Research On Key Technologies In The Video Copy Detection Based On Bag Of Visual Words

With the development of Internet technology and intelligent terminal technology, the video information with its intuitive, vivid and strong expressive are widely used gradually, and become a mainstream form of information spread and communication on the network, the security issues of video information resulted from this have become increasingly prominent. Video copy detection technology plays an important role in video information management and copyright protection, and becomes a focus topic in multimedia information processing field in recent years.With the development of Video copy detection technology for many years, a lot of methods are proposed. Video copy detection technology based on visual dictionary becomes the main method as it has high efficiency and suitable for large sale and high dimension. This paper carried out research on two main issues of this: visual dictionary generation and high video feature matching, the main contributions are as follows:(1)A less quantify loss, higher building efficiency method of visual dictionary generation is proposed. the current mainstream method of visual dictionary based on SIFT feature has been researched deeply, with the difficulty of identifying visual dictionary scale, big information loss of feature quantification, thus affecting the retrieval accuracy, we presents a visual dictionary generation method based on an improved affinity propagation algorithm. Firstly, the Locality Sensitive Hashing algorithm is used to pretreat SIFT features. And then use the improved affinity propagation algorithm to generate the visual dictionary with clustering reference video library. Finally a word map chain with hamming embedding is introduced to retrieval, effectively reduces quantify loss of visual dictionary generation and feature retrieval process, improve the retrieval accuracy. The results show that: compared with the existing method, the dictionary generation of this can improve the video copy detection precision, and is scalable to new video added.(2)A less memory consumption, higher retrieval efficiency method of video feature indexing is proposed. With the difficulty of high video feature dimension, large data size and low retrieval efficiency, this paper carried out research on rapid indexing method based on Local Sensitive Hashing. In order to solve the serious memory consumption of Locally Sensitive Hashing, a fast query method for video frame features based on a combination of dynamic hash index is proposed. Firstly, extracting a certain amount of sample from the reference video library to train the hash functions, generating a fixed number of standard hash function in accordance with entropy maximization, to enhance the mapping and division effect on the video feature data of the hash function. Secondly, designing a combination strategy of dynamic hash function to improve the efficiency of feature point returns and greatly reduce the required memory. The experimental results show that: this method is more efficient in memory consumption when the scale and dimension of the data increases, which can effectively improve the retrieval efficiency of the video copy detection.(3)Based on the method or algorithm before, a hierarchical based video copy detection system is designed. Firstly a fast indexing method is used to filter the reference video library. Then an exact match is given to get final confirmation. Experiments show that the hierarchical based video copy detection system can effectively improve the detection efficiency while ensuring a good detection accuracy.