Near Duplicate Video Detection Based On Short Video

In recent years,with the rapid development of communication technology and the Internet,information dissemination has been greatly promoted,and multimedia-based applications have been developed rapidly.Video,which carries a large amount of information,has gradually become an indispensable means of communication and entertainment for people.However,due to the diversification of video acquisition methods and the simplicity of video editing,there are a large number of approximate duplicate videos on the Internet,which will not only cause copyright disputes,but also affect the user's retrieval experience and increase the storage pressure for service providers.These problems have brought us many new challenges.In recent years,scholars have proposed different methods to solve the near duplicate video retrieval problem.Early methods mainly define features by manual,and the feature representations of a video are generated by extracting global features,local features or fusing two features of the video' s key frames.Then near duplicated videos are retrieved by feature matching.AlexNet's outstanding performance in the ImageNet Competition in 2012 sparked a trend of deep learning.Scholars constantly put forward new models or ideas to improve and optimize the network model,which has been applied in many practical scenarios.Among them,convolutional neural network has achieved excellent results in image recognition and retrieval.In view of that video is played by multiple frames of pictures continuously,convolutional neural network is also being used in video-related research.In this thesis a near duplicate video retrieval method that combines style features and content features is proposed,and a video crawler system based on near duplicate video retrieval is designed.In near duplicate video retrieval,based on the middle layer of the pre-trained network,the style feature is represented by the calculation of the Gram matrix of the shallow network's output,and the content feature is represented by the deep network' s output,and then the video feature is represented by the way of feature aggregation.Experiments based on the CC_WEB VIDEO data set show that the proposed method has improved the overall performance compared with the current popular near duplicate video retrieval methods,with an average accuracy of 0.982.In the design of video crawler system,this paper describes the detailed design and implementation of video crawler,video storage,near duplicate video retrieval,system management,system monitoring and other modules.Finally we test the system and the results show that the system can work stablly for a long time and distinguish the near duplicate video well.