Example-Guided Identify Preserving Face Synthesis by Metric Learning

Abstract

Generative adversarial networks (GANs) are commonly applied to example-guided identify preserving face synthesis. A binary classifier is used as style consistency discriminator in GAN model in order to ensure the consistency of style. However, the over-fitting problem of a binary classifier downgrade its discrimination ability on style consistency. In this paper, we propose a style consistency discriminator based on metric learning, which performs better in keeping identity information and guaranteeing consistency in style between input examplar and result. Through separating the positive pairs form the negative, metric learning model can efficiently measure the similarity between the synthesis face and the genuine face. The experimental results indicate that the metric learning performs better than a binary classifier in terms of preserving style consistency.