细节特征保持的三维面部表情迁移方法  

Detailed Features-Preserving 3D Facial Expression Transfer

作　　者：于志平 迟静[1,2] 叶亚男 代福芸 Yu Zhiping;Chi Jing;Ye Yanan;Dai Fuyun(School of Computer Science and Technology,Shandong University of Finance and Economics,Jinan 250014;Shandong Provincial Key Laboratory of Digital Media Technology,Jinan 250014)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东省数字媒体技术重点实验室,济南250014

出　　处：《计算机辅助设计与图形学学报》2021年第2期186-198,共13页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61772309);山东省省属优青项目(ZR2018JL022);山东省重点研发计划(2019GSF109112);山东省高等学校青创科技支持计划(2020KJN007);山东省自然科学基金(ZR2019MF016);山东省高等学校优势学科人才团队培育计划.

摘　　要：在三维面部表情迁移中,针对保持目标模型丰富的细节信息以使生成的新表情真实自然,以及减少表情迁移的学习训练时间这2个热点问题,提出一种细节特征保持的三维面部表情迁移方法.首先提取三维面部模型的细节特征,获得滤掉细节后的基本表情;然后利用改进的有参无监督回归方法将源模型的基本表情传递给目标模型;最后利用提出的细节特征向量调整策略对具有源基本表情的目标模型进行细节恢复.在Windows 10系统的Matlab中,以重建精度和训练时间为评价指标,对COMA等三维面部数据集进行视觉对比和定量分析实验.结果表明,与非线性联合学习方法相比,该方法在将源模型的表情无损迁移到目标模型的同时,很好地保持了目标模型自身的个性细节特征,使生成的表情真实自然;有效地提高了面部表情迁移的训练速度.In the 3D facial expression transfer field,aiming at the two hot problems of preserving the rich detailed information of the target model to make the generated new expressions realistic and natural,and reducing the training time,this paper presents a new detailed features-preserving 3D facial expression transfer method.Firstly,the detailed features are extracted from 3D face models to obtain the basic expression models without details.Then,the basic expression of source model is transferred to the target model with the improved parametric dimensionality reduction by unsupervised regression.Finally,the detailed features of the target model are restored by using the proposed detailed feature vector adjustment strategy.The visual contrast and quantitative analysis experiments with reconstruction accuracy and training time as the evaluation indexes are conducted on the 3D facial datasets such as COMA in the Matlab software under the Windows 10 environment.The results illustrate that compared with the nonlinear co-learning method,the method can not only transfer the expression of the source model to the target model without losses,but also well preserve the personalized detailed features of the target model,so it can make the generated expressions more realistic and natural.The method also effectively improves the training speed in facial expression transfer.

关 键 词：表情迁移 有参无监督回归 细节特征提取 细节特征恢复 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]