基于特征增强的人脸属性转换  

Face attribute transfer based on feature enhancement

作　　者：魏子涵 王慧[1,3] 王晶[1,3] 高岚[3,4] 张伟功 WEI Zi-han;WANG Hui;WANG Jing;GAO Lan;ZHANG Wei-gong(Beijing Advanced Innovation Center for Imaging Theary and Technology,Capital Normal University,Beijing,100048,China;College of Electronical and Information Engineering,Xidian University,Xian,710126,China;College of Information Engineering,Capital Normal University,Beijing 100048,China;Beiing Key Laboratory of Electronie System Reliability and Prognostics,Capital Normal University,Beiing 100048,China)

机构地区：[1]首都师范大学北京成像理论与技术高精尖创新中心,北京100048 [2]西安电子科技大学电子工程学院,陕西西安710126 [3]首都师范大学信息工程学院,北京100048 [4]电子系统可靠性技术北京市重点实验室,北京100048

出　　处：《微电子学与计算机》2020年第10期38-41,47,共5页Microelectronics & Computer

基　　金：国家自然科学基金(61772350);北京市高水平教师队伍建设计划(CIT&TCD201704082);北京市科技新星计划项目(Z181100006218093);北京未来芯片技术高精尖创新中心科研基金资助项目(KYJJ2018008)。

摘　　要：采用生成对抗网络及其扩展模型可以实现人脸属性转换,但生成图像存在分辨率低及特征损失的问题.本文基于GeneGAN网络,提出一种基于特征增强的人脸属性转换算法,在生成网络中增加网络深度,建立高低维特征耦合通道,通过对特定属性深度特征提取比较,优化生成图像的身份及属性特征.在CelebA数据集上的实验表明,本文算法得到的高分辨率图像不仅在保留图像身份信息上有较好的改善,同时在视觉效果和客观指标上都有较大的提高.The current facial attribute transformation based on generative adversarial networks(GAN)and GAN-based method has problems of low resolution and feature loss of the generated image.Based on the GeneGAN network,this paper proposes a face attribute conversion algorithm based on feature enhancement.By increasing the network depth in the generation network,establishing a high-and low-dimensional feature coupling channel,and extracting and comparing specific attribute depth features to optimize the identity and features of the generated image.Experiment based on CelebA dataset shows that the high-resolution images generated by the proposed algorithm have a better improvement in maintaining image identity information,and have a greater improvement in visual effects and objective indicators.

关 键 词：生成对抗网络 人脸属性转换 特征增强 深度特征 反向传递 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]