预训练网络引导的人脸图像超分辨率重建  

Super-Resolution Reconstruction of Face Images Guided by Pre-Training Network

作　　者：徐文博 孙广玲[1] 陆小锋[1] 

机构地区：[1]上海大学通信与信息工程学院,上海201900

出　　处：《工业控制计算机》2020年第6期36-38,共3页Industrial Control Computer

摘　　要：成像设备分辨率的不足或者成像距离较远等因素,导致获取的图像中人脸区域的分辨率较低,这样显然不利于后续对于该人脸的身份识别、表情识别、属性分析等高层任务。为此,提出预训练网络引导的人脸超分辨率重建方法,并且在人脸身份认证的任务上验证所提方法的有效性。具体地,使用残差结构的超分辨率重建网络Super-resolution Reconstruction Network(SRN)和预训练网络FaceNet形成一个用于图像超分辨率重建的训练框架。高分辨率图像和仿真的低分辨率图像构成有监督训练数据集;SRN网络接收低分辨率图像,输出超分辨率重建图像;FaceNet提供高分辨率和超分辨率重建图像的特征,以建立损失函数,用于引导SRN网络的监督学习。在测试阶段,SRN网络接收低分辨率图像,输出的超分辨重建图像用于解决身份认证问题。仿真实验结果表明,该方法对于提升低分辨率图像的人脸认证任务的性能,具有一定的优越性。This paper proposes a pre-training network-guided face super-resolution reconstruction method,and verify the effectiveness of the proposed method on the task of face identity verification.Specifically,the Super-resolution Reconstruction Network(SRN)using the residual structure and the pre-training network FaceNet form a training framework for image superresolution reconstruction.High-resolution images and simulated low-resolution images constitute a supervised training data set.SRN network receives low-resolution images and outputs super-resolution reconstructed images.FaceNet provides features of high-resolution and super-resolution reconstructed images to establish losses Function,used to guide the supervised learning of SRN network.In the test phase,the SRN network receives low-resolution images,and the super-resolution reconstructed images are used to perform the identity verification task.Empirical experiment results show that our method has certain advantages for improving the performance of face verification tasks for low-resolution images.

关 键 词：低分辨率 超分辨率重建 预训练网络 人脸认证 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]