Face verification with a kernel fusion method

In this paper, the fusion of information from several data representations in a face verification environment has been analyzed. We have considered three different face-based biometric representations of each subject in a subset of the Face Recognition Grand Challenge (FRGC) database: 2D texture images, 2.5D range data and mean curvature images. From these representations, linear and Gaussian kernel matrices have been defined. Fusion techniques have been applied to obtain a unique kernel from the individual kernels. The resulting kernel has been used to train Support Vector Machines (SVMs) for verification tasks. The proposed classifier outperforms the individual kernels results and the results of classical fusion techniques (feature-level and score-level methods) in different security level systems.