Research On The Technology Of Scene Summarization In Video Retrieval

With the rapid development of multimedia technology, multimedia data are increasing exponentially. Consequently, How to store,segment and retrirval the multimedia data effectively have become a urgent problem to solve. Video summarization is an efficient method to solve those problems, and it also can support the establishment of a video retrieval and index system, so it is meaningful for us to study the techology of video summarization.In this work, we firstly discuss the background of our research and then analyze the main existing content-based video summarization algorithms, we propose a key-frame extraction algorithm based on swarm intelligence and k-mean clustering algorithm, a scene fragmention approach and a method of generating video summrization based on visual attention model.Then we design the prototype system.The works are described as below:(1) Key-frame extraction algorithm based on swarm intelligence and k-mean clustering algorithm is presented. First of all, we introduce the swarm intelligence clustering algorithm,and combine it with k-mean clustering algorithm,we propose a key-frame extraction algorithm based on swarm intelligence and k-mean clustering algorithm, Firstly,the clustering based on swarm intelligence is conducted,and an initial cluster is obtained.Secondly,k-mean clustering algorithm is executived to optimize the results and accelerate the algorithm convergence.Finally,the key frames which represent the shot are obtained.our method has solved the problems of the traditional clustering algorithms of key frame extraction effectively,such as sensitive to a premature, easy to a precocious and need to use knowledge to determine the number of categories.(2) In this paper, a scene division method based on the motion character and time information of the shot is proposed. The traditional scene division methods usually only use the visual fenture of shot,it caused the results inaccurate,to solve this problem,when we define the similarity of shots, we combined the visual feature, the motion feature and the time information of the shot. And then use the chain clustering algorithm to cluster the shots,and the similar shots are in the same scene,a better scene division is obtained.(3) A scene summarization method based on visual attention model is presented. The visual attention model is introduced into the scene summarization, by extracting the static and motive significant regional, the visual attention degree of the shot is obtained by combine the static and motive significant degree,and we considered the time information of the shot, then we combined the visual attention degree and the continued degree by the weighted sum, and obtained the important degrees of the shot.According to the important degree of shots,we selected the biggest ones,then output the key frames of their chronologically as the scene summarization.(4) A prototype system of scene summarization by using the object-oriented ideas is designed and implemented.The system includes the shot border detection, key frames extraction, the scene division, and the generation of scene summarization.And the effectiveness of the above algorithms.have been proved through the experiment contrasts.