基于深度学习的三维人脸重建抗遮挡网络  

3D Face Reconstruction Anti-occlusion Network Based on Deep Learning

作　　者：李杏清[1] 王志兵[3] 杨润丰[2] 曾德生[1] 聂影影 LI Xingqing;WANG Zhibing;YANG Runfeng;ZENG Desheng;NIE Yingying(School of Information Engineering,Guangdong Innovative Technical College,Dongguan 523960,China;School of Architecture,Dongguan Polytechnic,Dongguan 523808,China;School of Electronic Information,Dongguan Polytechnic,Dongguan 523808,China)

机构地区：[1]广东创新科技职业学院信息工程学院,广东东莞523960 [2]东莞职业技术学院建筑学院,广东东莞523808 [3]东莞职业技术学院电子信息学院,广东东莞523808

出　　处：《现代信息科技》2024年第11期22-25,共4页Modern Information Technology

基　　金：广东省教育厅2022年度普通高校科研平台特色创新类项目(2022KTSCX385);2021年广东省普通高校创新团队项目(2021KCXTD082);2022年东莞市社会发展科技面上项目(20221800903482)。

摘　　要：对人脸单遮挡模型和人脸多遮挡模型进行了研究,提出了一种基于深度学习的三维人脸重建抗遮挡网络,实现了对遮挡人脸的有效重建。改进的单遮挡模型通过预训练和权重的修改有效地实现了人脸图像上下文信息的捕获。改进的多遮挡模型通过特征扭曲和变换,使用分布的损失函数和不同的微分器得出重建的人脸图像。实验结果表明,提出的方法能够在多种遮挡情况下生成更为准确的三维人脸模型,具有更好的鲁棒性和抗遮挡能力。This paper researches both the face single-occlusion model and the face multiple-occlusion model.It proposes a 3D face reconstruction anti-occlusion network based on Deep Learning,realizing the effective reconstruction of occluded face.The improved single-occlusion model effectively realizes the capture of contextual facial information through pre-training and weight modifications.The improved multi-occlusion model employs a distributed loss function and distinct differentiators to achieve reconstructed facial images through feature distortion and transformation.Experimental results validate that the proposed method can generate more precise 3D facial models across various occlusion scenarios,demonstrating superior robustness and anti-occlusion capabilities.

关 键 词：深度学习 三维人脸重建 单遮挡模块 多遮挡模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]