Maximum Margin Local Learning Projection for Face Recognition

In this paper a novel subspace learning method called Maximum Margin Local Learning Projection (MMLLP) is proposed for pattern classification. Firstly the local structures and the dissimilarities between the manifolds are characterized, and then a linear transformation that can maximize the dissimilarities between all manifolds while simultaneously minimizing the local estimation error is proposed. MMLPP embeds the discriminative information as well as the local geometry of samples into the objective function and it can explore as much underlying knowledge inside the samples as possible. Thus the abilities of preserving the local structure in each manifold and classification are combined into MMLLP. Extensive experiments on face databases demonstrate the effectiveness of MMLLP.