For Web Video Research And Implementation Of Data Mining And Retrieval

With the advances in the Internet and multimedia technologies, there is a widerange of information in the web. Especially the visual information covered any aspectsof social life making it more vivid. How to retrieve the video we need from the vastvideos is becoming an important problem. The traditional method to retrieve the videois based on the text, that means people will text annotation for the videos artificially,and then use traditional information retrieval technology to retrieve text. However, thereare several limitations in the traditional retrieval method because of the ambiguousnature of visual information. While the Content-based video retrieval (CBVR)technology can very well solve this problem.CBVR is the technology mainly refers to using features of videos to retrieve,according to shot detection, keyframe extraction and feature extraction. However, thereis a large amount of video data on the Web, and how to provides a more efficient andeffective video retrieval on the basis of CBVR? In order to meet the needs thatmentioned above, we need to organize and index the videos effectively. Therefore, theimportance of data mining on video has become increasingly prominent.This dissertation briefly summarizes CBVR system，and researches some keytechniques which specially focuses on shot detection, keyframe extraction and featureextraction．On the other hand, this dissertation researches the clustering algorithm indata mining to achieve data mining on Web-oriented video data. In this thesis we usekeyframe-based approach for video retrieval. Before the search, the keyframes will bepre-classification, so that the effect of retrieval can further improve. The prototypeimage retrieval system that belongs to our laboratory used artificial methods for imageclassification. It built a basic image gallery with dividing the images into20categories.While in this thesis we propose a new method for keyframe classification, which iscategorizing automatically by computer, but not artificial. Computer will base on thevisual features and the corresponding keyword, using K-means clustering algorithm tocategorize the keyframes.The experiment that based on the image library which has been automatically classified show that we have the same high retrieval precision as the prototype imageretrieval system does. The result demonstrates that this method proposed in thedissertation is effective, and it can overcome the insufficient of the artificialclassification.In the end, a web-oriented video retrieval system is constructed, and the systemmodules、features are discussed too. The experiment with the system shows that thealgorithms proposed in this thesis are efficient and effective.