Ensemble-Based Discriminant Manifold Learning for Face Recognition

The locally linear embedding (LLE) algorithm can be used to discover a low-dimensional subspace from face manifolds. However, it does not mean that a good accuracy can be obtained when classifiers work under the subspace. Based on the proposed ULLELDA (Unified LLE and linear discriminant analysis) algorithm, an ensemble version of the ULLELDA (En-ULLELDA) is proposed by perturbing the neighbor factors of the LLE algorithm. Here many component learners are generated, each of which produces a single face subspace through some neighborhood parameter of the ULLELDA algorithm and is trained by a classifier. The classification results of these component learners are then combined through majority voting to produce the final prediction. Experiments on several face databases show the promising of the EnULLELDA algorithm.