Patch Based Face Recognition via Fast Collaborative Representation Based Classification and Expression Insensitive Two-Stage Voting

Small sample size (SSS) is one of the most challenging problems in Face Recognition (FR). Recently the collaborative representation based classification with 12-norm regularization (CRC) shows very effective face recognition performance with low computational cost. Patch based CRC (PCRC) also could well handle the SSS problem, and a more effective method is conducted PCRC on different scales with various patch sizes (MSPCRC). However, computation of reconstruction residuals on all patches is still time consuming. In this paper, we devote to improve the performance for SSS problem in face recognition and decrease the computational cost. First, fast collaborative representation based classification (FCRC) is proposed to further decrease the computational cost of CRC. Instead of computing reconstruction residual on all classes, FCRC computes the residual on a small subset of classes which has a big coefficient, such a category full make use of the discrimination of representation coefficients and decrease the computational cost. Our experiments results show that FCRC has a significantly lower computational cost than CRC and slightly outperforms CRC. FCRC is especially powerful when it is applied on patches. To further improve the performance under varying expression, we use a two-stage voting method to combine the recognition outputs of all patches. Extended experiments show that the proposed two-stage voting based FCRC (TSPFCRC) outperforms many state-of-the-art face recognition algorithms and have a significantly lower computational cost.