Novel approaches to one-directional two-dimensional principal component analysis in hybrid pattern framework

Abstract

In this paper, we present variations of one-directional two-dimensional principal component analysis (2DPCA) in hybrid pattern framework. Using hybrid pattern framework, we propose two novel methods, namely extended sub-image principal component analysis (ESIMPCA) and extended flexible principal component analysis (EFLPCA). The ESIMPCA operates on sub-image and full image at a time and captures the local and global variation of images. The dimensionality problem of ESIMPCA feature matrices is eliminated by further applying 2DPCA on two-dimensional ESIMPCA feature matrices to generate EFLPCA feature matrices. The summarization of variances, time and space complexities of the proposed methods and their relationship with some existing variations of one-directional 2DPCAs are addressed. The experiment is conducted on ORL and YALE face databases with different image resolutions. The experimental results, using EFLPCA, show superiority performance in terms of feature dimensionality, recognition accuracy and speed with reasonable space requirement over some existing variations of one-directional 2DPCA.