一种快速有效的网络视频拷贝检测方法

Effective and efficient copy detection algorithm for Web video

为了快速有效地检测真实网络中的拷贝视频,针对现有基于顺序度量特征的检测算法存在鲁棒性不足和相似性度量不准确的问题,提出基于核心区域顺序度量特征和转换距离的视频拷贝检测方法。该方法在统计分析真实网络中拷贝视频特点的基础上,首先选取拷贝视频中相对稳定的核心区域提取顺序度量特征;其次提出基于最小转换代价的度量标准,并设计相应的顺序度量特征快速匹配方法;最后采用简化的最长匹配子序列算法进行特征序列匹配,检测查询视频中的拷贝片段。基于真实网络数据和MUSCLE-VCD-2007数据的实验结果显示,相对于现有基于顺序度量特征的拷贝检测方法,本方法鲁棒性更强,检测效率更高。

To efficiently detect near-duplicate Web videos and deal with the problems of poor robustness and bad similarity measurement in the exiting algorithms based on ordinal measures, this paper proposed a method based on core area ordinal measures and the transformation distance. After analyzing Web near-duplicates, it selected the comparatively invariant area of copy video to extract ordinal measure fingerprints. Then it used the minimum cost of the transformation to measure the distance between the ordinal measures, and designed the effective algorithm for fingerprint matching. Finally, the proposed method used the simplified longest same sub-sequences algorithm to match the sequences of features and detect the copy clips. Experimental results based on real network and MUSCLE-VCD-2007 datasets show good robustness of core area ordinal measures and excellent performance of copy detection.