融合高频滤波和伪影损失的人脸超分辨率重建  

Face image super-resolution fusing with high-frequency filtering and artifact loss

作　　者：孙红[1] 宋冬豪 陈玉娟 Sun Hong;Song Donghao;Chen Yujuan(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&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：《计算机应用研究》2023年第6期1906-1911,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61472256,61170277,61703277)。

摘　　要：人脸超分辨率重建的需求愈发强烈,针对现有方法在恢复图像时高频信息丢失严重导致平滑,同时伴随着伪影的问题,提出了融合高频滤波和伪影损失的重建方法。该方法能够获取人脸高频信息,在不影响细节纹理的情况下去除伪影,以生成对抗网络模型为框架,引入自适应残差结构以减少计算成本,使用Ranger优化器来缓解训练的不稳定。实验中,使用不同缩放因子,该方法相较于其他方法拥有更高的PSNR值和SSIM值。2倍、4倍、8倍缩放时在CelebAMask-HQ数据集上的PSNR值分别为37.88 dB、32.50 dB、29.51 dB,同时模型收敛速度较快,表明该方法的高效性与稳定性。The demand for super-resolution reconstruction of faces is becoming more and more intense.In light of the problem that the existing method causes smooth due to severe loss of high-frequency information and has artifacts when restoring images,this paper proposed a reconstruction method that combined high-frequency filtering and artifact loss.This method could obtain high-frequency information of the face and remove artifacts without affecting the detailed texture.Using the framework of generative adversarial network,it introduced an adaptive residual structure to reduce computational costs and the Ranger optimizer to alleviate the instability of training.In the experiment,using different scaling factors,the method has higher PSNR and SSIM values than other methods.The PSNR values on the CelebAMask-HQ dataset at×2,×4,and×8 scaling are 37.88 dB,32.50 dB and 29.51 dB,respectively,the model converges faster,indicating the efficiency and stability of the method.

关 键 词：高频滤波 伪影损失 自适应残差 生成对抗网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]