Research On Video Copy Detection Technique Based On Multiple Feature Fusion

With the rapid development of web video services, the web video data presents an explosivegrowing. Among these huge volumes of videos, there exists large numbers of duplicate andnear-duplicate videos, which makes copyright infraction and data redundancy become more andmore abominable. Therefore, how to detect copy videos accurately and efficiently on the Internetis becoming a very important issue to be researched. On the basis of further researches on themeanstream video copy detection algorithms based on global feature and local feature, this paperimproves the existing algorithms from effectivity and efficiency. Then, a video copy detectionsystem based on multiple feature fusion is designed to solve the problem which the single featureisn’t robustness and distinctive enough to detect copy videos of variant transformations. Themain contributions of this paper are as follows:(1)In global feature area, the video copy detection algorithm based on the core area ordinalmeasure feature and Transformation distance is proposed. Firstly, after analyzing near-duplicateson real network and selecting the comparatively invariant area of copy video to extract ordinalmeasure feature, which can enhance the robustness of ordinal measure. Secondly, according totransformation mechanism of video copy, the ordinal measure feature similarity measurementmethod based on Transformation distance is proposed and corresponding fast matching approachis designed. Experiment results show that this algorithm can detect copy video of commontransformation types fast and effectively.(2)In local feature area, the video copy detection algorithm based on multiple bags of visualwords model and E2LSH are proposed. Firstly, the AP algorithm and random grouping are usedto cluster large-scale SURF features in this method, which builds multiple bags of visual wordsand overcomes the problem of synonymity and ambiguity on visual words. Secondly, accordingto the distance ratios between a SURF feature and visual words, the mapping range is enactingindividually and the SURF feature is mapped to all the visual words in this range to improveaccuracy of feature mapping. Finally, the E2LSH is used to improve time efficiency of framematching, and the principium of the temporal coherence property inherent is followed to fusematched frames into video clip to detect and locate the duplicates. Experiment results show thatmultiple bags of visual words model can increase the possibility of feature matching and videocopy detection method based on multiple bags of visual words model and E2LSH is better thanthe existing method.(3)In multiple features area, the video copy detection system based on multiple featuresfusion is designed. First of all, the method based on multiple bags of visual words model and E2LSH is used to mainly detect copy videos of non-linear transformations in query videos andfilters out the most of legal video. Then the method based on the core area ordinal measurefeature and transformation distance is used to mainly detect copy video of linear globaltransformations in doubtful videos, and the system uses the feedback information of first-leveldetection to accereate the core area ordinal measure feature matching. Experiment results showthat multi-feature video copy detection system can effectively promote detection accuracycompared to the method based on singe feature, and the detection efficiency is higher comparedto the similar algorithm.