Face Authentication Using Client Specific Fisherfaces

In the paper we propose a one dimensional client specific fisher face representation for personal identity verification. This novel LDA approach contrasts with the conventional LDA representation which involves multiple shared fisher faces. The method provides two measures for authentication: a distance to the client template, and a distance to the mean of impostors. These two decision scores are combined to achieve significant performance gains. The method is tested on the XM2VTS database according to the internationally agreed Lausanne protocol. The demonstrated performance superiority is not the only advantage of the proposed method. Additional features of practical significance include the simplicity of training, as for large user databases the proposed technique requires only a matrix multiplication of the client mean vector. Moreover, the client enrollment is insulated from the enrollment of other clients. This opens the possibility to use other than the centralised architecture for the personal identity verification system and in fact smart card processing becomes a reality without any need to restrict the representation framework and therefore the representational capacity of the system. Finally the speed of probe testing is more than two orders of magnitude faster than that achieved by conventional PCA and LDA methods as the proposed techniques involves only a single fisher face per client. These attractive properties make the method ideally suited for both representation and authentication in personal identity verification systems.