Hierarchical neural network for facial attribute transfer. Issue 7 (March 2020)

• Record Type: Journal Article
• Authors: Guo, You; Shang, Yanlei; Zhang, Xiaofang
• Is Part Of: IOP conference series. Volume 768:Issue 7(2020)
• Publication Date: 2020-03
• Subjects: Materials science -- Periodicals; 620.1105
• DOI: 10.1088/1757-899X/768/7/072088
• Languages: English
• ISSNs: 1757-8981

Abstract

Abstract: The facial attribute transfer refers to generating a face image with desired attributes while preserving other details. In the existing methods, some of the them consider the independence among attributes but neglect the integrity. It may result in information loss and lead to distorted generation. The others ensure the integrity at the cost of insufficient independence limits. The generated images will be partially blurred. In this paper, GAN and Variational Autoencoders (VAE) structure are incorporated to preserve the details while attribute transfer occurs. The concept of hierarchical latent representation is introduced to realize the attribute independence. These methods work with each other forming an effective network for facial attribute transfer, referred as Hi-GAN. Experiments on the CelebA dataset show that our model can output clearer and more realistic images on facial attribute transfer.