基于多尺度互信息量的数字视频帧篡改检测

Detection of Frame Forgery in Digital Video Based on Multi-scale Mutual Information

针对单镜头视频时域篡改问题,提出一个以内容相似性为基础的视频篡改被动盲检测算法。通过高斯金字塔变换获得视频帧的3种尺度视觉内容,根据信息论定义相邻两帧的归一化平均互信息,采用线性组合构建多尺度归一化互信息描述子,实现相邻两帧多尺度视觉内容相似性的度量。利用局部离群点检测算法计算视觉内容相似性异常度,使用阈值法检测视频篡改位置。实验结果表明,该算法不仅能有效地检测出视频帧删除、复制以及插入3种篡改的位置,而且适用于不同编码格式视频间和同源的篡改。在检准度和检全率上优于现有的时域篡改检测算法。

Aiming at the time tampering problem in the single video shot,a new algorithm based on the content similarity is proposed to detect the tampers of frame duplication,deletion and insertion. Firstly,the three-scale visual contents are obtained by using the Gaussian pyramid transform on every frame. Then,the normalized mutual information is defined on the single scale visual content of adjacent frames based on information theory. And the descriptor of multiscale normalized mutual information is computed by linear combination. Thirdly,the abnormal degree of content similarity is computed by local outlier detection algorithm. Finally,the forgery places are detected by threshold. Experimental results show that the proposed algorithm can effectively localize the tamped position of the frame duplication,insertion,deletion tampers,and can be fit for the forgery of different coding formats and different cameras. The results also show that the proposed algorithm outperforms the existed algorithms in terms of precision and recall.