Research On Effective Video Copy Detection Based On Shot

With the development of computer technology, the number digital media increaseamazingly, followed by a series of problem about copyright. In application about copydetection, contented based copy detection show better performance in both robustness andapplicability, thus it is considered as substitution of watermarking technology. In thispaper, we propose shot based video copy detection, which apply the model ofbag-of-words in text retrieval field to detect copies.Firstly, according to the realistic condition of shot segmentation, we delete or mergesome shots to make the result of shot segmentation more reasonable, and decrease thenumber of index, accelerating the speed of copy search. After that, we extract local featureof every key frame in every shot, and experimentally choose the dimensionality ofdimensionality-reduction, which reduce the dimensionality as much as possible whilemaintain good discriminability. Using this way of dimensionality reduction to generate thekey-words in the model of bag-of-words guarantees that the generation of each feature’scorresponding key words only rely on the content themselves, which enhance therobustness of key word.Due to some special attacks and the drawback of current shot segmentation algorithm,the length of copy video and original video may differs a lot, which leads to the situationthat two video with very different length are copies of each other. To tackle the problemabove, we propose two kinds of methods for similarity evaluation in detection stage. Sincemost result shot are matched properly, we adapt method which enlarges the differencebetween similarity of copy shots and similarity of non-copy shots, increasingdiscriminability and enhancing recognition capability. For the situation that length of shotsare very different, we use method which decreases the influence of shot length onsimilarity for evaluation, improving the precision and recall of our system.The experiment shows that our methods of dimensionality reduction and visual wordgeneration do well in video content representation. Besides, only one similarity evaluationmethod has achieved good performance and the application of the other method for secondjudgment improves the recall and precision to some extent.