Random Projection on Sparse Representation based Classification for Face Recognition

Sparse Representation based Classification for Face Recognition (SRC-FR) has becomes popular, because of its ability to overcome several problems in FR such as occlusion and image corruption. Given this advantages, this method suffers from heavy computational process. In this paper we propose dimensionality reduction of image samples to reduce the computational burden. This reduction is performed by multiplying the feature matrix with random projection matrix
$(\Phi)$
of smaller size than feature matrix A. Two random projection matrices are generated using Gaussian and Uniform distribution. Several reduction factor in matrix
$\Phi$
are verified which are from 1 to 256, are evaluated. Higher reduction factor indicates higher dimensionality reduction. As a reference we compared the proposed reduction method to the classical linear down scaling the image. The simulation results on AT&T Dataset that consist of 400 images shows that the proposed method with reduction factor of 8 to 256, achieve recognition rate higher than the classical linear down-scaled method. In addition, the proposed method also shows a better recognition rate up to 5% to the original SRC method.