Contribution to facial authentication and synthetic datasets for Edge-AI applications
Yao, Wang
Yao, Wang
Loading...
Repository DOI
Publication Date
2024-09-27
Type
doctoral thesis
Downloads
Citation
Abstract
Secure authentication of low-power consumer devices such as doorbell cameras is a unique and significant challenge. This challenge arises from the complexity of the facial images captured by these devices, which encompass different lighting conditions, various postures, and individuals of different ages. These factors collectively present hurdles to the accurate recognition capabilities of these devices. The development of robust face recognition systems requires a broad and diverse dataset containing images depicting a variety of scenarios such as various lighting conditions, head pose variations, and age. However, current publicly available datasets are insufficient to meet this demanding standard. Moreover, the process of building such a comprehensive dataset is challenging because it requires laborious and time-consuming efforts. The introduction of stringent data protection regulations, such as GDPR in Europe, further makes this task more complex by introducing additional compliance and restrictions.
To address these challenges, this work explores the use of synthetic data as a viable solution. The first endeavor of this study involves quantifying and compensating for the impact of lighting and pose on the performance of the facial recognition system by introducing synthetic images generated through GAN-based portrait relighting and head pose generation algorithms. Next, we quantified the performance of facial recognition algorithms across different age groups and age intervals. Further, synthetic age images were introduced, evaluated, and utilized to compensate for the performance of facial recognition across age intervals. Finally, motivated by the DAVID smart-toy project, we have investigated the use of different generative models and designed GAN-based and Diffusion-based models to generate photo-realistic child images. The work presented in this dissertation overcomes the unavailability of real data by employing synthetic data as a data augmentation technique, investigating the use of synthetic data to improve the performance of face recognition and generate high-definition synthetic face images of children. Extensive experiments reveal that while current synthetic data still exhibits deficiencies in comparison to real data, its efficacy as a data augmentation technique remains significant.
Funder
Publisher
University of Galway
Publisher DOI
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International