摘要
针对现有的视频人脸识别方法不能很好地学习局部模型特定协方差的问题,为了更好地识别视频中的人脸,提出了基于异方差概率线性判别分析(PLDA)的外观流形建模(AMM)算法。首先,借助于高斯分布集合,对训练集中所有的人脸分别进行外观流形建模;然后对从视频中采集到的人脸进行聚类,并使用异方差PLDA模型学习聚类结果,从而获得表征分布的参数;最后,通过点到模型距离对测试人脸的每一帧到训练集的所有聚类进行融合匹配,并根据匹配得分最高原则完成人脸的分类。在两大通用视频人脸数据库Honda及MoBo上的实验验证了所提算法的有效性及稳定性,实验结果表明,相比其他几种较为先进的视频人脸识别算法,所提算法明显提高了识别率,并且大大降低了计算复杂度,有望应用于实时视频人脸识别系统。
For the issue existing video face recognition methods can not well learn specific covarianee of local m（alel, appearam＇e manifold modeling based on heteroscedastic Probabilistic Linear Discriminant Analysis （PLDA） is proposed to recognize video face better. Firstly, all faces in training sets are appearance manifold modeled respectively by Gaussian distribution collection. Then, faces got from video are clustered and beternscedastic PLDA is used to learn the clustering results so as to get characterization distribution parameters. Finally, distances between points and models are used to fusion and match all clusters from each frame of testing face and training sets, and criterion of getting highest matching scores is used to finish classification. The effectiveness and stability of proposed algorithm is verified by experiments on Honda and MoBo, experimental results show that proposed algorithm has improved recognition aeeuracy and decreased the computing complexity comparing with several advanced video face recognition algorithms, so it is expected to be applied into real-time video face recognition systems.
出处
《电视技术》
北大核心
2014年第9期218-222,227,共6页
Video Engineering
关键词
视频人脸识别
异方差概率
线性判别分析
外观流形建模
高斯分布
video face recognition
heleroscedastic probabilistic
linear discriminan｜ analysis
appearance manifold modeling
Gaussian distribution
