Unsupervised Facial Image Synthesis Using Two-Discriminator Adversarial Autoencoder Network

Recent years have witnessed the unprecedented success in single image synthesis by the means of convolutional neural networks (CNNs). High-level synthesis of facial image such as expression translation and attribute swap is still a challenging task due to high non-linearity. Previous methods suffer from the limitations that being unable to transfer multiple face attributes simultaneously, or incapability of transferring an attribute to another by a continuously changing way. To address this problem, we propose a two-discriminator adversarial autoencoder network (TAAN). The latent-discriminator is trained to disentangle an input image from its original facial attribute, while the pixel-discriminator is trained to make the output image attach to the target facial attribute. By controlling the attribute values, we can choose which and how much a specific attribute can be perceivable in the generated image. Quantitative and qualitative evaluations are conducted on the celebA and KDEF datasets, and the comparison with the state-of-the-art methods shows the competency of our proposed TAAN.