Fast Sparse Representation Classification Using Transfer Learning

摘要: Under certain conditions, the sparsest solution to the combination coefficients can be achieved by L1-norm minimization. Many algorithms of L1-norm minimization have been studied in recent years, but they suffer from the expensive computational problem, which constrains the applications of SRC in large-scale problems. This paper aims to improve the computation efficiency of SRC by speeding up the learning of the combination coefficients. We show that the coupled representations in the original space and PCA space have the similar sparse representation model (coefficients). By using this trick, we successfully avoid the curse of dimensionality of SRC in computation and develop the Fast SRC (FSRC). Experimental results on several face datasets illustrate that FSRC has comparable classification accuracy to SRC. Compared to PCA+SRC, FSRC achieves higher classification accuracy.