Feature extraction using two-dimensional neighborhood margin and variation embedding

In this paper, we introduce a novel linear discriminant approach called Two-Dimensional Neighborhood Margin and Variation Embedding (2DNMVE), which explicitly considers the modes of variability among nearby images and the discriminating information. To be specific, we construct an adjacency graph to model the intra-class variation, which characterizes the modes of variability of the face images, of the values of face images from the same class, and inter-class variation which encodes the discriminating information, and then incorporate the modes of variability and discriminating information into the objective function of dimensionality reduction. Thus, 2DNMVE is robust to intra-class variation and has better generalization capability on testing data. Experiments on four face databases show the effectiveness of the proposed approach.

► We construct an adjacency to model the intra-class and inter-class variation.
► Intra-class variation characterizes the most important modes of variability of patterns.
► Inter-class variation characterizes the between-class separability.
► We incorporate the intra-class variation and inter-class variation into the objective function.
► Our approach well encodes the discriminating information and unfolds the manifold structure.