摘要
针对传统的三维(3D)人脸识别算法仅考虑特征提取而不能很好地运用于实际视频人脸识别系统的问题,提出一种基于格拉斯曼流形谱聚类的动态3D视频全自动识别系统。首先通过去除孤立点、均匀采样、剪裁、姿势纠正等过程将3D视频数据集进行规范化;然后从训练视频的不同位置提取出可变长度的局部视频片段,使用基于谱聚类的高效算法将其表示为格拉斯曼流形上的点;最后,将所得到的聚类中心和测试视频中的点相匹配,并且利用基于表决的策略来完成测试视频的人脸识别。在大型通用3D视频数据库BU4DFE上的实验验证了该算法的有效性。实验结果表明,与几种较为先进的视频人脸识别算法相比,该算法取得了更好的识别效果。
Traditional three-dimensional face recognition methods only consider feature extraction but can not be well applied in actual video face recognition systems. Aiming at this problem,we propose the dynamic 3D video automatic recognition system which is based on spectral clustering on Grassman manifold. Firstly,it standardises 3D video datasets through outliers removing,uniform sampling,clipping and posture correcting. Then,it extracts local video clips which lengths are variable from different positions of the training video,and uses spectral clustering-based efficient algorithm to represent them as the points on Glassman manifold. Finally,it matches the derived clustering centre with the points in testing video,and completes face recognition of test video using vote-based strategy. The effectiveness of the proposed method is verified by experiments on large-scale popular 3D video database BU4DFE. Experimental results show that the proposed algorithm has better recognition effect than several rather advanced video face recognition algorithms.
出处
《计算机应用与软件》
CSCD
北大核心
2014年第5期168-171,共4页
Computer Applications and Software
关键词
人脸识别
格拉斯曼流形
谱聚类
三维视频
面部表情
Face recognition Grassman manifold Spectral clustering Three-dimensional video Facial expression
