Robust face recognition via automatic grouping sparse representation

Recently, sparse representation based classification
^[4]
was successfully applied to face recognition. In SRC, the testing samples are represented as a sparse linear combination of the training samples and then classified according to the reconstruction error. The key of SRC is solving a Lasso problem. However, Lasso tends to select only one sample from a group of correlated training samples. Indeed, group-clustered sparsity rather than only sparsity is preferred, especially in highly correlated samples. In this paper, we proposed an automatic grouping sparse representation based face recognition algorithm. In this algorithm, the basic sparse representation model Lasso is replaced with the automatic grouping sparse representation model OSCAR, which make the non-zero coefficients of the solution grouped automatically on the highly correlated samples and keep the sparsity as well. Experiments on the benchmark datasets Extended Yale B and AR show that the algorithm proposed in this paper achieve better recognition rate and is more robust than the basic SRC.