基于高频小波损失的生成对抗人脸补全方法研究  

GENERATIVE ADVERSARIAL NETWORK FOR FACE COMPLETION BASED ON HIGH FREQUENCY WAVELET LOSS

作　　者：黄俊健 朱煜[1] 林家骏[1] 郑兵兵[1] 韩飞 Huang Junjian;Zhu Yu;Lin Jiajun;Zheng Bingbing;Han Fei(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学信息科学与工程学院,上海200237

出　　处：《计算机应用与软件》2021年第2期233-238,共6页Computer Applications and Software

基　　金：上海市科学技术委员会科研计划项目(17DZ1100808)。

摘　　要：基于生成对抗网络架构设计一种新的人脸补全模型。在生成网络中使用空洞卷积以增加特征图的感受野,提升网络性能;针对生成补全图像模糊,提出基于小波分解的损失函数设计方法,将图像转换到小波空间,提取高频信息作为l_1小波损失,有效提升人脸图像补全的质量。对VGGFace2人脸数据集下半部分人脸进行遮挡,作为训练数据集,以LFW数据集遮挡,进行人脸补全测试结果分析。实验结果表明,所设计算法的网络补全后的人脸结构相似性(SSIM)达到0.803 4,峰值信噪比(PSNR)达到20.946 7,有效提升了人脸补全的效果。This paper proposes a new face completion model based on the generative adversarial network architecture.The dilated convolution was used in the generation network to increase the receptive field of the feature map and improve the network performance.For the problem of generating the complement image blur,we proposed a loss function design method based on wavelet decomposition.The image was transformed into wavelet space,and the high frequency information was extracted as l 1 wavelet loss to improve the quality of face image completion effectively.We occluded the lower half of the VGGface2 to be a training dataset,and the LFW data set was occluded to analyze the results of face completion test.The experiment results show that the SSIM and PSNR reach 0.8034 and 20.9467 respectively.And this algorithm can effectively improve the effect of face completion.

关 键 词：人脸补全 生成对抗网络 空洞卷积 小波分解 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]