基于双目视觉的人脸三维重建  

Face reconstruction based on binocular stereo vision

作　　者：林琴[1,2,3] 李卫军 董肖莉[1,2,3] 宁欣 陈鹏 LIN Qin;LI Weijun;DONG Xiaoli;NING Xin;CHEN Peng(Laboratory of Artificial Neural Networks and High-peed Circuits,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China;School of Microelectronics,University of Chinese Academy of Sciences,Beijing 100029,China;Cognit-ive Computing Technology Wei Fu Joint Lab,Beijing 100083,China)

机构地区：[1]中国科学院半导体研究所高速电路与神经网络实验室,北京100083 [2]中国科学院大学电子学院,北京100029 [3]认知计算技术威富联合实验室,北京100083

出　　处：《智能系统学报》2018年第4期534-542,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(90920013);国家公派留学基金项目(201404910237)

摘　　要：基于双目立体匹配算法Patch Match算法,提出了一种获取人脸三维点云的算法。该算法对局部立体匹配算法Patch Match进行了优化。该方法既不需要昂贵的设备,也不需要通用的人脸三维模型,而是结合了人脸的拓扑结构信息以及立体视觉局部优化算法。此方法采用非接触式的双目视觉采集技术获取左右视角的人脸图像,利用回归树集合(ensemble of regression trees,ERT)算法对人脸图像进行关键点定位,恢复人脸稀疏的视差估计,运用线性插值方法初步估计脸部的稠密视差值,并结合局部立体匹配算法对得到的视差结果进行平滑处理,重建人脸的三维点云信息。实验结果表明,这种算法能够还原出光滑的稠密人脸三维点云信息,在人脸Bosphorus数据库上取得了更加准确的人脸重建结果。In this paper, we propose a binocular stereo algorithm called PatchMatch for generating a 3D dense pointcloud of the human face. The proposed algorithm optimizes a local stereo matching method, also known as PatchMatch,which combines topological information of the human face with a local optimization algorithm for stereo vision and re-quires neither expensive equipment nor generic face models. With this method, by applying a non-contact binocular vis-ion selection technology, face images at both left and right visual angles are obtained. We use an ensemble of regressiontrees （ERT） algorithm to position key points of a face image and estimate the sparse disparity of facial landmarks. Then,we use a linear interpolation method to make a preliminarily estimation of the dense facial disparity, and by using thelocal stereo matching algorithm, we can smooth the obtained visual disparity results and use the three-dimensional pointcloud information to rebuild the human face. The experimental results with the Bosphorus database show that the pro-posed algorithm can recover dense facial three-dimensional point cloud information and obtain more accurate face re-construction results than other methods on Bosphorus database.

关 键 词：人脸拓扑结构 立体匹配 线性插值 稠密视差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]