Monogenic features based single sample face recognition by kernel sparse representation on multiple Riemannian manifolds

Abstract

This paper presents a novel method for single sample face recognition, characterized by the use of grayscale monogenic features of the face images, and the kernel sparse representation on multiple Riemannian manifolds. To indicate regional face discriminability, the multi-scale extended monogenic features are firstly locally extracted from different regions of an image, according to a specific face partition scheme, and then the intrinsic subspace of the feature vector set corresponding to each region is further extracted and modeled as a point of a Grassmann manifold. The congeneric local feature scheme is also exploited to entire image for the extraction of co-occurrence distributions of the grouped feature images of intersectional dimensions, based on a special binarization scheme applied to the resultant feature images. This derives the auxiliary marginal distribution-based descriptors residing on the closures of multiple multinomial manifolds. To train the kernel sparse representation classifier using the combined descriptors for a recognition task, the strategy of kernel alignment combining column L2-norm-based kernel matrix normalization are adopted for multiple kernel fusion, where the used kernels are all derived from Riemannian geometries of two types of manifolds. The superiority of our method is demonstrated through extensive experiments.