Face-based biometric systems with deep learning in non-homogenous settings

The modern performance of such facial verification / re-identification systems has made them ubiquitous: from being embedded in smart devices to operating at the nation-state / population level. State-of-the-art models require millions of images for training, and the acquisition of that data can be time-consuming / logistically difficult, and in the specific case of facial images for use in surveillance settings, fraught with potential ethical and legal violations.<br><br>This thesis therefore focuses on analysing and reducing the performance gap between face-based biometric experiments, and real-world systems which often perform worse than the literature would otherwise suggest.

This is attempted holistically, with an eye to lessening the aforementioned logistical and ethical issues throughout.<br><br>Firstly, the performance of some typical deep learning models trained for facial verification / re-identification is investigated with regards to real-world cross-domain / multi-modal scenarios. This is achieved by contributing a novel dataset designed for the benchmarking of such systems. Data augmentation techniques, including the incorporation of synthetically generated data, are then explored for the purpose of mitigating the performance penalty when operating in cross-domain / multi-modal settings.<br><br>Secondly, domain adaptation techniques are employed to overcome the domain gap in non-homogeneous settings.<br>Initially, this is achieved by utilising domain transfer techniques during the post-training verification / re-identification stage, leading to a 12% performance improvement over baseline.