2D-ONPP: Two Dimensional Extension of Orthogonal Neighborhood Preserving Projections for Face Recognition

Abstract

This paper considers the problem of orthogonal neighborhood preserving projections (ONPP) in two-dimensional sense. Recently, ONPP was proposed as a projection based dimensionality reduction technique, attempting to preserve both the intrinsic neighborhood geometry of the data samples and the global geometry. Concerned with two dimensional data, such as face images, often vectorized for ONPP algorithm to find the intrinsic manifold structure. However, ONPP can't be implemented effectively due to the high dimensionality. Therefore, a novel method, two-dimensional orthogonal neighborhood preserving projections (2D-ONPP), directly based on 2D image matrices instead of ID vectors, is proposed for face recognition society. It finds an embedding that preserves neighborhood geometrical features and detects the intrinsic image manifold structure. The performance of the proposed algorithm is compared with existing 2D-PCA and ONPP methods on ORL and Yale B databases. Experimental results show the efficient computation performance and the competitive average recognition rate of our 2D algorithm.