非对称周期推理循环渐进的人脸修复算法研究  

Asymmetric periodic inference cyclic progressive face completion algorithm research

作　　者：李雅倩[1] 张旭曜 李岐龙 Li Yaqian;Zhang Xuyao;Li Qilong(School of Electrical Engineering,Yanshan University,Qinhuangdao Hebei 066004,China)

机构地区：[1]燕山大学电气工程学院,河北秦皇岛066004

出　　处：《计算机应用研究》2022年第8期2496-2499,2531,共5页Application Research of Computers

基　　金：国家自然科学基金项目(62106214);河北省自然科学基金项目(F2019203195)。

摘　　要：针对生成对抗网络中修复网络无法兼顾图像的全局一致性和局部一致性,且计算负载较大的问题,在非对称U-Net网络架构的基础上引入渐进修复的思想。首先,提出了非对称周期特征推理模块,增加图像修复内容与周围已知像素之间的关联性,提高了修复图像的全局一致性表现;其次,提出新型的U-Net结构生成器网络,避免了编码器中的未知像素进入解码器,从而破坏解码器中特征的问题;最后,引入了感知损失和风格损失,进而提高了网络在主观评价下的修复效果。在人脸图像数据集上的实验表明,该算法在主观视觉效果和客观指标上都有显著的提高。To address the problem that the inpainting network in generative adversarial networks cannot take into account the global consistency and local consistency of images and has a large computational load,this paper introduced the idea of incremental inpainting based on the asymmetric U-Net network architecture.Firstly,this paper proposed an asymmetric periodic feature inference module to increase the correlation between image inpainting content and surrounding known pixels,which improved the global consistency performance of restored images.Secondly,this paper presented a novel U-Net structured generator network to avoid the problem of unknown pixels in the encoder entering the decoder and thus corrupting the features in the decoder.Finally,it introduced perceptual loss and style loss,which improved the inpainting effectiveness of the network under subjective evaluation.Experiments on the face image dataset show that the proposed algorithm shows significant improvements in both subjective visual effects and objective metrics.

关 键 词：生成对抗网络 渐进式修复 非对称周期特征推理 图像修复 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]