Sparse Representaion via l1/2-norm Minimization for Facial Expression Recognition

The l(1/2)-norm regularizer is shown to have many promising properties such as unbiasedness, sparsity and oracle properties. By exploiting these properties of l(1/2)-norm regularizer, we propose a novel model of sparse representation based classification via l(1/2)-norm minimization (em-SRC) for facial expression recognition in this paper. In l(1/2)-SRC, we use l(1/2)-norm minimization as an alternative to l(0)-norm minimization instead of using l(1)-norm minimization in the traditional l(1)-SRC. By adopting l(1/2)-norm minimization, we can find a sparser and more accurate solution than the l(1)-SRC, and the optimization problem of l(1/2)-norm minimization is much easier to be solved than that of l(0)-norm minimization. Furthermore, an active-set based iterative reweighted algorithm is proposed to solve the l(1/2)-norm minimization problem. The experimental results on JAFFE and Cohn-Kanade databases testify the efficiency of l(1/2)-SRC.