一种新的时序一致性特征的近重复视频检测算法  

A New Method for Detecting Near-duplicate Videos with Temporal Consistency Feature

作　　者：郭丁云 薛峰 郭旦萍 GUO Ding-yun, XUE Feng, GUO Dan-ping （Anhui Sanlian University, Hefei 230601, China）

机构地区：安徽三联学院电子电气工程学院,安徽合肥230601

出　　处：《电脑知识与技术》2016年第11期160-163,共4页Computer Knowledge and Technology

基　　金：安徽三联学院院级科研项目（20142008）;安徽省高等教育振兴计划新专业建设项目（2013zvtz082）.

摘　　要：针对基于镜头关键帧层次进行近重复视频检测耗时的问题,该文提出了一种新的视频镜头的时序一致性特征,使用该特征对已有研究的级联式近重复视频检测算法进行改进。首先在进行关键帧特征提取之前,在镜头层次利用镜头的时序一致性特征初步滤除掉一些完全不相同的视频,然后再使用全局分块颜色特征和SURF特征,并使用k-d树型索引结构来实现快速最近邻检测,最终逐步检测与查询视频重复或近重复的视频。通过在CC_WEB_VIDEO标准数据集上进行实验,并与现有常用的三种方法比较,结果显示本文方法的检测效果更好,尤其对于那些画面变化较大以及运动场景复杂的视频更明显。Given the existing near-duplicate video detection algorithms are processed by matching features directly in the key frame level, which is very time-consuming, this paper introduces a new cascaded near-duplicate video detection approach using the temporal consistency feature in the shot level to preliminarily filter out some dissimilar videos before extracting features, and combining global blocking color features and SURF feature. Then we use the k-d tree indexing structure to achieve fast detec- tion speed and thus obtain the ultimate videos that are duplicated with the query video step by step. We have verified the ap- proach by experimenting on the CC_WEB_VIDEO dataset, and compared the performance with the existing three commonly used methods. The analysis results show the proposed method can achieve better detection effect, especially for the videos with great frame changes and complex motion scenes.

关 键 词：近重复视频检测 时序一致性特征 分块颜色特征 加速鲁棒特征 k-d树型索引结构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术;自动化与计算机技术—计算机科学与技术]