Single sample face recognition using deep learning: a survey

Face recognition has become popular in the last few decades among researchers across the globe due to its applicability in several domains. This problem becomes more challenging when only a single training image is available and is popularly known as single sample face recognition (SSFR) problem. SSFR becomes even more complex when images are captured under varying illumination conditions, different poses, occlusion, and expression. Further, deep learning methods have shown performance at par with humans recently. Due to the emergence of deep learning methods in the last decade, it has been made possible to recognize faces with excellent accuracy even in a single sample scenario. In this paper, we present a comprehensive survey of SSFR using deep learning. We also propose a novel taxonomy and broadly divide these methods into three categories viz. virtual sample generation, feature-based, and hybrid methods. Performance comparison of these methods as reported in the literature has also been performed. Finally, we review publicly available databases used by the researchers and give some important future research directions which will help aspiring researchers in this fascinating area.

Automated summary

Face recognition is widely applied in various domains, especially in the challenging scenario of single sample face recognition problem. The complexity increases when there are variations in illumination, pose, occlusion, and expression. Deep learning methods have shown comparable performance to humans, enabling accurate recognition even with a single sample. This paper presents a comprehensive survey of single sample face recognition using deep learning, with a novel taxonomy dividing methods into virtual sample generation, feature-based, and hybrid methods. Performance comparison and future research directions are also discussed.