Face Aging Using Generative Adversarial Networks

Abstract

AbstractWhen computationally aging a face, it is desirable that the age output is close to the expected age and the individual’s characteristics are maintained. Traditionally, there have been two kinds of modeling techniques used in this task: prototype-based and model-based methods. The first calculates the mean difference between age groups, and the latter uses parametric models to simulate change over time. Both approaches fail to keep individual characteristics when transforming a face from a younger domain to an aged one. With advances in computer vision, generative models have been applied to perform this task, especially generative adversarial networks (GANs). With them, it becomes possible to generate realistically aged faces of specific individuals. These networks can encode latent information, by enabling a generator model to create new images conditional on an input image. This could lead to improvements in biometric systems, assist in the search for missing people and in the identification of criminals in an automated way. In addition, generative models are currently used in multiple applications in entertainment. In the last decade, an increasing number of publications focused on applying state-of-the-art generative models on facial aging. Most of these works were done by using a general architecture and customizing it to fit the aging problem’s needs. This work enumerates the most frequent architectures, some of the challenges involved in the task, the available benchmark databases, and their applications. Additionally, it compares different face aging architectures, by using the three most frequent databases: FG-NET, UTKFaces, and CACD. Finally, age estimation and face verification techniques were used to measure the effectiveness of the resulting faces.