基于改进Pix2Pix网络的去口罩遮挡修复研究  

Research on Improved Pix2pix-Based Network for Unmasking of Masked Face

作　　者：吴雅琴[1] 侯云峰 陈林 WU Ya-qin;HOU Yun-feng;CHEN lin(School of Mechanical Electronic and Information Engineering,China University of Mining and Technology(Beijing),Beijing 100083,China)

机构地区：[1]中国矿业大学(北京)机电与信息工程学院,北京100083

出　　处：《计算机仿真》2023年第10期242-248,共7页Computer Simulation

基　　金：国家自然科学基金重点项目(52130409);中央高校基本科研业务费专项(2021YJSJD14)。

摘　　要：针对人脸去口罩遮挡修复难题,提出一种由SVR模型预测被遮挡的面部特征点,引导改进后的Pix2Pix网络进行图像修复的算法。训练了一个高精度SVR回归器,解决了现有模型因遮挡区域特征丢失而修复失败的问题。采用16层U-Net生成器,提高模型对深层特征的提取和还原能力;采用联合PatchGan判别器抑制噪声,提高图像的整体性;引入Smooth L1损失函数加强模型在训练后期的学习能力,提高了Pix2Pix模型的修复能力。实验表明模型修复效果接近真实图像,在PSNR和SSIM两项指标上分别提升了6.0%和9.2%。Aiming at the problem of unmasking masked faces in the image,missing coordinates in face 68 landmarks were predicted by SVR model,which can guide the improved pix2pix network for image inpainting.A highprecision SVR regressor was trained to solve the problem of inpainting failure of existing models due to insufficient features.A 16-layer U-Net generator was used to improve the model's ability to extract and restore deep features.The joint PatchGan discriminator was used to suppress noise and improve the integrity of the image.The Smooth L1 loss function was introduced to strengthen the learning ability of the model in the later stage of training,improving the inpainting ability of the pix2pix model.Experiments show that the model inpainting effect is close to the real image,and the PSNR and SSIM are improved by 6.0%and 9.2%.

关 键 词：去口罩遮挡 图像修复 面部特征点补全 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]