基于改进CycleGAN的视频监控人脸超分辨率恢复算法  

Improved video surveillance face super-resolution recovery algorithm based on CycleGAN

作　　者：陈贵强 何军[1] 罗顺茺 Chen Guiqiang;He Jun;Luo Shunchong(College of Computer Science,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《计算机应用研究》2021年第10期3172-3176,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(U1836103);四川省科技重点研发项目(18ZDYF2039);四川省重大科技专项(2017GZDZX0002)。

摘　　要：针对有监督超分辨率算法训练过程需要大量成对图像、处理真实低分辨率图像视觉恢复效果差等问题,提出了一种基于改进CycleGAN的半监督算法Cycle-SRNet。首先,利用退化模型获得与真实低分辨率人脸相似的图像,用于训练网络参数;其次,通过重建模型恢复出具有真实效果的高分辨率人脸图像;最后引入感知损失函数保持人脸结构相似性,以更好地恢复面部特征。实验结果表明,该算法不需要成对的图像进行网络训练,在视觉效果上能够将模糊的视频监控低分辨率人脸图像恢复成清晰可辨的人脸图像,在FID、PSNR和SSIM指标上超越了SRCNN、SRGAN、CinCGAN等方法。Supervised super-resolution algorithms require plentiful paired-image for networks training,resulting in poor visual recovery effects when processing real-world low-resolution images.To solve this problem,this paper proposed an improved semi-supervised algorithm based on CycleGAN,donated as Cycle-SRNet.Firstly the degradation model generated images similar to real world low-resolution faces for network training,and then the reconstruction model restored high resolution face image.Furthermore,this paper introduced perceptual loss function to ensure the structural similarity of the faces so as to recover the facial features better.Experimental results show that the algorithm can restore fuzzy low-resolution video surveillance face images into recognizable face images,and it exceeds SRCNN,SRGAN and CinCGAN in terms of FID,PSNR and SSIM indicators.

关 键 词：单幅图像超分辨率恢复 生成对抗网络 CycleGAN 半监督学习 人脸超分辨率 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]