基于时空域信息融合的视频拷贝检测算法研究

为了有效利用视频独有的时空域特性来提高视频拷贝检测算法的鲁棒性和精度,提出一种基于时空域信息融合的快速拷贝检测算法。该算法包括基于时空域信息融合的指纹提取算法、基于倒排索引的匹配搜索算法和结合异步滑窗策略的基于匹配状态机的匹配搜索算法。指纹提取算法首先将视频分段形成时空域信息帧,然后对该信息帧进行分块,提取DCT系数后,利用其中值进行阈值化得到视频指纹。基于倒排索引的搜索算法根据指纹的二值性特点建立倒排索引表,然后通过索引表快速查询指纹。结合异步滑窗策略的基于匹配状态自动机的搜索算法,利用与最近邻之间的匹配状态来改变搜索范围和步长,而异步滑窗策略通过对在线和离线过程采用不同的提取策略,减少搜索量,加快搜索速度。实验结果表明,提取的指纹对噪声模糊、添加字幕、空间偏移、旋转、掉帧具有较好的鲁棒性,同时提出的搜索方案在时间效率上也有较大的提升。

Research on Video Copy Detection Algorithm Based on Spatial-Temporal Domain Informative Fusion

In order to effectively utilize unique spatial-temporal domain characteristics of video to enhance the robustness and accuracy of copy detection algorithm,this paper proposed a fast video copy detection algorithm based on spatial-temporal domain informative fusion,which includes a fingerprint extraction algorithm based on spatial-temporal domain informative fusion and two kinds of matching search algorithms of which one is based on inverted-file index and the other is based on matching state machine with asynchronous window strategy.The fingerprint extraction algorithm firstly forms the spatial-temporal domain informative frame by video segmentation,then partitions the informative frame into blocks and extracts DCT coefficient with its median value as threshold to obtain video fingerprint.The matching search algorithm based on inverted-file index sets up inverted-file index table by binary characteristics of fingerprint,then quickly queries fingerprint according to the index table.Combining the matching search algorithm based on matching state machine with asynchronous window strategy,we can change search scope and step size by matching state with nearest neighbor.Meantime,asynchronous window strategy can adopt different extract strategies in online and offline process to accelerate the whole search.The experimental results show that our fingerprint extraction algorithm is robust in the case of Gaussian noise,adding subtitles,spatial shift,rotation and frame drop,and the proposed schemes tend to have great improvement in time efficiency.