基于图像集匹配(ISM)的正则化最近点法在视频人脸识别中的应用

Application of Regularized Nearest Points Algorithm Based on ISM in Video Face Recognition

针对视频人脸识别中传统的图像集算法受光照、表情、姿态及遮掩等变化而影响识别性能的问题,提出了一种图像集匹配的正则化最近点算法。首先,将图库图像集和探针图像集建模成正则化仿射包,利用迭代器自动确定两个图像集间的正则化最近点;然后,利用最近子空间分类器最小化正则化最近点;最后,根据正则化最近点之间的欧氏距离及结构计算RNP集之间的距离,并利用最近邻分类器完成人脸识别。在Honda/UCSD、CMU Mobo和YouTube三大视频人脸数据库上的实验验证了所提算法的有效性及可靠性,实验结果表明,相比其他几种图像集人脸识别算法,所提算法取得了更好的识别效果,同时,大大减少了训练及测试总完成时间。

The recognition performance of traditional image set face recognition algorithms is impacted by variation of illustration,expression,pose and occlusion seriously in video,for which regularized nearest points algorithm based on image sets matching is proposed. Firstly,gallery image sets and probe image sets are modeled as regularized affine hulls,and iterator is used to conform regularized nearest points between the two sets. Then,recent subspace classifier is used to minimize the regularized nearest points. Finally,distance between RNP sets is calculated by Euclidean distance and structure of RNP and nearest neighbor classifier is used to finish face recognition. The effectiveness and reliability of proposed algorithm has been verified by experiments on the three video face databases Honda / UCSD、CMU Mobo and YouTube. Experimental results show that proposed algorithm has better recognition efficiency,less training time and testing time than several face recognition algorithms based on image sets.