The Retrieval Method Of Shot Based On Video Content Under Triple Play

Because of the development of network technology, the number of online videoincreases sharply. Under the triple-play, facing a large number of video files how to findthe desired video quickly from the video database have become a key issue. Method ofretrieval on video content splits the video to shot, describe the shot and retrieve thevideo through a shot case.In order to gain a deeper understanding of the content-based retrieval method，analysis video structure deeply, study the video’s three aspects: the theory of key frameextraction, the extraction measure of shot time-spatial feature and the way of measuringshot similarity.(1) Propose an improved ant and agglomeration algorithm measure toextract video feature. Keyframe plays an important role in video retrieval. Use themeasure to extract the color feature, shape and texture feature, and cluster the feature bythe improved ant algorithm to form an initial clustering. Then, apply the agglomerationto optimize the initial clustering to get the final clustering. Last,find the vector in thecenter of the vectors and extract the referred image frame as keyframe.(2) Analysis thecharacteristics of the shot, develop a method to extract temporal feature and spatialfeature. Use the shot’s color conversion to describe the time’s feature and form a colorhistogram by weight.On shot’s Spatial feature,use Wavelet transform matrix todescribe the shot’s properties color,texture, entropy and so on. Process the frame intoblocks according to the color, calculate the barycentre of each block, and the standardaspect ratio of the area in X and Y direction and then weight these statistics to form ahistogram.(3) Discuss the probabilistic distance method to measure the shot’s similarity.Map the vectors of Spatial feature on a high-dimension space with a function ofnonlinear and make the vectors follow to Gaussian distribution.Then according thevectors distribution in the high-dimensional space, calculate the Probabilistic distances of each feature vector groups. The smaller the distance, which means that the higher thespatial similarity. Obtain the temporal similarity by the intersection of the correspondingcolor histogram. Weight the spatial and temporal similarity to draw the overallsimilarities.Through experiments and analysis, content-based video retrieval method canretrieve the desired video well, and more accurate than the method based on thekeyframe. Provide a better service for the video content supervision of the triple play.Then offer some new opinions to further exploration of video retrieval.