| Video,as a carrier of information,is becoming more and more popular and widely applied in various fields.In the foce of massive video data,how to quickly and accurately fied the required video resources is the current research bctspot.Against this background,videoa sumenarization technology emerges,and has made great progress in recent years.This thesis is aimed at Improving the performance of video summarizaiton.The main work of the thesis Includes the following three parts.(1)A video summarizaiton(?)ethod based on changepoing detection and K-means clusiering is proposed.Because change detection requires smooth data in time,and there is linle difference berween adjacent franes in vidco,so change detection ean be used to shot segmentation.Secondly,considering the temperal characteristics of video franes,a time factor is added to measure shot simailarity in shot merging,and the redundant shots caused by the fast motion ofobjects of the dimming of light are removed Then,the 3σ caiterion is used to determing the namber of clusters of video frames in the shot,and the K-means clustering algonithem is combined to extract the keyframes in each sbct,meanwhile,the mcoochrome frames and similar frames are cilminased,Finaily,the extracted keyframes are formed to video summarization.In addision,our method is evaluated on a benchmark dataset,which has fifty vidcos and comes with five human-crassed summarizations.The evaluation results show that our nethod can generade high-quality summarizations,which compared with human-creale summarizaiton and OV,DT,STIMO and VSUMM 2 methods.(2)A vidco summariation generation method based on feature fusion and Bayesian odine change detection is proposed.Firstly,the HSV quartitative color feature of video fraunes is estracted and a new festure MEP is corstructed.The Euclidcan distance is used to calculate thecomprehensive similarity of two adjacent frames.Then,Bayesian online change polin detestion algorithm is applied to achieve sbot seumentaiton by the comprchensive similarity.The position of change point detected is the location of shot segmentation in video,which realizes real-time shot segmentation and avoids the determination of traditional shot boundary threshold.Secondly,keyframes in shots are extracted by combining frame rate and K-means clustering.Finally,monochrome frames are recognized based on rank of image matrix and similar frames are eliminated.The experimental results show that the quality of video summarization generated by this method is higher than that of other algorithms.(3)A video summary library is established,and videos in the library are classified into three categories by classification algorithm.For each category,a summary of the classified video is generated.Firstly,focusing on single-view video,a summarization library containing 410 videos is established,which contains 14 different types of videos,such as animal videos,road traffic,documentary,home videos,and lasts from 1 minute to 10 minutes.Each video has annotation results from five different users.Secondly,the random forest algorithm is employed to classify the video in the library.According to the video characteristics,the video in the library is divided into three categories:edited video,artificially captured video and surveillance video.The accuracy of the classification reaches 84.15%.Finally,three different video summarization methods are proposed by using corresponding video features for three different types of videos.The experimental results show that each new video summarization method has good performance. |
