Research On Distributed Collaborative Processing And Retrieval For Massive Videos

With the rapid development of video sharing website, the number of videos has been increasing dramatically in recent years. How to effectively process and retrieve the massive videos has become a challenging task in the multimedia application field.The main problems of distributed collaborative processing and retrieval for massive videos include:(1) The collaborative scheduling of computing tasks and computing resources in multimedia content processing, that is, how to predict resource requirements of the task and match with appropriate computing resources;(2) The fast similarity-based video retrieval problem for large-scale video database;(3) The fast and effective video copy detection for large-scale video database. The main objective of this paper is to address these issues. Specifically, the main contributions and innovations are summarized as follows:1. With the increasing multimedia content production business, there are more and more compute-intensive and data-intensive computing tasks. We design and implement a distributed computing middleware which automatically manage computing tasks and resources. It uses an execution time estimation algorithm based on Extreme Learning Machine for self-optimizing. It can be used as a standard component in the processing of multimedia content, for the sake of simplication in using lots of computing resources. The middleware has been integrated into a multimedia content production platform. The experimental results show that the makespan of execution time estimation algorithm based on Extreme Learning Machine is reduced by at least26%compared with that of the opportunistic load balancing algorithm for the rendering, transcoding and special effect synthesis tasks.2. To achieve fast similar video retrieval for large-scale video database, we propose a fast video retrieval method based on MD-LSH, which solves scalability and efficiency problems. Firstly, the feature vector set is extracted for every video and the indexes are built by using a distributed memory hash structure, called MD-LSH. Secondly, the similarity of the relevant videos is calculated according to the returned similar frame sets. Finally, the sorted similar video list is returned as the query result. The proposed method can achieve fast video retrieval for large-scale video database in the distributed multi-node environment. The experiment results show that the proposed method is at least four times faster than the existing methods. At the same time, the method adopts distributed architecture to meet the requirements of the scalability.3. To achieve fast copy detection for large-scale video database, we propose an efficient fast copy detection method based on SimHash. This method first extracts spatiotemporal feature based on discrete cosine transform, and then constructs a robust compact signature based on SimHash for every video. Finally, video copy detection is achieved by using efficient hamming distance retrieval in signature tables. The proposed method can achieve efficient fast video copy detection for large-scale video database in the distributed multi-node environment. The experiment on the dataset, which contains10335videos with a size of272.48GB, shows that the F value of the proposed method is0.952and the average retrieval time is at least15.2times faster than the existing methods.The first part of the above research achievements has been applied in the "collaborative computing and resource scheduling" subsystem, which is a part of the National863Project "development of distributed broadband services collaborative making environment".The second and the third parts of the above research achievements have been applied in the "multimedia search in the enhanced search system" subsystem, which is a part of the National Key Technology R&D Project "research on enhanced search system architecture, key technology and test specification".