Combining L1-norm and L2-norm based sparse representations for face recognition

This paper for the first time proposes to combine conventional sparse representation, i.e. the L1-norm based sparse representation, and a novel L2-norm based sparse representation for face recognition. The proposed method is able to exert the advantages of these two different sparse representation methods. The underlying reasons why the proposed method can perform very well are as follows. First, the L2-norm is able to avoid outliers and the L1-norm is helpful for achieving the sparseness, which are both beneficial to accurate classification. Second, the devised novel L2-norm based sparse representation itself is able to obtain satisfactory performance. Third, the scores generated from these two methods have low correlation and can provide complementary information for recognizing the face. Extensive face recognition experiments illustrate that the proposed method obtains high accuracy, and it is more robust than both previous conventional sparse representation and L2-norm based representation.