基于属性法向量修改的人脸编辑算法

Facial image editing through manipulating attribute normal vector

根据用户需求修改给定人脸图像,在美妆造型设计、异质人脸合成等领域具有重要的应用价值.目前已有研究基于生成对抗网络(GAN)模型实现人脸的语义属性迁移或者定性修改,但少有算法能够对用户指定的属性进行定量修改.针对人脸多个属性的分离(解纠缠)以及定量编辑两大问题,提出了一个基于属性法向量修改的定量人脸属性编辑算法.首先,端对端求解出隐藏空间中每一个面部属性的最优控制法向量,实现人脸多属性的解纠缠;然后通过调整法向量的移动系数,控制单个属性的变化程度从而控制该属性的定量修改;最后通过融合多个单属性控制向量的方式,达到对人脸多个属性的综合修改效果.实验结果表明,该算法能够同时适用于生成的虚拟人脸图像和真实人脸照片,在保持生成图像质量的同时,实现了人脸属性定量修改的功能.

Modification of given facial images according to users'demands has secured wide applications in areas such as makeup design,criminal portrait.Current generative adversarial network(GAN)based algorithms can transfer or qualitatively change semantic attributes,but can hardly attain the quantitative change of user-specified attributes.Aiming at these problems of attributes separation(entanglement)and quantitative editing,we propose a quantitative facial attribute-editing algorithm based on attribute normal vector modification.First,the optimal control vector corresponding to each attribute is solved end-to-end in the hidden space,and each attribute is entanglement.Then,by adjusting the displacement factor of normal vector,the variation degree of single attribute is controlled.Finally,through fusing multiple control vectors,the comprehensive modification effect of changing hybrid facial attributes is obtained.Experimental results show that the proposed method can be applied to quantitatively editing both the composite and real face images,while maintaining the quality of resulting images.