基于可变步长关键帧提取的网络视频拷贝检索

Web video copy detection based on variable-length step key frame selection

为快速有效地检测网络中的拷贝视频,针对现有关键帧提取算法时间复杂度高、关键帧不具有代表性的缺点,提出一种可变步长提取关键帧提取方法。根据视频变化连续性特点,用相似的两近邻帧代表它们之间的视频片段;该方法首先选取关键帧中的核心区域与受影响较小的边缘区域,对不同的区域取权值并通过转换距离度量分块灰度顺序特征(OM)来判断两帧间相似度;然后利用滑动窗口来查找最大相似匹配,从而检测出查询视频中的拷贝片段。在网络数据和MUSCLE-VCD-2007数据上的实验结果表明,该方法相对于现有的基于OM特征拷贝检测法而言,其鲁棒性更强,检测效率更高。

To weaken the influence of unrepresentative frames and time complexity on detecting the copy of the video on the network, a quick and efficient variable-length step algorithm was proposed to select the most representative key frames.According to the characteristics of the video continuous change, the highly similar adjacent frames could be used to replace the corresponding video between them. Firstly, the algorithm selected core region and less affected edge region of key frame and allocated different areas with different weights so as to use the Ordinal Measures（ OM） transformation distance measurement to judge the similarity between two frames. Then the sliding window was used to find the most similar match to detect the copy clips of query videos. The experimental results on actual network and MUSCLE-VCD-2007 dataset show that the proposed algorithm has stronger robustness and higher detection efficiency compared with the existing algorithms of copy detection feature measure based on OM.