人脸超分辨率网络(FSRNet)的改进  

Improvement of Face Super Resolution Network(FSRNet)

作　　者：段燕飞 王瑞祥 咬登国 张航 DUAN Yanfei;WANG Ruixiang;YAO Dengguo;ZHANG Hang(College of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学软件工程学院,成都610225

出　　处：《重庆科技学院学报（自然科学版）》2021年第3期89-94,共6页Journal of Chongqing University of Science and Technology：Natural Sciences Edition

基　　金：国家自然科学基金项目“新一代产品几何量技术规范的范畴纤维化模型及智能专家系统的研究与开发”(61203172);四川省科技计划项目“机器视觉深度学习技术在猕猴桃分选中关键技术研究与应用”(20ZDYF0008)。

摘　　要：人脸超分辨网络(FSRNet)使用人脸几何先验信息优化人脸超分辨率,可以从低分辨率人脸图像生成逼真的高分辨率人脸图像,但FSRNet生成的超分辨率图像存在伪影。对其关键模块进行了改进,并引入了新的损失函数。直接输入16×16像素的低分辨率图像,最后使用转置卷积函数放大图像,降低了计算复杂度,提升了粗略超分辨网络的性能。通过两步训练法,解决网络训练时调参困难的问题。引入热图损失、面部注意力损失和对抗性损失训练,提高超分辨率人脸图像的质量。实验结果证明,采用改进后的方法,可以生成面部细节更加清晰的高质量人脸图像。Face Super Resolution Network(FSRNET)uses face geometric prior information to optimize face super-resolution,and can generate realistic high-resolution face images from low-resolution face images.However,there are artifacts in FSRNET super-resolution images.The key modules are improved and a new loss function is introduced.The low resolution image of 16×16 pixels is directly input,and the transpose convolution function is used to enlarge the image,which reduces the computational complexity and improves the performance of the rough super-resolution network.Through the two-step training method,the difficulty of adjusting parameters in network training is solved.Heat map loss,facial attention loss and confrontational loss training are introduced to improve the quality of super-resolution face images.Experimental results show that the improved method can generate high quality face images with clearer facial details.

关 键 词：人脸超分辨率网络 注意力损失 先验信息 生成对抗网络 神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]