Research On The Extraction Of Fingerprint Features For Video Management

With the rapid development of the Internet,people’s demand for videos is increasing which results in the exploding of the amount of Internet videos.As richer videos satisfying people’s daily demand,it’s also a great challenge for us to manage and index them.Due to the large video database,it’s rather hard to protect the copyright of videos and detect the health and legality of the content in videos.Besides,for the high openness of Internet,people can upload videos to video website re-edited by advanced techniques of clipping and handling,which brings great obsession to video indexing.It comes into new difficulty for supervision of videos in network effectively in masses of videos.Feature extraction based-on content is becoming the main mean to detect and index videos.By handling the features can produce a unique ‘fingerprint’ that can represent videos.For the well discrimination of the ‘fingerprint’,we can use it to manage and supervise the video library.In this paper,we get features of video content by SURF and then use visual vocabulary model to generate video fingerprint.A video is composed by a group of continuous pictures,which contains abundant content and great amount of data.For solving the problem of re-editing,the video fingerprint must be robust and precise.But,in practice,it needs to be real-time to generate and index fingerprints.In this paper,we analyze the algorithm existed and discuss the insufficient of them.In order to improve the efficiency of fingerprint generation,we do some work as follows:First,we analyze the trends of network videos and problems we face at present.And,we explaine the background and meaning of this paper.At the same time,we introduce status of research and development about managing and indexing huge amounts of videos at home and abroad.Second,we explain the techniques about video managing and matching based-on content.And we discuss the application in copyright protecting of digital watermarking and digital fingerprint that are both based-on information hiding respectively.Then,we use shot detection and SURF to complete extraction of key frames and features.Third,in order to speed up the generation of video fingerprint and decrease the counts and storage of computing,we use sparse coding to handle SURF features so that we can use little nonzero values to represent original data.Then,we verify our algorithmby Matlab platform and test the performance of the result.Fourth,we use Kmeans clustering to train the sparse features which were extracted from library of Caltech-101.To every feature by sparse coding,it can find a visual word from the dictionary to represent itself in the frame.Then,we statistics the frequency of all word in one frame,and we hashed the frequency result to a sequence that is called fingerprint of this frame.We cascade all key frame fingerprint in chronological order and we get the fingerprint of this video.In the end of this paper,we use Matlab and Visual Studio to test the robust and accuracy of improved results.Last,we build a platform based on model of browser to server by Java and MySQL for indexing video fingerprint sequences.The technologies of extraction for video fingerprint explored in this paper and the sequence alignment existed are added into the platform.The users can complete the search of videos based on content in explorer.Besides,we offered administractors a port to update the videos and the fingerpr--int.