基于DECA的单目人脸三维重建研究  

Research on Three-Dimensional Reconstruction of Monocular Face Based on DECA

作　　者：王承伟 张翠军 王振凯 李昊渊 张一帆 

机构地区：[1]河北地质大学信息工程学院,河北 石家庄 [2]河北地质大学人工智能与机器学习研究室,河北 石家庄

出　　处：《计算机科学与应用》2023年第12期2500-2508,共9页Computer Science and Application

摘　　要：针对基于单目图像的DECA模型在人脸三维重建时精度不高,且容易出现过拟合的问题,提出用Vision Transformer (ViT)改进DECA模型的特征提取器部分,增强模型的局部和全局理解能力,提取更高维的特征,以提高人脸特征点的检测精度和人脸重建的精确性。进一步,引入DropKey策略,将ViT中的Key作为Drop对象,惩罚注意力峰值,以改善训练过程中的过拟合问题。实验结果表明,在引入ViT和DropKey策略后,人脸三维重建的效果有明显的提升。In view of the low accuracy of the monocular image-based DECA model in face 3D reconstruction and the problem of overfitting, Vision Transformer (ViT) is proposed to replace the feature extractor part of the DECA model to enhance the local and global understanding ability of the model and extract higher-dimensional features. To improve the accuracy of face feature point detection and face reconstruction. Further, the DropKey strategy is introduced, and the Key in ViT is used as Drop object to punish the attention peak, so as to improve the overfitting problem in the training process. The experimental results show that after the introduction of ViT and DropKey strategies, the effect of face 3D reconstruction has been significantly improved.

关 键 词：人脸重建 深度学习 Vision Transformer DropKey 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]