计算机与现代化

• 图像处理 • 上一篇    下一篇

动态人脸表情合成的模型特征驱动算法综述

  

  1. (深圳亚联发展科技股份有限公司研发中心,广东深圳518057)
  • 收稿日期:2019-02-20 出版日期:2019-07-05 发布日期:2019-07-08
  • 作者简介:陈松(1988-),男,四川阆中人,助理工程师,本科,研究方向:计算机图形学,机器学习,E-mail: 578079004@qq.com; 袁训明(1965-),男,山东寿光人,高级工程师,硕士,研究方向:计算机图形学,数字图像处理,E-mail: yxm123@asialink.com。

A Survey of Dynamic Human Facial Expression Synthesis 〖JZ〗Approaches Driven by Model Features

  1. (Research and Development Center, Shenzhen Asia Link Technology Development Co., Ltd., Shenzhen 518057, China)
  • Received:2019-02-20 Online:2019-07-05 Published:2019-07-08

摘要: 分析人脸模型的动态表情合成方法并依据它们内在特点进行分类描述。尽管这个领域已经存在较多文献,但是动态人脸表情合成仍然是非常活跃的研究热点。根据输出类型的不同,分类概览二维图像平面和三维人脸曲面上的合成算法。对于二维图像平面空间合成人脸表情主要有如下几种算法:主动表情轮廓模型驱动的人脸表情合成算法,基于拉普拉斯算子迁移计算的合成方法,使用表情比率图合成框架的表情合成算法,基于面部主特征点offset驱动的人脸表情合成算法,基于通用表情映射函数的表情合成方法和近来基于深度学习的表情合成技术。对于三维空间人脸合成则主要包括:基于物理肌肉模型的合成,基于形变的表情合成,基于三维形状线性回归的表情合成,基于脸部运动图的表情合成和近来基于深度学习的三维人脸表情合成技术。对以上每一种类别讨论它们的方法论以及其主要优缺点。本工作有望帮助未来研究者更好地定位研究方向和技术突破口。

关键词: 面部表情生成, 表情变形移植, 表情夸张, 脸部衰老效果

Abstract: This paper surveys the overall methodologies for generating dynamic facial expression on 2D and 3D human face models and categorizes them into several classes based on their intrinsic properties. Even though there exists a considerable body of previous works, this topic is still gaining very active research attention. According to the dimensionality of their final output, the methods are categorized into class working on 2D image plane and class working on 3D facial manifold surfaces. The general approach of synthesizing 2D human facial expression includes: Feature-points driven facial expression synthesis method, generation methods based on facial proportion map, Laplacian transfer based expression generation method, methods based on general expression mapping functions, synthesis methods based on active expression model and the latest approaches using deep learning techniques. On the aspect of synthesizing 3D facial expression, the classifications contain: 3D shape regression-based methods, synthesis methods based on deformation, approaches based on pseudo muscle models, algorithms based on a motion-vector analysis, and the latest approaches using deep learning techniques. For each category, we describe its basic methodology along with its advantages and limitations. This survey paper is expected to help researchers to better position their potential future direction in the context of existing solutions.

Key words: facial expression generation, expression deforming transfer, facial expression exaggeration, facial aging effect

中图分类号: