利用属性控制的人脸图像修复  

Face Image Inpainting Using Attribute Guided

作　　者：张繁[1,2] 叶凯威 王鹿鸣 刘泽润 王章野[2] Zhang Fan;Ye Kaiwei;Wang Luming;Liu Zerun;Wang Zhangye(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023;State Key Laboratory of CAD&CG,Zhejiang University,Hangzhou 310058;College of Art and Design,Zhejiang University of Science&Technology,Hangzhou 310023)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023 [2]浙江大学CAD&CG国家重点实验室,杭州310058 [3]浙江科技学院艺术设计学院,杭州310023

出　　处：《计算机辅助设计与图形学学报》2022年第7期1085-1094,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家重点研发计划(2017YFB1002703);国家自然科学基金(61772456,U1736109).

摘　　要：人脸图像修复在刑事侦查、医疗美容、安防等领域有着重要的应用价值,但是传统已有工作大多基于扩散和纹理方法,均未有效地利用人脸图像已有的属性知识和语义信息,图像真实性和语义不一致问题较为突出.因此,提出一种基于属性控制的人脸图像修复方法.首先通过改进生成对抗网络,加入额外人脸属性信息以及分类网络,实现了在修复人脸图像的同时控制人脸面部特征;然后加入基于多尺度特征融合的注意力模块,在几乎不增加复杂度的同时通过增大神经网络的感受野获得更多全局特征,以提升人脸图像修复质量.结合MAE,PSNR,SSIM和FID这4项评价指标,在CelebA数据集上的实验结果表明,本文方法通过有效的语义引导,不仅可以保证人脸图像修复质量,还可以基于属性修改人脸面部特征.Face image inpainting is becoming more and more significant in the practical fields of criminal in-vestigation,cosmetic plastic surgery,and security protection.But most of traditional existing work is based on diffusion and texture methods,failing to effectively use the embedding attribute knowledge and semantic in-formation of face images.It is difficult to ensure the authenticity and preserve semantic consistency at the same time.To tackle this problem,a face image inpainting method based on attribute control is proposed.First,a modified generative adversarial network is proposed to restore face images and control facial features by add-ing additional face attribute information and a classification network.Then an attention module based on multi-scale feature fusion is proposed,through increasing the receptive field of the neural network to obtain more global features to improve face image inpainting quality without increasing the arithmetic complexity.Finally,combining the four evaluation metrics,MAE,PSNR,SSIM,and FID,the experiments on the CelebA dataset show that the proposed method can not only generate high-quality face image inpainting results steered by semantic information,but also generate corresponding facial features based on attribute control.

关 键 词：人脸图像修复 属性控制 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]