An alternative to face image representation and classification

  • PDF / 1,733,211 Bytes
  • 9 Pages / 595.276 x 790.866 pts Page_size
  • 117 Downloads / 211 Views

DOWNLOAD

REPORT


ORIGINAL ARTICLE

An alternative to face image representation and classification Qi Zhu1,2 · Ning Yuan1 · Donghai Guan1 · Nuoya Xu1 · Huijie Li1 Received: 30 May 2017 / Accepted: 26 February 2018 © Springer-Verlag GmbH Germany, part of Springer Nature 2018

Abstract Sparse representation has brought a breakthrough to the face recognition community. It mainly attributes to the creative idea representing the probe face image by a linear combination of the gallery images. However, for face recognition applications, sparse representation still suffers from the following problem: because the face image varies with the illuminations, poses and facial expressions, the difference between the test sample and training samples from the same subject is usually large. Consequently, the representation on the probe face image provided by the original gallery images is not competent in accurately representing the probe face, which may lead to misclassification. In order to overcome this problem, we propose to modify training samples to produce an alternative set of the original training samples, and use both of the original set and produced set to obtain better representation on the test sample. The experimental results show that the proposed method can greatly improve previous sparse representation methods. It is notable that the error rate of classification of the proposed method can be 10% lower than previous sparse representation methods. Keywords  Face recognition · Alternative face image · Sparse representation

1 Introduction In face recognition, the success of sparse representation methods (SRM) [1, 2] has attracted tremendous attention. Actually, SRM has been viewed as the breakthrough of face recognition [3]. Besides face recognition, SRM has also been applied to action recognition [4], image denosing [5], image super-resolution [6, 7], image recovery [8], video analysis [9], and illumination recovery [10] etc. The sparsest solution of SRM can be obtained by using the sparse representation method with l0 norm minimization [11, 12]. However, solving the linear optimization problem with l0 norm minimization is NP-hard problem [1, 2, 13]. Recently, the methods proposed in [14–16] have demonstrated that when the solution obtained using l1 norm minimization is sparse enough, it can be equivalent to the solution obtained by l0

* Donghai Guan [email protected] 1



College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China



Fujian Provincial Key Laboratory of Information Processing and Intelligent Control, Minjiang University, Fuzhou 350121, China

2

norm minimization. Besides these kinds of methods, various other SRMs such as SRMs with l2 norm minimization [17–19] and l2,1 norm minimization [20] are also available. l2,1 norm minimization was also used to improve discriminant analysis [21] and embedding learning [22]. Extensions of sparse representation such as kernel sparse representation [23, 24] and sparse component analysis [25–27] have been proposed. In this