Face recognition via discriminative atom decomposition and linear subspace learning

A novel sparse coding based discriminative decomposition method is proposed to decompose facial image into different components, which are used to guide linear subspace learning for face recognition. A dictionary is learnt from the training samples and each training sample is sparsely represented by atoms in dictionary. And our idea is that discriminative atoms, i.e., atoms which are infrequently used by with relatively large coefficient in sparse coding, tend to carry more discriminative information. Therefore, we decompose a facial image into discriminative component (using discriminative atoms in sparse coding) and indiscriminative component (without using discriminative atoms in sparse coding). During subspace learning, the discriminative component is preserved while the indiscriminative component is suppressed. The experimental results on benchmark face image database suggest that the proposed method achieve good performance.