Non-iterative generalized low rank approximation of matrices

As an extension to 2DPCA, generalized low rank approximation of matrices (GLRAM) applies two-sided (i.e., the left and right) rather than single-sided (i.e., the left or the right alone) linear projecting transform(s) to each 2D image for compression and feature extraction. Its advantages over 2DPCA include higher compression ratio, superior classification performance, etc. However, GLRAM can only adopt an iterative rather than analytical approach to get the left and right projecting transforms and lacks a criterion to automatically determine the dimensionality of the projected matrix. In this paper, a novel non-iterative GLRAM (NIGLRAM) is proposed to overcome the above shortcomings. Experimental results on ORL and AR face datasets and COIL-20 object dataset show that NIGLRAM can get not only so-needed closed-form transforms but also comparable performance to GLRAM.