基于卷积神经网络的人脸图像超分辨率重建方法  

Face Image Super-resolution Reconstruction Method Based on Convolutional Neural Network

作　　者：倪若婷 周莲英[1] NI Ruoting;ZHOU Lianying(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212013)

机构地区：[1]江苏大学计算机科学与通信工程学院,镇江212013

出　　处：《计算机与数字工程》2022年第1期195-200,共6页Computer & Digital Engineering

基　　金：市校级科研项目“慢病本体知识库开发”(编号:8421170004);市校级科研项目“市智慧妇幼信息平台本体知识库系统”(编号:20180477)资助。

摘　　要：为了恢复极低分辨率人脸图像的轮廓及更多细节,论文提出一种结合基于特征的反投影方法以及人脸语义分割的方法。该方法有四部分,包括初步重建网络、精细编码网络、人脸信息先验估计网络以及精细解码网络。首先通过初步重建网络,将图像进行4倍上采样。其次使用DBPN的反投影网络进行精细重建,对初步重建网络的结果进行两倍上采样。与此同时,引入U-net网络对初步重建网络输出结果进行语义分割,得到人脸分割信息。最后将精细网络的输出结果与人脸分割信息进行维度拼接并卷积,完成最后的重建。论文还在MSE损失函数的基础上引入人脸先验信息损失函数与VGG损失函数,有效引导人脸轮廓、五官重建。实验结果验证了该方法的有效性,该方法比基于双三次插值方法的PSNR值和SSIM值分别提升了3.6db和13.34%,最终重建出的人脸图像取得了较好的效果。In order to recover the contour and more details of extremely low-resolution face images,this paper proposes a method that combines feature-based back-projection method and face semantic segmentation.The method consists of 4 parts,in⁃cluding coarse net,fine SR encoder,facial prior estimation network,fine decoder SR network.First,the image is up-sampled 4 times through the coarse net.Secondly,deep back-projection network in DBPN is used for fine reconstruction,and the results of coarse net are up-sampled twice.At the same time,a U-net network is introduced for semantic segmentation on the result of coarse net to obtain face segmentation information.Finally,the output of fine SR encoder and facial prior estimation network are concatenat⁃ed and convolved to complete the final reconstruction.This paper also adds face prior information loss function and VGG loss func⁃tion,which effectively guides the recovery of face contours and features.The results verify the effectiveness of this method.Com⁃pared with reconstruction method based on bicubic,PSNR and SSIM value of this method are increased by 3.6db和13.34%respec⁃tively,and the reconstructed face image has achieved better results.

关 键 词：人脸超分 神经网络 语义分割 损失函数 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]