Disentanglement Learning of Emotions and Identities from Facial Image

Abstract

Learning disentangled representations of emotions and identities from facial images is important for understanding human faces. This is because the human face consists of emotions and identities, and the emotions affect the face as expressions. However, because these factors interact with each other, it is difficult to learn to decompose them from facial images. Such entangled representations lead to failure in facial recognition and facial image synthesis tasks. Therefore, we aim to disentangle emotions and identities from the facial image. To achieve this goal, we propose a novel method for learning disentangled representations of emotions and identities. In the experiment, we confirmed that our method achieves to learn disentangled representation of emotions and identities, and allows us to perform an emotion-controllable image synthesis by swapping two facial images.