基于全局感知的人脸属性编辑  

Facial attribute editing based on global perception

作　　者：顾广华[1,2] 刘畅 孙文星 窦轶阳 GU Guanghua;LIU Chang;SUN Wenxing;DOU Yiyang(School of Information Science and Engineering,Yanshan University,Qinhuangdao 066004,Hebei China;Hebei Key Laboratory of Information Transmission and Signal Processing,Qinhuangdao 066004,Hebei China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]河北省信息传输与信号处理重点实验室,河北秦皇岛066004

出　　处：《华中科技大学学报（自然科学版）》2024年第11期117-124,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62072394);河北省自然科学基金资助项目(F2024203049);河北省重点实验室资助项目(202250701010046)。

摘　　要：针对正确编辑人脸属性的同时保留其他属性的细节的问题,提出一种端到端的基于全局感知人脸属性编辑模型——基于全局感知的生成对抗网络(GP-GAN).编码器设计全局感知模块来代替传统的卷积编码器,使其同时具有全局感知能力和局部感知能力,进而更好地编辑全局属性和局部属性;同时,针对属性解耦问题提出一种具有多尺度属性特征融合模块和多层感知机转换器的属性转换器.多尺度属性特征融合模块可以在潜在空间中生成丰富的特征向量,为属性解耦提供条件;多层感知机转换器将人脸属性特征编码分为属性相关和属性不相关两部分,并施加正交约束保证属性解纠缠.实验结果表明:本研究模型可以在保留其他属性细节的条件下更好地编辑人脸属性,使生成的图像更加自然真实、质量更高.Aiming at the issue of correctly editing facial attributes while preserving the details of other attributes,a end-to-end facial attribute editing model—global perception generative adversarial networks(GP-GAN)was proposed based on global perception.The encoder designed a global perception module to replace the traditional convolutional encoder,so that it had both global perception and local perception,and could better edit global and local attributes.Meanwhile,for the problem of attribute disentanglement,an attribute transformer with multi-scale attribute feature fusion module and multilayer perceptron transformer was proposed.The multi-scale attribute feature fusion module could generate rich feature vectors in the potential space,providing conditions for attribute disentanglement.The multilayer perceptron transformer divided the facial attribute feature coding into two parts,which were related attribute and unrelated attribute,and applied orthogonal constraints to ensure the attribute disentanglement.Experiment results show that the proposed model can better edit facial attributes while retaining other attribute details,so that the generated image is more natural and real with higher quality.

关 键 词：人脸属性编辑 多层感知机 全局感知 局部感知 多尺度特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]