摘要
为解决传统立体匹配算法匹配低纹理人脸图像时极易产生误匹配的问题,提出一种基于区域生长的人脸立体匹配算法。该算法利用级联回归树算法提取的人脸特征点将人脸划分为不同区域以分别限制各区域的视差搜索范围,从而避免在全局范围上查找匹配点;同时利用人脸的局部形状特性,采用局部曲面拟合的方式筛除误匹配种子点并生成大量可靠种子点用于区域生长;最后,分别在实验室环境采集的人脸图像和FRGC v2.0人脸数据库上进行定性和定量实验。实验结果表明,与传统算法相比,所提算法能够重建出更加准确的三维人脸模型。经点云配准后与人脸点云真实值的均方根误差在2 mm以内,且不同光照、姿态、表情下人脸图像的重建表明所改进的立体匹配算法具有较好的鲁棒性。
To address the problem of mismatches produced by traditional stereo matching algorithms when matching low-texture face images,this paper proposed a face stereo matching algorithm based on region growing.This algorithm used the facial feature points extracted by the ensemble of regression trees to divide the face into different regions and limited the disparity search range of different regions,thereby avoided finding matching points on the global scale.Meanwhile,it used the local shape characteristics of the face and the local surface fitting to filter out mismatched seed points and generate a large number of reliable seed points for region growing.Finally,this paper performed qualitative and quantitative experiments on face images acquired in laboratory environment and the FRGC v2.0 face database respectively.Experimental results show that compared with the traditional algorithm,the improved algorithm can reconstruct a more accurate 3D face model,the root mean square error is within 2 mm after the point cloud registration.In addition,the reconstructions of face images under different lighting,poses and expressions show that the improved stereo matching algorithm has better robustness.
作者
夏颖
盖绍彦
达飞鹏
Xia Ying;Gai Shaoyan;Da Feipeng(School of Automation,Southeast University,Nanjing 210096,China;Key Laboratory of Measurement&Control of Complex Systems of Engineering of Ministry of Education,Southeast University,Nanjing 210096,China)
出处
《计算机应用研究》
CSCD
北大核心
2021年第3期932-936,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(51475092)
深圳市知识创新计划基础研究项目(JCYJ20180306174455080)。
关键词
三维人脸重建
视差图
双目视觉
立体匹配
区域生长
人脸特征点
3D face reconstruction
disparity map
binocular vision
stereo matching
region growing
face feature point
