| The amount of video clips on the web is huge and is still growing very rapidly. There are many duplicates or near-duplicates among them, which take up extra storage space and affect the efficiency of video retrieval. In order to solve this problem, near-duplicate web video detection is necessary. In Addition, this technology also can be used in news search, topic detection and tracking, and copyright detection and so on.This thesis focuses on the near-duplicate web video detection. Traditional near-duplicate web video detection algorithms are mostly text-based, and a large amount of manual annotation is needed. Content-based detection algorithms include global-feature-based methods and local-keypoint-based methods. The global-feature-based method is fast and its performance is good for almost exactly the same videos or videos with simple scenarios. But for duplicate videos with a little changes in contents, such as changes in color or illumination, removal or insertion of frames, it is difficult to detect them by global-feature-based method. The local-keypoints-based method works accurately in detecting videos with complex scenarios, but it is time-consuming. Video retrieval system requires not only accurate rate, but also real-time. To balance accuracy and efficiency, in this paper, we propose a method to detect near-duplicate web videos based on shot level comparison (SLC+NNS). We define a new global feature on shot level--Shot Signature, and a similarity based on shot level comparison (SLC). Furthermore, in order to further improve the detection rate of SLC+NNS method, we use Locality Sensitive Hashing (LSH) algorithm to fast query the approximate nearest shot, and also we propose fast method to detect near-duplicate web video based on shot level comparison and locality sensitive hashing (SLC+LSH). Experiments show that the performance of our SLC+NNS method is better than Sig-Ch method and CONT+CONX method in accuracy and efficiency, especially for videos with complex scenarios. Our method is approximately identical to HIRACH method in accuracy, but ten thousands of times faster than HIRACH method. Further, our SLC+LSH method is many times faster than SLC+NNS with no apparent decline in speed. 30figures,15tables,60references. |
