DOMAIN TRANSFER SPARSE REPRESENTATION FOR SINGLE SAMPLE FACE RECOGNITION

In this paper, we propose a new single sample face recognition approach under the widely used sparse representation-based classification (SRC) framework. Previous work has shown that SRC only works well when there are sufficient number of training samples per person and not suitable for SSFR. To address this, we propose a domain transfer sparse representation-based classification (DT-SRC) method by using an auxiliary dataset to learn intra-class variations and transferring them into the single-sample training set. Since the auxiliary and training sets are likely captured in different environments, we apply the dictionary learning technique to learn a meta-space to transfer intra-class variations from the auxiliary set to the training set. To achieve this, we minimize the distribution difference of these two datasets in the meta-space so that such information can be effectively transferred. We extend DT-SRC to discriminative DT-SRC (DDT-SRC) by making use of the label information samples in the auxiliary set to exploit more discriminative information in the learned meta-space. Experimental results on three face benchmark datasets demonstrate the effectiveness of the proposed approach.