Multi-expression Generative Adversarial Networks for Facial Expression Synthesis

Abstract

Facial expression synthesis has always been a research hotspot in the field of computer vision and graphics. The facial expressions are complex and vary from person to person. It is a challenging task to synthesize a rich and diverse facial expression. In this paper, we propose a novel network framework: Multi-Expression Generative Adversarial Network (MEGAN). We combine automatic encoders with generative adversarial networks, and innovatively integrate labels into encoders and decoders. With just a set of networks and a discriminator, we can synthesize faces with many different expressions in an orderly fashion. In addition, we use the encoder to find a low-dimensional representation of the facial image in the potential space, and to transfer the expression under the condition that most facial features are preserved. We don't need pairs of face images, and huge amounts of data. The results of the later experiments can show that the framework we propose is able to achieve better quality and more realistic results in facial expression synthesis.