Sveinn Pálsson

Semester Work
Supervisors: Eirikur Agustsson, Dr. Radu Timofte

Generative Adversarial Style Transfer Networks for Face Ageing

What will you look like in 10 years? Or how could that missing person look like today? In this thesis we look at the problem of face aging which relates to rendering an image of a face to change its apparent age. This task involves synthesizing images and modelling the aging process which both are problems that have recently enjoyed much research interest in the field of computer vision. We propose to look at the problem from the perspective of image style transfer, where we think of the age of the person as the underlying style of the image. We propose a novel learning-based method for face aging where we directly involve a pre-trained differentiable face age estimator in the loss function. We build our method on CycleGAN, a general purpose framework for unsupervised image-to-image translation between two image domains. In this thesis we as well apply CycleGAN directly to face aging with predefined age groups. We see that it can be very effective in learning the aging process as well as generating high quality images but this approach is limited to only large age groups. Our proposed method addresses this issue and is shown to be able to produce smaller aging effects. It is however limited by quality of the age estimator. We compare our results with some state-of-the-art face aging techniques and show that both CycleGAN and our method achieve competitive results.