Wasserstein Divergence GAN With Cross-Age Identity Expert and Attribute Retainer for Facial Age Transformation

Abstract

We propose the Wasserstein Divergence GAN with an identity expert and an attribute retainer for facial age transformation. The Wasserstein Divergence GAN (WGAN-div) can better stabilize the training and lead to better target image generation. The identity expert aims to preserve the input identity at output, and the attribute retainer aims to preserve the input attribute at output. Unlike the previous works which take a specific model for identity and attribute preservation without giving a reason, both the identity expert and the attribute retainer in our proposed model are determined from a comprehensive comparison study on the state-of-the-art pretrained models. The candidate networks considered for identity preservation include the VGG-Face, VGG-Face2, LightCNN and ArcFace. The candidate backbones for making the attribute retainer are the VGG-Face, VGG-Object and DEX networks. This study offers a guidebook for choosing the appropriate modules for identity and attribute preservation. The interactions between the identity expert and attribute retainer are also extensively studied and experimented. To further enhance the performance, we augment the data by the 3DMM and explore the advantages of the additional training on cross-age datasets. The additional cross-age training is validated to make the identity expert capable of handling cross-age face recognition. The performance of our approach is justified by the desired age transformation with identity well preserved. Experiments on benchmark databases show that the proposed approach is highly competitive to state-of-the-art methods.

Keywords

• Generative adversarial network
• facial age transformation
• face recognition