Research On Video Summarization Technology Based On Content Analysis

With the continuous development of computer vision,video summary technology has been widely used in all aspects of people's lives.At present,the performance of computer hardware constantly improving,and the video data generated in daily life has increased dramatically,providing a good opportunity for video summarization based on massive data research.Video summary refers to extracting representative parts by analyzing the structure and content of the video,and then combining them into a summary in some way.It fully expresses the main content of the video.In the research of video summarization,the lack of description ability of features is likely to cause improper detection of key content,which makes the summarization effect worse.Moreover,the selection of the video summary needs to meet the needs of different users.How to make the generated summary include both the key content of the video and the information that meets the needs of different users.This is the challenge that summarization algorithms have always faced.Aiming at many problems of video summarization,this thesis proposes a video summarization method based on content analysis.Through the analysis of the key steps in the video summary process,the video summary algorithm is improved from many aspects,as follows:(1)We analyzed and summarized the feature construction methods commonly used in current video summarization algorithms,and elaborated on the advantages and disadvantages of some methods.Aiming at the problem of incomplete description of video content,this thesis proposes a method for constructing features of convolutional neural networks,which improves the characterization ability of features.(2)This thesis analyzed a variety of video summarization methods and illustrates the key role of supervised learning methods in algorithms.Aiming at the subjectivity of the abstract and the complicated problem of supervised model training,this thesis proposes an algorithm combining non-parametric supervised learning and adding user supervision information to make the abstract more meet the user's needs.(3)In the steps of video key content extraction,this thesis proposes to use the label distribution learning algorithm to extract key content.At the same time,this thesis uses the same type of video to have the similar video structure,so that the algorithm can handle multiple types of video,which solves the problem of limited video categories to some extent.The experimental results show that the proposed algorithm has good robustness,and the generated summary contains important key content of the video and information that meets the user's needs.