基于局部相似性的视频镜头分割

Video shot segmentation based on local similarity

为了进一步提高视频镜头的分割精度,提出了一种基于局部相似性的视频镜头分割方法。首先为了有效地进行视频特征降维,提出了改进的保局投影算法,利用仿射传播聚类算法得到具有相同模式的相似样本,根据相似样本构建连接矩阵,并根据降维前后模式的相关系数确定最佳降维维数,该算法有效地保留了数据的局部分布信息;然后利用具有相同模式的相似样本构建局部支持向量机检测镜头边界。实验结果表明,该方法利用样本的局部相似性特点,在视频特征提取和镜头边界检测两个阶段提高了镜头分割精度。

In order to improve the precision of video shot segmentation, this paper presented a new method for video shot seg-mentation based on local similarity. The improved locality perserving projections （ LPP）, which preserved the local distributioninformation of the data manifold, was proposed to extract effective viedo feature. It obtained similariy samples with the samemodel using affinity propagation clustering, and then built the connection matrix of LPP based on similariy samples, determinedthe optimal projection dimension by the correlation coefficient of model between dimension reduction before and after. The algo-rithy proposed local support vector machine （LSVM） baesd on similariy samples to detect video shot boundary. The resultsshow that the proposed method based on local similarity improves the precision of video shot segmentation in the stage of viedofeature extracting and shot boundary detecting.