基于3D先验特征的人脸超分辨率重建算法  

Facial Super-resolution Reconstruction AlgorithmBased on 3D Prior Features

作　　者：姚汉群 刘广文[1] 王超[2] 杨依宁 才华[1] 付强[4] YAO Hanqun;LIU Guangwen;WANG Chao;YANG Yining;CAI Hua;FU Qiang(School of Electronic Information Engineer,Changchun University of Science and Technology,Changchun 130022,China;National and Local Joint Engineering Research Center for Space Optoelectronics Technology,Changchun University of Science and Technology,Changchun 130022,China;National Key Laboratory of Electromagnetic Space Security,Tianjin 300308,China;School of Opto-Electronic Engineer,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]长春理工大学空间光电技术国家与地方联合工程研究中心,长春130022 [3]电磁空间安全全国重点实验室,天津300308 [4]长春理工大学空间光电技术研究所,长春130022

出　　处：《吉林大学学报（理学版）》2024年第4期895-904,共10页Journal of Jilin University:Science Edition

基　　金：国家自然科学基金重大项目(批准号:61890963);吉林省科技发展计划项目(批准号:20210204099YY)。

摘　　要：为有效解决复杂环境下人脸超分辨率特征恢复的问题,提出一种全新的人脸超分辨率网络.该网络通过融合3D渲染先验知识和双重注意力机制,增强了对人脸空间位置和整体结构的理解,同时提高了细节信息的恢复能力.在数据集CelebAMask-HQ上的实验结果表明:对放大4倍下采样的人脸,该算法在峰值信噪比和结构相似性上达到28.76 dB和0.827 5;对放大8倍下采样的人脸,峰值信噪比和结构相似性评价指标达到26.29 dB和0.754 9.与同类的SAM3D算法相比,该算法在处理放大4倍下采样时的峰值信噪比和结构相似性上分别提升了4.09,1.93个百分点,在处理放大8倍下采样时上述两个指标分别提升了2.02,4.54个百分点.从而证明该算法的优越性,也表明在实际应用中人脸的超分辨率恢复能获得更真实和清晰的视觉效果.In order to effectively solve the problem of facial super-resolution feature recovery in complex environments,we proposed a novel facial super-resolution network.By integrating 3D rendering prior knowledge and a dual attention mechanism,the network enhanced the understanding of the facial spatial position and overall structure while improving the ability to recover detailed information.The experimental results on the CelebAMask-HQ dataset show that the proposed algorithm achieves peak signal-to-noise ratio and structural similarity of 28.76 dB and 0.8275 for downsampled faces magnified by 4 times,and 26.29 dB and 0.7549 for downsampled faces magnified by 8 times.Compared with the similar SAM3D algorithm,the proposed algorithm improves the peak signal-to-noise ratio and structural similarity by 4.09 and 1.93 percentage points when dealing with 4 times downsampling,and by 2.02 and 4.54 percentage points when dealing with 8 times downsampling,respectively.This proves the superiority of the proposed algorithm and also indicates that facial super-resolution recovery can achieve more realistic and clear visual effects in practical applications.

关 键 词：机器视觉 人脸超分辨率 3D先验 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]