Robust classification with available virtual samples and grouped sparse representation strategy in face recognition

ABSTRACT

Due to the variation of the images of a face, limited training samples have high uncertainty for representing a test sample. Moreover, traditional grouped sparse representation method only considers the scores from different groups, not consider the difference between different groups. To address this, in this paper, a novel method is proposed to reduce the uncertainty in face recognition. For a test sample, our method first selects its K nearest training samples to form a sub-training set, where K is smaller than the number of the training samples. Then, each selected training sample and its nearest training sample are simultaneously used to generate two virtual samples. All virtual samples form a virtual sub-training set. Then, an improved grouped spare representation method is derived on the two sets to generate two residuals. Finally, two residuals and a score are fused to classification the test sample. The score is defined as the $l_2$-norm of the cross product of two residuals. Experimental results on four databases demonstrate that our method is robust and can obtain higher recognition accuracy than the state-of-the-art approaches.

KEYWORDS:

• Sub-training set
• virtual samples
• grouped spare representation method

Additional information

Funding

This work is supported in part by National Natural Science Foundation of China (Grant 61501147), Natural Science Foundation of Heilongjiang Province (Grant F2015040), China Postdoctoral Science Foundation (Grant 2016M601438) and The Fundamental Research Funds of Heilongjiang Province, China (No. F2016024).