Research On Video Summarization Algorithm

In recent years,with the rapid development of Internet and multimedia technology,the data of multimedia information has surged.Digital video,as the major carrier of multimedia information,is comprehensively applied to every aspect of life.On one hand,a mass of videos bring convenience to life,allowing people to obtain information in more abundant forms;yet on the other hand,they also bring about huge pressure to video storage,transmission,archiving and retrieval.Therefore,video abstraction technology comes into being.Similar to text abstraction,video abstraction is the summary and generalization of the source video contents,where it picks up meaningful contents to constitute a compact abstraction,by analyzing the source video data flow.Video abstraction can combine with video annotation technology to be applied to video retrieval,and also can be employed as an independent product application,such as movie trailers,to apply in our life.Video abstraction technology is the research hotspot in the field of computer vision at present.There exist two disadvantages in the traditional algorithm of generating video abstraction based on cluster analysis: First,it cannot adaptively work out optimal number of clusters according to different durations of input and different types of videos.Second,only the color features of images are abstracted in the algorithm,but the texture and shape features are ignored,the single feature can't have a comprehensive expression of the visual image information,but also can't eliminate the image noise effectively,this make the poor quality of video abstraction.Against the disadvantages above,the paper proposes a generation algorithm of adaptive optimal cluster number and multi-feature fusion video abstraction.After the video is decomposed into an image sequence,pre-sample processing is conducted first.Then the color,HOG and LBP features of the pre-sampled frames are abstracted,and fused to represent an image.After that,the zero mean is used to normalize the cross-correlation indexes as the standard of similarity measurement between frames,so that the continuous similar frames can be divided into several shots,thus to acquire optimal cluster number.Next,improved k-means++ algorithm is utilized to cluster all the frames,and the nearest frame from the center of the cluster is chosen as the key frame.At last,the normalized variance of the color histogram and gradient directionhistogram of all key frames will be figured out respectively,with filtering out meaningless ones.The quality of video abstraction is measured in the paper based on two complementary indexes of matching rate and error rate.And the experimental results on the data set of TRECVID2007 show that the proposed algorithm owns strong robustness,which has further improved the quality of generated video abstractions.