Using component features for face recognition

We explore different strategies for classifier combination within the framework of component-based face recognition. In our current system, the gray values of facial components are concatenated to a single feature vector which is then fed into the face recognition classifier. As an alternative, we suggest to train recognition classifiers on each of the components separately and then combine their outputs using the following three strategies: voting, sum of outputs, and product of outputs. We also propose a novel Bayesian method which weighs the classifier outputs prior to their combination. In experiments on two face databases, we evaluate the different strategies and compare them to our existing recognition system.