Feature extraction via prototype margin distance maximizing criterion for subspace learning

Abstract

Locality sensitive discriminant analysis (LSDA) is an effective feature extraction algorithm, which captures the local structure and label information to build margin for classification. However, the margin of LSDA is easily affected by outliers. To overcome the limitations of LSDA, we propose a new margin-based feature extraction algorithm called prototype margin distance maximizing criterion (PMDMC). Specifically, for each class, we first find its either heterogeneous nearest marginal samples to build the between-class marginal patch. While one side does not characterize the margin. Therefore, for between-class marginal patch, we turn back to find its corresponding within-class marginal patch in the original class. With the between-class marginal patch and within-class marginal patch, we build a novel margin. Finally, we use the Fisher-like criterion to find the optimal feature subspace. In this feature subspace, the margins between different classes are enlarged by maximizing the distance between two heterogeneous marginal patches, and simultaneously minimizing the distance between the prototypes (class means) and their corresponding within-class marginal patches. The obtained margins are more discriminative than before. What's more, under this margin design, the outliers which are usually lying in margins can be pulled close into their respective homogeneous domains. From the comparative experiments performed on the ORL, Yale, and AR face databases, we prove the effectiveness of the proposed PMDMC.