Singular value decomposition based virtual representation for face recognition

• Published in: Multimedia tools and applications Vol. 77; no. 6; pp. 7171 - 7186
• Main Authors: Zhang, Guiying, Zou, Wenbin, Zhang, Xianjie, Zhao, Yong
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.03.2018; Springer Nature B.V
• Subjects: Classification; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Face recognition; Facial recognition technology; Mathematical analysis; Matrix algebra; Matrix methods; Multimedia Information Systems; Representations; Singular value decomposition; Special Purpose and Application-Based Systems; Training; Sparse representation; Collaborative representation; Face recognition; Virtual samples; Singular value decomposition (SVD); Image classification
• DOI: 10.1007/s11042-017-4627-8
• ISSN: 1380-7501

Sparse representation, which uses a test sample to represent a linear combination of an entire set of training samples, has achieved great success in face recognition, and it results in good performance when sufficient training samples exist. However, the available number of images of a subject’s face is usually limited in real face recognition systems. In this paper, to obtain more facial representations, we propose a novel method that applies singular value decomposition (SVD) to produce virtual images from original images. The obtained virtual images not only enlarge the size of the set of training samples but also represent relatively stable low frequency facial information; thereby improving the robustness and classification accuracy. We also integrate these virtual samples with the original samples, providing more available information for object classification and, consequently, achieving better performance. To the best of our knowledge, this paper is the first work to use the product of a singular value matrix and right singular vectors to generate virtual samples for face recognition. Experiments on the most widely used and challenging benchmark datasets demonstrate that our method obtains better accuracy and is more robust compared with previous methods.