Research On Large Scale Near-duplicate Video Retrieve Based On Contents

In recent years, with the rapid development of computer and networkcommunication technology, the type and quantity of information on the Internet isgrowing. Large-scale content-based video retrieval approximate repeat is becomingincreasingly important. Although many methods have been proposed to come to solvethis problem, they are mainly concentrated in the accuracy of the index. Not a lot ofresearch done for the real-time search of the database of the large-scale network video.In this case, this paper researching focuses primarily on the retrieval speed and accuracyof the feature extraction and database indexing techniques in order to achieve real-timelarge-scale network video retrieval.In the first part we describe the basic technology for video feature extraction.Video data is unstructured data, which is different from structured data (text). Here weintroduce video data from the color characteristics, texture characteristics, shapecharacteristics and regional target. Next we do an overview of the video and indexingtechnology. This section describes the video features index and related concepts, thebasic structure of the index, the index model. Then we propose order feature extractionmethod, and the introduction of LBP (local binary pattern) modal to do the furtherprocessing of the order characteristics. Keyframes are divided into primary part andsecondary areas part. The primary part contains main message and does not easy to becontaminated. Order characteristics based on the LBP extraction methods is not onlygood robustness but also contains a unique video time characteristic color change. Timecharacteristics are got by taking two key frames’ main regional behind a key frame. Thenext section describes the comment indexes, the characteristics of the index and thefield index method and characteristics. Then introduced the excellent performance ofthe text index inverted index method, in combination with the histogram intersectioncore (histogram intersection kernel) introduced this approach to video index. Then theidea of using fast histogram intersection (fast histogram intersection) to improve thestructure of the inverted index, thus reduce the query comparison operation, therebyimproving the retrieval speed. In the part of system’s implementation, the first describesthe architecture of the system, and then elaborated key frame extraction, featureextraction, video indexing, and query module implementations and specifics.Finally, the experimental part in order to prove the effectiveness and efficiency ofthis system, I am in an open network video library (probably about ten thousand video)to evaluate the system. Next, from the two aspects of accuracy and complexity wecompare this system with four existing methods. Is very difficult because theestablishment of large-scale video database, analyzed the theoretical the million levelvideo database, under general hardware conditions (1G memory) approximate thepossibility of repeating video retrieval system. System using the proposed method to achieve both accuracy and speed should bein the same open network video database is better than several other methods, and thetime can achieve real-time effects.