Subspace Ensembles for Face Recognition

In many cases, the statistical learning methods used for solving the face recognition problem tend to be overfitted to the training data. It is because of small number of individual faces and large number of features in the face space. Utilizing a single subspace, these methods can also suffer from different intra-personal variations. One way for handling these problems is to construct multiple subspaces from the original face space. In this paper a novel ensemble framework based on random subspace method is presented. We focus on Fisherface-based face recognition approach and propose to use classifier selection methodology to improve the random subspace LDA technique. Our method uses a small subset of complementary classifiers from the pool of randomly generated classifiers and by combining their decisions via a trainable combiner (Mixture of Experts), tries to achieve high classification accuracy with lower operating complexity. Experiments on the Cohn-Kanade Expression Database demonstrate the satisfactory performance of the proposed method as compared to Eigenface, Fisherface and random subspace LDA methods.