摘要
为提高视频人脸聚类性能,解决视频中的人脸易受到光照强度、物体遮挡和背景变化等因素的干扰的问题,提出一种基于视频先验知识、多视图和轨迹信息约束的人脸聚类研究方法。对聚类样本进行多特征提取,利用稀疏子空间表示算法获取人脸稀疏系数表示矩阵,使用轨迹信息和KNN重构系数矩阵,结合协同谱聚类算法获得人脸聚类结果。通过Notting Hill库和电影轨迹人脸库两个数据集验证该方法的可行性,实验对比结果表明,该方法对于视频中的人脸聚类具有较好的性能。
Aiming at improving the performance of video face clustering,and solving the problem that faces in videos can be disturbed by illumination intensity,object occlusion and background changes easily,a face clustering method was proposed based on video prior knowledge,multi-view and constrained information.Multiple features of images were extracted,and sparse subspace representation algorithm was used to achieve the sparse coefficient matrix.The constrained track matrix and KNN were used to reconstruct the coefficient matrix.The clustering result was obtained by multi-view co-trained spectral clustering.The effectiveness of the method was demonstrated by two datasets,Notting Hill dataset and movie trailer face dataset.The experiment comparison results show that the proposed method has good performance on video face clustering.
作者
卞佳丽
梅雪
张晋
BIAN Jia-li;MEI Xue;ZHANG Jin(College of Electrical Engineering and Control Science,Nanjing Tech University,Nanjing 210000,China)
出处
《计算机工程与设计》
北大核心
2019年第11期3219-3224,共6页
Computer Engineering and Design
基金
江苏省研究生科研创新基金项目(KYCX17_0925)
关键词
人脸聚类
多视角
稀疏子空间表示
约束矩阵
协同算法
face clustering
multi-view
sparse subspace representation
constrained matrix
co-training