三维双线性多分辨率形变模型研究  

Research on 3D Bilinear Multidimensional Morphable Models

作　　者：刘碧霞[1] 李绍滋[1,2] 郭锋[1,2] 苏松志[1,2] 

机构地区：[1]厦门大学信息科学与技术学院 [2]福建省仿脑智能系统重点实验室(厦门大学),福建厦门361005

出　　处：《厦门大学学报（自然科学版）》2011年第3期509-514,共6页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金项目(60873179);深圳市科技计划基础研究项目(JC200903180630A);高等学校博士学科点专项科研基金(20090121110032)

摘　　要：形变模型在计算机视觉和计算机图形学等领域均有广泛的应用.但现有形变模型的建立或依赖于不稳定的人脸图像对应光流算法,或需要大量的人机交互,而且基于随机梯度下降算法的模型匹配过程常收敛于局部最优值.针对这些缺陷,提出三维双线性多分辨率形变模型.首先基于人脸关键特征分割和网格重采样建立原始人脸的自动稠密对应;然后经紧致凸松弛将模型匹配问题转化为两个双线性规划问题,最后通过基于内点方法的全局优化算法求解.基于模拟和真实数据的实验表明:该模型在最优性、模型匹配速度、收敛性和对噪声异常点的鲁棒性优于传统的形变模型.Morphable model has widespread applications in computer vision and computer graphics.The current model construction method either depends on the unstable optical flow algorithm or it requires lots of human-machine interactions.The model matching procedure based on stochastic gradient descent algorithm often gets stuck in the local optimum.In this paper,3D bilinear multidimensional morphable models are proposed.Firstly,dense correspondences between prototypic faces are constructed by face segmentation based on key feature and the mesh resampling.And then the model matching problem is converted into two bilinear programs after tight convex relaxations.Finally,the bilinear programs are efficiently solvable by modern interior point methods.Experiments with synthetic data and real image data validate that the 3D bilinear multidimensional morphable models outperform the morphable models in optimality,model matching speed,convergence rate and robustness to noise and outliers.

关 键 词：形变模型 稠密对应 三维人脸重建 双线性规划 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]