基于卷积神经网络的人脸超分辨率重建  

Face super-resolution reconstruction based on convolutional neural network

作　　者：王容 张永辉[1] 张健 张帅岩 WANG Rong;ZHANG Yong-hui;ZHANG Jian;ZHANG Shuai-yan(College of Information Science and Technology,Hainan University,Haikou 570228,China)

机构地区：[1]海南大学信息科学技术学院

出　　处：《计算机工程与设计》2019年第9期2614-2619,共6页Computer Engineering and Design

基　　金：海南省自然科学基金项目(618MS027);海南大学大学生创新创业基金项目(Hdcxcyxm201704)

摘　　要：针对监控环境中的人脸图像分辨率低、辨识度差的问题,提出一种基于卷积神经网络的人脸超分辨率重建算法。设计包含42个卷积层的网络,在高层采用Inception结构与残差网络的组合学习残差,与输入相加得到输出。Inception结构增加网络宽度,增加网络非线性,残差网络加快网络的收敛速度,将不同放大因子的低分辨率人脸融合到一个训练集中训练,解决不同放大因子的重建问题。实验结果表明,相比Bicubic和SRCNN,该方法在峰值信噪比(PSNR)和结构相似性指数(SSIM)上分别提升了2.4 dB/2.1%和1.33 dB/1.05%,网络的收敛速度得到较大提高。Aiming at the problem of low resolution and poor recognition of face images in surveillance environments,a face super-resolution reconstruction algorithm based on convolutional neural network was proposed.A deep convolutional network with 42 convolutional layers was designed,at the upper level,the residuals were learned using the combination of the Inception structure and the residual network,and added to the input to obtain the output.Using Inception structure increased the network depth and width,which increased network nonlinearity,and using the residual network speeded up network convergence.The low-resolution human faces with different amplification factors were combined into one training set for training,which solved the problem of reconstruction of different magnification factors.Experimental results show that compared with the Bicubic and SRCNN,this method improves PSNR and SSIM by 2.4 dB/2.1%and 1.33 dB/1.05%respectively,and the convergence speed of the network is greatly improved.

关 键 词：人脸超分辨率 卷积神经网络 视频监控 残差网络 Inception结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]