On the Development of Face Recognition using Synthetic Data and Unsupervised Representation Learning

Current face recognition (FR) model development has benefited from recent advances in training deep learning models and is based on the availability of large-scale identitylabeled training datasets. However, due to increasing ethical and legal concerns, many of these large-scale datasets are retreated by their creators. Hence, employing privacyfriendly synthetically generated data as an alternative to real-world data for FR model development has recently attracted increasing attention in the research community. Yet, current synthetic datasets proposed in the literature suffer from either low real-world variation, are limited in variation to specific attributes, or lack identity separability. This thesis proposes two contributions to evade the need for large-scale identity-labeled synthetic training datasets. First, this thesis proposes an unsupervised face recognition training paradigm based entirely on privacy-friendly synthetic data (USynthFace). Distinctive representations are learned by enhancing the similarity of two different views of the same instance. Since this concept depends mainly on augmentations to create different views, extensive color, geometric, and GAN-based augmentations are considered. Furthermore, detailed sensitivity studies on all of USynthFace components are provided. USynthFace shows competitive performance on several benchmarks compared to recent state-of-the-art (SOTA) FR models trained on GAN-based synthetic data. Second, this thesis proposes a concept for disentangled identity representations of GANs’ latent space for identityspecific face image generation (DIRGAN). The latent space is disentangled by identifying a boundary for each unconditional latent code. Each boundary divides the latent space into two identity classes from which new latent codes can be sampled. Applying DIRGAN to SOTA generator models can create datasets that tend to have real-world variation while maintaining the highest identity separability in comparison to recent SOTA conditional identity generative models. High identity separability is achieved, although attribute variation is not restricted to a predefined set of attributes. FR models achieve competitive results on several benchmarks when trained on these newly created datasets compared to GAN-based synthetic FR models. On some benchmarks, a DIRGAN-based FR model 3 even surpasses previous best performances, reaching, for example, 72.85% verification accuracy on AgeDB-30 and 78.60% verification accuracy on CA-LFW.