基于StyleGAN的可编辑高清人脸生成方法

Editable High-Definition Face Generation Method Based on StyleGAN

人脸由于身份认证的特殊性,是众多计算机领域的研究对象。人脸生成任务不但可以扩展人脸数据集,还有广阔的商用价值。随着生成式对抗网络的蓬勃发展,人脸生成任务逐步转向了高清的人脸生成。本文针对人脸生成任务里的高分辨率人脸定向生成的子任务,提出了三种方法,分别面向易划分二分平面属性类别(如性别)的人脸定向生成、抽象属性类别(如黄种人、明星脸)的人脸定向生成和人脸属性编辑三种使用场景,最终实现了基于StyleGAN的可编辑高清人脸的定向生成模型,并使用CelabA、FFHQ两种开源数据集验证了模型的可靠性。

Due to the particularity of identity authentication,human face is used as the research object in many computer fields.Face generation task can not only expand the face data set,but also have broad commercial value.With the rapid development of generative adversarial networks,the task of face generation has gradually shifted to high-definition face generation.In this paper,three methods are proposed for the sub-task of high-resolution face orientation generation in the face generation task.They are respectively for the face orientation generation of easily-divided binary plane attribute categories(such as gender category),the face orientation generation of abstract attribute categories(such as yellow race,star face)and face attribute editing.Finally,an editable high-definition face oriented generation model based on StyleGAN is implemented,and the reliability of the model is verified by CelabA and FFHQ open source datasets.