一种视频数据代表选择框架方法

Representative Selection Framework Approach for Videos

为有效处理视频数据问题,提出一种识别海量数据集中代表子集的方法,即代表选择方法,经选择后的小容量的数据代表完全可以代表原始大数据集的结构特征。对于给定的大数据集,首先生成相应1-norm非负稀疏图,然后利用一种谱聚类算法基于所生成的稀疏图将大数据反复划分直至形成聚类簇。代表选择过程中,将每个聚类看作Grassmann流形中的一个点,然后基于测地距衡量这些点间的距离,接着利用min-max算法分析距离以提取出较优的聚类子集。最后,通过分析被选集类的一个稀疏子图,利用主成分集中性方法探测出数据代表,称此过程为基于非负稀疏图与Grassmann流形测地距的代表选择框架。为验证所提出的框架,将之应用于视频分析中,从一长段的视频流中识别出少数的几个关键帧,实验效果通过人工判断与标准评价方法进行评价,并与现有的几种方法的效果进行比对,结果证明所提出的代表选择框架方法具有更好的效果与可行性。

In order to solve the process problem of massive videos data,a representative selection method of identifying the optimal subset of data points as a representative of original massive dataset was proposed.The selected data points of subset can represent inner structure of original massive dataset.And the novel representative selection method is based on l_1-norm non-negative sparse graph for the original massive dataset.The massive data points are partitioned into some clusters by using a spectral clustering algorithm based on the non-negative sparse graph generated in previous steps.Each cluster is viewed as a point in the Grassmann manifold,and the geodesic distances among these points are measured.By using a min-max algorithm,geodesic distances are analyzed to build an optimal subset of clusters.Finally,the principal component centrality method is used to detect a representative after analyzing the sparse graph of selected clusters.The proposed framework is validated on the problem of video summarization,where a few key frames should be selected in long video clips which contain massive frames.The comparison of the results obtained between the proposed algorithm and some state-of-the-art methods was producted.Result indicates the effectiveness and feasibility of the proposed framework.