计算机工程与应用2026，Vol.62Issue(17)：270-282,13.DOI:10.3778/j.issn.1002-8331.2507-0102
改进ResNet-3D的多视角三维人脸重建
Multi-View 3D Face Reconstruction Based on Improved ResNet-3D
摘要
Abstract
In multi-view 3D face reconstruction,the traditional CNN has insufficient reconstruction accuracy due to the limited ability of feature extraction.Aiming at the limitations of traditional multi-view 3D face reconstruction methods in feature extraction and geometric modeling,an end-to-end multi-view 3D face reconstruction system based on improved ResNet-3D network is proposed.The system extracts multi angle facial spatial features based on the standard ResNet-3D architecture,and introduces the spatial attention gating(SAG)module and dual stream architecture to improve the pro-cess.Then,in the feature fusion stage,a virtual view generation strategy based on spherical interpolation is proposed to improve the view generalization ability,to build a hybrid loss function with view rendering consistency and geometric smoothness constraints to optimize the accuracy of geometric modeling,and to cooperate with 3DMM parameter regres-sion to achieve high-precision 3D reconstruction.Experimental results show that on the BU-3DFE dataset,this method outperforms traditional methods in terms of the following indicators:(1)NME reduction by 25.3%;(2)reconstruction accuracy of 95%for 68 keypoints;(3)cross-view generalization error reduction by 58.2%.This method effectively im-proves the accuracy of multi-view reconstruction.关键词
三维人脸重建/三维残差网络(ResNet-3D)/多视角融合/三维可变形模型(3DMM)/注意力机制Key words
3D facial reconstruction/3D residual network(ResNet-3D)/multi-view fusion/3D morphable model(3DMM)/at-tention mechanism分类
信息技术与安全科学引用本文复制引用
卢政霖,许锋..改进ResNet-3D的多视角三维人脸重建[J].计算机工程与应用,2026,62(17):270-282,13.基金项目
"十三五"国家重点研发技术项目(2017YFC0821004) （2017YFC0821004）
2022辽宁省教育厅基本科研重大攻关项目(LJKZZ20220007) （LJKZZ20220007）
中央高校基本科研业务重大培育项目(3242022004). （3242022004）
