Fast Compressed Domain Copy Detection with Motion Vector Imaging

With an increasing number of videos uploaded to the Internet, how to fast detect copy videos in compressed domain has been paid greater attention to. Many researchers have tried using information in motion vector to be the feature. However, in these methods motion vectors are used as histogram, which lacks structural information in detail. To address this problem, in this paper we propose a new way of using Motion Vector Imaging. We first extract motion vector from a compressed video, and then project them onto a canvas to generate a MVI which contains detail motion information. Based on these MVIs, a siamese deep neural network is utilized to train on pairs from dataset and one side of the network is applied to extract features. Finally, a cascade system using MVI model and I frames is used to do fast copy detection. Results on public dataset CC_WEB_VIDEO show that MVI can achieve high recall rate and precision rate at a high speed.