渐进式生成对抗网络的人脸超分辨率重建  

Progressive GAN for Face Image Super-resolution

作　　者：胡德敏[1] 王揆豪 林静 HU De-min;WANG Kui-hao;LIN Jing(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2021年第9期1955-1961,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61170277,61472256)资助;上海市教委科研创新重点项目(12zz137)资助;上海市一流学科建设项目(S1201YLXK)资助。

摘　　要：人脸幻构是图像超分辨率重建领域的一个子领域,用于恢复面部基本特征且不变形.现有方法着重于恢复本身细节相对丰富的图像,本文针对高频细节已丢失严重的人脸图像提出了一种渐进式生成对抗网络的人脸超分辨率重建方法(P-FSRGAN),可生成逼真的8倍超高分辨率人脸图像.采用渐进式生成方法,通过分阶段拆分训练的方式来保证训练过程的稳定.Inception-ResNet结构的引入增加了网络的宽度;加快了网络收敛速度.引入语义分割网络获得人脸的边缘轮廓信息和面部特征.实验结果表明,在8倍放大尺度因子下,P-FSRGAN的峰值信噪比达到25.83dB、结构相似性指标达到0.7735、多尺度结构相似性指标达到0.8989,均优于其他算法,表明了本文方法的有效性.Face hallucination(FH)is a subfield in the field of image super-resolution reconstruction(SR),which is used to recover basic facial features without distortion.Existing methods focus on recovering images that are relatively rich in their own details,and this paper proposes a progressive CAN for face image super-resolution(P-FSRGAN)method to generate realistic 8 x super-high resolution face images for face images where high-frequency details have been severely lost.The progressive generation method is used to ensure the stability of the training process by splitting the training in stages.The Inception-ResNet structure is introduced to increase the width of the network;to speed up the convergence of the network.The semantic segmentation network is introduced to obtain the edge contour information and facial features of a face.The experimental results show that the peak sigmal-to-noise ratio of P-FSRGAN reaches 25.83 dB,the structural similarity index reaches 0.7735,and the multiple-scale structural similarity index reaches 0.8989 under 8-fold amplification scale factor,which are all better than other algorithms indicating the effectiveness of the method in this paper.

关 键 词：人脸超分辨率 语义分割 Inception-ResNet结构 生成对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]