Research On Video Copy Detection Methods Based On Non Negative Sparse Coding

With the rapid development of Internet and digital media, network video has been showing explosive growth, how to manage and effectively protect the copyright of network video has attracted wide attention. Content-Based Copy Detection(CBCD), as an important method to(solve the problem), has become a research hotspot in the fields of multimedia information processing.In the field of video copy detection, the bag of words model and spatial Pyramid matching model have been applied widely and have made a certain progress. The bag of words model and spatial Pyramid matching models are typically using K-means method to generate visual dictionary and using vector quantization(VQ) technique for feature mapping.However, because K-means clustering method is so sensitive to the initial center that it’s not easy for it to get the global optimal solution,and that would seriously affect the formation of visual dictionary. The vector quantization techniques we commonly used have a big constraint on experiment results cause its mapping accuracy. The sparse coding,as a kind of neural network methods, has been successfully applied to the image field and obtained a high accuracy of image classification. This paper presents a copy detection method of non negative sparse coding based on video. The main innovations are as follows:(1)We propose a non-negative sparse dictionary learning method to solve some problem that exist in visual word bag model such as the instability of clustering and the synonymity and ambiguity problem between visual words.Due to the experiments on holiday image data sets we can see,compared to the visual words which formed by clustering or some improved algorithms,the method in this article has a better solution on synonymity and ambiguity problem in visual words.The experiment results show that compared to the KM-BoVW-HA and KM-BoVW-SA methods,our method has respectively improved it’s query accuracy to6%,2.2% and it’s recall accuracy to 12.8%,7%.(2) While using the visual word bag model for feature extraction towards key frames,we cannot get a high accuracy,for solving this problem,we used non-negative sparse coding model and pooling method,and with the application of non-negative sparse dictionary,we also used non-negative sparse expressions on key frames.Compared to the vector quantization method in word bag model,our method had a better performance on our data set experiments,it’s cost of normalized detection NDCR decreased 23%, and1 F improved about 7.79%.