Research On Key Technologies Of Video Copy Detection

With the continuous development of information technology,video capture equipment and video editing software,the network videos grow explosively accompanied by a large number of copy videos.The appearance of a large number of copy videos brings a lot of technical challenges to video content supervision,video copyright protection,video search engine ranking and so on.Based on this background,video copy detection technology emerges.On the basis of in-depth understanding of the current research of copy video detection,this paper analyzes the shortcomings of the key technologys of copy video detection.Research further in the following aspects,and make some progress.First of all,aiming at the drawbacks of the need to repeatedly read from the reference library of video keyframes and feature extraction of keyframes when one using the local feature of the one-to-one matching process,this paper proposes to use the local feature of image one-to-many matching algorithm,the algorithm extracts all the features of keyframes in reference library and preserves them in a matcher data structure.For the testing keyframe,extract feature and use approximate nearest neighbor matching search algorithm for keyframe matching.Experiments show that using the proposed algorithm the speed has increased by more than 10 times compared with one-to-one matching algorithm while the recall ratio and precision are basically unchanged.Secondly,because the time consumption of local feature similarity matching algorithm is too large,this paper applies AlexNet convolutional neural network model to extract the keyframe features,and based on the features to match similarity.Experiments show the Top-5 accuracy rate using AlexNet model similarity matching algorithm reaches 97.8% and is far more than the classic SIFT and SURF with the combination of BOF,the speed also has a huge upgrade compared with using local feature similarity matching algorithms.At the same time,according to the special application background of keyframe matching application only required to return Top-1,the fusion of ORB and AlexNet feature similarity matching algorithm is proposed,the Top-5 results returned by AlexNet model are re-ranked based on the use of ORB features,experiments show that this algorithm can improve the accuracy of Top-1.Finally,in view of the problem that traditional copy detection algorithms can not solve the combination of copy video clips and non copy fragments,based on keyframe similarity matching results,this paper proposes a video sub-sequence matching algorithm based on diagram,the algorithm is verified by the the reference library of video set composed of CC_WEB_VIDEO and the first audio and video retrieval recognition challenge in 2014,the Score value can reach 87.26%,the total time to detect 3000 piece of 1 minute test videos is 3.17 hours,each video only requirs 3.8s.