Near-duplicate Detection In Large Scale Video Dataset

In recent years, near-duplicate video detection has become an active researchin video retrieval, and also has broad application prospects in copyright protection,video surveillance and topic detection and tracking. Given its applicationrequirements, it is quite necessary to research a near-duplicate video detectionalgorithm with higher accuracy and faster detection speed. In this thesis, majorwork for the goals and requirements are completed as follows:1. The overview of the research significance, research status at home andabroad of near-duplicate video detection and the existing three algorithms is given.The implement steps of near-duplicate video detection are introduced after the keytechnologies of image similarity detection. Shot boundary detection, key frameextraction, feature extraction and indexing structure are introduced in details.2. A cascaded near-duplicate video detection algorithm with temporalconsistency feature in the shot level is proposed. After a further study based on theanalysis of the existing algorithms, we borrow ideas from combining global andlocal features to get a cascaded detection. We define a new temporal consistencyfeature to represent the shot based on the inherent continuous consistency of videosequences. we first extract the temporal consistency features in the shot levelbefore we extract features of key frames to preliminarily filter out some videosdissimilar with the query video. Then we combine global and local features tofurther accurately detect the remaining videos.3. We use the improved edge contour difference method to detect shotboundaries. An overlapped image blocking method and the SURF algorithm areintroduced to extract global and local features respectively for better accuracy. Wealso use the k-d tree indexing structure to achieve fast nearest neighbor search.Through three forms of experiments, we verify the usefulness and accuracy ofthe proposed method. We experiment the near-duplicate video detection on thelarge scale video dataset CC_WEB_VIDEO. From the performance comparisonswith the existing three commonly used methods and the analysis results, theproposed method can achieve better detection accuracy, especially for the videoswith complex motion scenes and great frame changes.