A Hybrid Network for Facial Age Progression and Regression Learning

Facial age transformation is an attractive application on an entertainment or amusement robot. With this application, the robot can transform an input face to the same face but in different ages. We propose a new algorithm for age transformation. Due to recent progresses made by state-of-theart deep learning approaches, the facial age progression and regression has become an attractive research topic in the fields of computer vision. Many existing approaches require paired data which refer to the face images of the same person at different ages. As the cost of collecting such paired datasets is expensive, some emerging approaches have been proposed to learn the facial age manifold from unpaired data. However, the images generated by these approaches suffer from the weakness or loss in generating some age traits, for example wrinkles and creases. We propose a hybrid network that is composed of a generator and two discriminators. The generator is trained to disentangle the age from the identity of the face so that it can generate a face of the same identity as of the input face but at a different age. One of the discriminator is designed for handling multitasks, including the identification of real vs. fake (generated) faces and the classification of the identities and ages of the faces. The other discriminator is designed to make the latent space satisfy the requirement so that the generated image can be made more realistic. Experiments show that the proposed network can generates better facial age images with more age traits compared with other state-of-the-art approaches.