Personalized face verification system using owner-specific cluster-dependent LDA-subspace

We propose an owner-specific cluster-dependent linear discriminant analysis (OSCD-LDA) method, and apply it to develop a personalized face verification system. Before the owner enrollment, our system first divides all the training face images into a number of clusters, each containing a subset of face images having similar characteristics. Once the owner completes the enrollment procedure, the system assigns the owner to the cluster that contains faces most similar to the owner's training faces. Then, the system uses the training faces in this most similar cluster to determine the OSCD-LDA subspace for computing the matching score. This OSCD-LDA subspace can be considered as a personalized subspace, trained specifically for this owner in order to best discriminate this particular owner from other non-owners. Our experimental results have shown that the proposed OSCD-LDA method outperforms the conventional LDA method, and can reduce false acceptance rate and false rejection rate by about 40 percent when using the XM2VTS database.