Deep learning (DL) methods, and in particular convolutional neural networks (CNNs), provide the most used and most effective face recognition systems. However, when “each identity has only a single sample available for training,” performance drastically drops.
The authors of this survey consider, in Section 2, how DL solutions are applied to the single-sample face recognition problem. They categorize the methods as either virtual sample based, which work by generating virtual images or virtual features to improve training, or generic learning methods, which exploit multi-sample sets using diverse strategies as deep features for machine learning, new loss functions, or improved network structures.
Data is relevant to DL methods that need large quantities of training data. Four popular datasets (AR, Extend Yale B, LFW, and MS-Celeb-1M) containing up to a million faces under different conditions, created for face recognition, can be used in single-sample recognition. Those datasets provide benchmarks for comparing the performance of models, as reported in Section 3.
The discussion in Section 4 points out concerns about possible limitations and risks. Identity information preservation when using virtual sample methods can be lost. Another concern relates to domain adaption when applying generic learning methods, which are trained on generic datasets containing images taken in very different contexts. Finally, a solution for single image face recognition can arrive from unsupervised methods. There is also a methodological concern about the nature of the DL models, which work differently from human vision and seem to not use semantic information.
Even though the reported accuracy of today’s methods can be as high as 98 percent, not a clear best choice emerges. The discussion offers ideas for improvements. Indeed, this paper is for researchers. Readers new to the field should consider additional materials in order to understand Section 2.