Research And Implementation Of A Faster Video Retrieval Method

In recent years,there has been a great increase in multimedia data aswellasinmultimediaaccesstechnology.Thechangescanbefelteverysingleday.However,it remains a great difficulty to search out from the numerousdatabase the veryinformation one needs.There exist nowadays a good manysearch engines for text information,but not for audio or videoinformation.Therefore,it has become a great challenge to reach the videoinformation fast and effectively.The traditional KBVR(Keyword-basedVideo Retrieval),because of its limited description ability,strong sense ofsubjectivity,manual notation request and some other limitations can nolonger satisfythe user's need.future multimedia services should be rapid andeffective content-based retrieval services.CBR(Content BasedRetrieval)which focuses on the color,texture,shape and spatial relation of thesubject to establish the media vectors.This is based on that we can make useofthemedia'smulti-dimensionalcharacteristicsforsimilaritymatching.Videos contain not only what the stationary pictures contains,but alsodynamic information and information that changes as necessary.Being abasic building block of multimedia technology,video has perfected theexpression of information.It will be useful and helpful if one can get thevideo information he needs fast and effectively.Furthermore,video hasbecome one of the main ways through which people communicate with theoutside world.Besides combining image and sound,it can also expressdynamic information.These advantages are perfectly manifested in digitallibrary,in which the readers are able to get detailed information of the wholedatabase.What is mentionedaboveis just a few commonuseofvideos,but isalready sufficient to prove its value.Being so important,video retrieval hasalways been the focus of multimedia retrieval,among which the CBR hasbecomeoneofthemajorbranches.The CBR consists of image inquiry,sample inquiry,video browsing andretrieval,graph and text inquiry.There has already been some test systemsand practical use of CBR,but much more needs to be done for further studyand its commercialization.Take the Informedia Digital Library Program inCarnegie Mellon University as example,only 1000 hours'video database isplanned to be established and that is far from enough for commercial use.InChina,sofartherehasnotbeenasystemwhichisabletosatisfytheuserbothinaccuracyandspeed.Video retrieval is mainly made up of scene segmentation,key framesselection,feature extraction and similarity matching.Certain progresses havebeen made from feature extraction to model retrieval,as far as temporaryresearch work is concerned.But due to its difficulty and complexity,most ofthe efforts are put on the structure analysis of the video such as lensdivision,key frame withdrawing and scene structuring and so on,while thestudy of video retrieval is relatively less,which is undoubtedly the mostimportantinpracticaluse.In this paper,a detailed study of sample inquiry,one of the practical useof video retrieval is carried out and also a fast and effective way of sampleinquiry is proposed.Sample inquiry is that the user submits an image orvideo in order to search out other similar videos.The current sample inquirywill first Structured treatmented of both the sample video and the videos inthe database,then make a comparison of the eigenvalue of their key framesandfinally,takingall the factors that wouldeffect thesimilaritydegreeofthefragments in view to feedback to the user the results according to thesimilarity degree.Peng proposes a maximum matching algorithm based ongraph method and Zhao proposes a Sliding Window Method which is basedon equivalence theory.Both of the methods above work quite well in similarfragment retrieval but need much more improvement in speed in accuratefragment retrieval.(An accurate fragment retrieval needs not only thesimilarity in content and characteristic with the sample video but also thecomplete similarity in chronological order.The similar fragment retrieval isto search out all video fragments that only share similarity with the samplevideo in content,that is to search out all the videos adopted from the samplevideo by slowing down or speeding up or inserting advertisements andsomeothervideoeditingmethods).Afterthestructure analysis andthefeatureextraction,inordertonarrowthe search scope and enhance the retrieval speed,those videos that do notmatch with the sample video should be excluded.In the paper,a synthesisconsiderationofthecharacteristicsofthesamplevideoistaken.Includingthelens length and its distribution as well as its relation in size,the distancebetween two continual lens,the number of key frames in each lens and theaverageofinterframe gapineachlens.That helps toalarge extent totickoutthe non-matching videos.Then one can use the Sliding Window Method tosearch the video out of what is left.In this way,the retrieval speed isremarkably increased but a high level of accuracy is also kept.It will workprettybetterwithlongsamplevideoswhichconsistoffourormorethanfourlens.As at the very beginning,80% of the videos would be excluded due tothe use of the lens length distribution,there will be a great increase inspeed.Forsamplevideos withtwoormorelens,it will not matteragreat dealto the accuracy whether the lens are precisely segmented.One can see fromthe results of the experiments in Chapter 5 that the accurate rate of samplevideos with two or more lens is always above 95%.For the single lens videoretrieval,since the maximum and average of interframe gap and the lenslength are all taken into consideration,with the aid of binary searchdefinitely,thespeedwillbeimproved.The sliding windows shot is to some extent,like arranging apre-selection before the retrieval and is now the most effective way amongsimilar methods.Through the comparison with the Sliding Window Methodonecanseethattheretrievalspeedis enhancedtoahighlevelinthepremiseofaccuracy.Unfortunately,it does not workthesamewell insimilarfragmentretrieval for which we using the Sliding Window Method to ensure highaccuracyinsimilarfragmentretrieval.