Content-Based Video Copy Detection And Trademark Number Recognition

With the development of network and multimedia technologies, a large number of video data was released and spread every day, thus copyright protection of digital video has gradually become an urgent problem to video provider. Digital watermarking, a traditional copyright protection method, is difficult to deal with such situations. Therefore, as an alternative technology, Content-based Video Copy Detection (CBCD) has received great attention.In this paper, a novel framework for CBCD is proposed. It mainly includes three parts:low-feature extraction, video matching including frame matching method and segment matching method, hierarchical fusion scheme for matching results.In the low-level feature selection and extraction, robust global features and local feature are extracted and combined to describe video contents. The Bag-of-visual-words algorithm is proposed to effectively quantify the high-dimensional key-point feature Speeded Up Robust Features (SURF) and the binary signature is proposed to improve the matching precision. In addition, a density sampling method is proposed to improve the generation of visual codebook. The experiment shows that our method is robust and effective.In the video matching strategy, global feature (Block-based Gradient Histogram) is used for coarse matching and SURF and Intensity Ordinal Measurements (IOM) are used for pure matching, then we combine the two matching strategy together as frame matching method. Besides, Smith-Waterman algorithm is used to realize segment matching.Layered fusion scheme is used to refine the detection results. In this paper, we propose a layered fusion scheme to improve final results precision. IOM and SURF are separately used in different fusion phases. Experiment in TRECVID 2008 shows that the algorithm is high effective and robust.A novel commodity trademark number recognition method based on SVM is provided in this paper, which can automatically locate trademark and recognize it. We extract some visual features such as Haar-like, block-based gradient histogram to train a multi-class SVM classifier. Experiments on 600 characters from 60 pictures show that our algorithm is effective.