Age Progression using Generative Adversarial Networks

Abstract

This study presents a technique to generate face age progression by adopting a conditional generative adversarial network based approach. The best model resulting from a five-fold cross validation has an accuracy of 91.93%, False Omission Rate of 0.45% and Negative Prediction Value of 99.55%. Building on prior work, this paper has three contributions. First, the use of uneven age clusters is presented to account for more rapid and drastic ageing in babies and toddlers than older individuals. Second, perceptual losses rather than per-pixel losses are considered to enable identity preservation. Third, a facial recognition system is applied to verify the identity of individuals upon ageing. Identity preservation was achieved and confirmed, with a facial recognition accuracy of 92.4%. Visual fidelity was also confirmed, with 95.2% subjects identifying ageing in the conducted survey.