| In order to make use of the increasing number of vast amount of video resourcesefficiently, not only a expected video retrieval technology needs to be found fast, butalso a video abstract technology that can brow the video content quickly is needed.The main use of video abstract includes fast browsing of monitor video importantinformation, video management of large websites, and other various fields. This paperrealized the layered extraction of video abstract, which is mainly divided into shotboundary detection, extraction of key frames, fast hierarchical expression and randombrowse of video abstract.For shot boundary detection, in order to improve the real-time performance andreduce complexity, this algorithm is mainly divided into three steps. Firstly, theblock-based color histogram is used as features to compute frame difference curvewhich is based on double multi-scale that is to use different block sizes in space anddifferent intervals in time, and the size of the block should match with time step. Thismethod can not only ensure enough tolerance for sports, also ensure enough spaceinformation. Secondly, in order to avoid the influence of sudden flash and camerashake, considering the stability of shot change on adjacent scales, the framesdifference product curves based on multiple small step length are used for themutation detection and the frame differential curves based on multiple big step lengthare used for the gradient detection, which can not only locate accurately, but alsoavoid the interference. Finally, the double thresholds are used to cut shots and locatethe potential gradual shots, and the triangular norm is used to verify the suspectedgradual fragment and locate it accurately.In the aspect of key frame extraction, considering the content similarity and timecorrelation, the solution is designed to extract the key frames using hierarchicalclustering. Firstly, time distance is introduced in the first clustering, using thetop-down split hierarchical clustering algorithm to get a micromesh clustering resultsin shots based on gray pattern and time correlation. Secondly, the second clustering was based on the clustering results of the first time, which not only considered thesimilarity of gray pattern, but also the duration information corresponding to eachcluster, which adopts bottom-up agglomerative hierarchical clustering, clustering thekey frames with the similar time and the separation of content together in to onecluster. In the process of clustering, we should consider the two clustering endconditions of the cluster number and the distance between the clustering. The pointwhich is closest to the clustering center was selected as key frames.In terms of video abstract extraction, after extracting key frames in the shots, thenumber of key frames is still very large, considering agglomerate hierarchicalclustering between the shot key frames in order to realize the layered expression ofthe key frames. On the design of the interface, using the tree control, layeredrepresentation and the random browsing of the key frames were realized. Theexperimental results show that the design has strong practical value. |