Combining Classifier for Face Identification at Unknown Views with a Single Model Image

Abstract

We investigate a number of approaches to pose invariant face recognition. Basically, the methods involve three sequential functions for capturing nonlinear manifolds of face view changes: representation, view-transformation and discrimination. We compare a design in which the three stages are optimized separately, with two techniques which establish the overall transformation by a single stage optimization process. In addition we also develop an approach exploiting a generic 3D face model. A look-up table of facial feature correspondence between different views is applied to an input image, yielding a virtual view face. We show experimentally that the four methods developed individually outperform the classical method of Principal Component Analysis(PCA)-Linear Discriminant Analysis(LDA). Further performance gains are achieved by combining the outputs of these face recognition methods using different fusion strategies. © Springer-Verlag 2004.