Template update algorithms and their application to face recognition systems in the deep learning era

Biometric technologies and facial recognition systems are reaching a very high diffusion for authentication in personal devices and public and private security systems, thanks to their intrinsic reliability and user-friendliness. However, although deep learning-based facial features reached a significant level of compactness and expressive power, the facial recognition performance still suffers from intra-class variations such as ageing, different facial expressions, different poses and lighting changes. In the last decade, several "adaptive" biometric systems have been proposed to deal with this problem. Unfortunately, adaptive methods usually lead to a growth of the system in terms of memory and computational complexity and involve the risk of inserting impostors among the templates. The first goal of this PhD thesis is the presentation of a novel template-based self-update algorithm, able to keep over time the expressive power of a limited set of templates. This classification-selection approach overcomes the problem of manual updating and stringent computational requirements. In the second part of the thesis, we analyzed if and to what extent this "optimized" self-updating strategy improves the facial recognition performance, especially in application contexts where the facial biometric trait undergoes great changes due to the passage of time. In contexts of long-term use, in fact, the high representativeness of the deep features may not be enough and this is usually overcome with a re-enrollment phase. For this reason, one of our goals was to evaluate how much an automatic template updating system could compete with human-in-the-loop in terms of performance. To simulate situations of long-term use in which the temporal variability of biometric data is high, we acquired a new dataset collected by using frames of some videos in YouTube related to Daily Photo Projects: people take a picture every day for a certain period of time, usually to show how their appearance is changing. The temporal information present in this new dataset allowed us to evaluate how long a facial feature can remain representative depending on the context and the recognition system. Extensive experiments on different datasets and using different facial features are conducted to define the contexts of applicability and the usefulness of adaptive systems in the deep learning era.