Sample pair based sparse representation classification for face recognition

• We proposed a method to use virtual available facial images for face recognition.
• We improved the linear regression classification method for face recognition.
• The classification accuracy on sample pairs are better than that on original samples.
• This method achieves lower classification error rates than many other methods.
• This method performs well even when there are few training samples of each class.

Sparse representation classification, as one of the state-of-the-art classification methods, has been widely studied and successfully applied in face recognition since it was proposed by Wright et al. In this study, we proposed a method to generate virtual available facial images and modified the well-known linear regression classification (LRC) and collaborative representation based classification (CRC) for face recognition. The new method integrates the original and virtual symmetry facial images to form a training sample set of large size. Experimental results show that the proposed method can achieve better performance than most of the competitive face recognition methods, e.g. LRC, CRC, INNC, SRC, RCR, RRC and the method in Xu et al. (2014). This promising performance is mainly attributed to the fact that the sample combination scheme used in the new method can exploit limited original training samples to produce a large number of available training samples and to convey sufficient variations of the original training samples.