Deep learning for facial recognition on single sample per person scenarios with varied capturing conditions

Single sample per person verification has received considerable attention because of its relevance in security, surveillance and border crossing applications. Nowadays, e-voting and bank of the future solutions also join this scenario, opening this field of research to mobile and low resources devices. These scenarios are characterised by the availability of a single image during the enrolment of the users into the system, so they require a solution able to extract knowledge from previous experiences and similar environments. In this study, two deep learning models for face recognition, which were specially designed for applications on mobile devices and resources saving environments, were described and evaluated together with two publicly available models. This evaluation aimed not only to provide a fair comparison between the models but also to measure to what extent a progressive reduction of the model size influences the obtained results.The models were assessed in terms of accuracy and size with the aim of providing a detailed evaluation which covers as many environmental conditions and application requirements as possible. To this end, a well-defined evaluation protocol and a great number of varied databases, public and private, were used