基于多尺度细节的高保真三维人脸重建  

High-Fidelity 3D Facial Reconstruction Based on Multi-Scale Details

作　　者：彭辰 Chen Peng(College of Management,University of Shanghai for Science and Technology,Shanghai)

机构地区：[1]上海理工大学管理学院,上海

出　　处：《建模与仿真》2025年第2期796-807,共12页Modeling and Simulation

摘　　要：三维重建算法在人脸识别、影视娱乐、医疗美容等领域有广泛应用,但在细节特征恢复方面仍存在局限。为有效提高人脸几何细节重建恢复模型的建模性能,提出了一种基于多尺度细节模型的高保真三维人脸细节重建修复算法。首先,设计了A2D-ResNet50网络,并在其中集成了残差可变核卷积模块,通过动态调整卷积核采样位置以增强细节特征提取。其次,引入了相对总变分(RTV)细节损失函数以增强局部几何细节恢复。此外,对特征点损失函数进行优化,并结合拉普拉斯变形约束,确保局部细节精确对齐和全局形状保持。实验结果显示,该方法在重建精度方面较现有技术有所提高,具有更高的鲁棒性和个性化细节恢复能力。3D reconstruction algorithms are widely used in fields such as face recognition,film and television entertainment,and medical aesthetics,but they have limitations in the recovery of detailed features.To balance global shape fidelity and local geometric detail recovery,a high-fidelity 3D facial reconstruction algorithm based on multi-scale details is proposed.The algorithm first designs an A2DResNet50 network and integrates a Residual Alterable Kernel Convolution Block within it,which enhances the extraction of detailed features by dynamically adjusting the sampling position of the convolutional kernel.Secondly,a Relative Total Variation(RTV)detail loss function is introduced to enhance the recovery of local geometric details.In addition,the feature point loss function is optimized and combined with Laplacian deformation constraints to ensure precise alignment of local details and preservation of global shape.Experimental results show that the method improves in detail accuracy compared to existing technologies and has higher robustness and the ability to recover personalized details.

关 键 词：三维人脸重建 可变核卷积 相对总变分 拉普拉斯变形 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]