theoretical need to properly handle npd kernels within feature extraction and classification frameworks. Recently, classifiers such as Support Vector Machines (SVMs) with npd kernels, Indefinite Kernel Fisher Discriminant Analysis (IKFDA) and Indefinite Kernel Quadratic Analysis (IKQA) were proposed. In this paper we propose feature ex- traction methods using indefinite kernels. In particular, first we propose an Indefinite Kernel Principal Component Analysis (IKPCA). Then, we properly define optimization problems that find discriminant pro jections with indefinite kernels and propose a Complete Indefinite Kernel Fisher Discriminant Analysis (CIKFDA) that solves the proposed problems. We show the power of the proposed frameworks in a fully automatic face recognition scenario.