A matrix-oriented method for appearance-based data compression - An idea from group representation theory

Abstract

Motivated by ideas of group representation theory, we propose a matrix-oriented method to dimension reduction for image data. By virtue of the action of Stiefel manifold, the original image representations can be directly contracted into a rather low-dimensional space. Experimental results show that the performance of PCA and LDA is significantly enhanced in the transformed space. In addition, the reconstructed images by proposed algorithm are better than those by 2DPCA. © Springer-Verlag Berlin Heidelberg 2004.