表情数据库无关的人脸表情转移  

Facial Expression Transfer without Expression Database

作　　者：吕培[1] 徐明亮[1] 谢蕾[2] 周兵[1] 

机构地区：[1]郑州大学信息工程学院,郑州450001 [2]浙江大学CAD&CG国家重点实验室,杭州310058

出　　处：《计算机辅助设计与图形学学报》2016年第1期68-74,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61379079;61202207;61472370;61502433);教育部博士点基金(20124101120005)

摘　　要：为了提高人像的脸部表情质量,提出一种面向同一人脸表情转移的方法.该方法的输入包括2幅人脸图像、中性人脸库以及人脸blendshape模型.首先根据人脸轮廓面积,从2幅输入人脸图像中选择一幅近正面人脸图像,并利用中性人脸库及人脸blendshape模板生成输入人脸的特定blendshape模型;然后利用blendshape模型生成与输入人脸图像匹配的三维人脸模型,并利用RBF网格变形算法对其进行调整;最后利用生成的三维人脸模型对2幅输入图像进行扭曲,并通过图像混合的方式生成表情转移图像.实验结果表明,文中方法能够生成自然的表情转移图像,有效地提高原有图像中的人脸表情质量.This paper presents a method of facial expression transfer using expressions from same person. Its target is to improve the quality of facial expressions in photos. The inputs of this method include two face images, neural face database and template blendshape face model. Firstly, according to the contour of face area, a near frontal face image is chosen from these two input face images. Secondly, based on the neutral face database and template blendshape face model, a specific blendshape model for this person is generated. Thirdly, using this blendshape model, the 3D face model of the input face is generated and refined using RBF mesh deformation. Finally, with these two generated 3D face models, the two input images are distorted and facial expression transfer image is generated. The experimental results show that the proposed method can generate natural expressions transferred images and improve the quality of facial expressions in original photos.

关 键 词：人脸表情 表情转移 网格变形 图像混合 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]