Learning a Synchronous MAP for Improved Face Recognition

Maximum A posteriori Probability (MAP) is a machine-learning method for classification of objects by their visual representations, such as humans by the images of their faces. Multiple-class recognition can be rendered in a two-class problem, through the attempt to categorize image differences as more likely under either the between- or the within-class hypotheses. When Gaussian models are learned for these two probability densities, Principal Component Analysis (PCA) is used to estimate them. This procedure leaves open the choice for the dimensionalities at which the two models are estimated-a 2-parameter learning problem. Synchronous MAP (sMAP) is proposed here to reduce the complexity to a single dimension by learning a synchronization between the models on the basis of their power spectra. Experiments are reported on the standard, publicly available FERET fafb Sep96 test protocol that demonstrate sMAP results which are almost twice better than the previously reported (non-synchronous) MAP results, and also improve significantly over the previous state of the art-another subspace approach, based on Linear Discriminant Analysis (LDA).