The Research Of Movie Video Summarization Algorithm Based On Integrated Semi-supervised Learning Framework

With the development of the Internet,the endless variety of smart devices and tools,people can upload data to the Internet or obtain data from Internet more easily.Among these,the video data has become the data which people pay more attention on,however,video data also have the characteristics of big data,rich content(including images,sounds,text,etc.),the temporality,and the properties bring big challenge to the video storage,video retrieval and many other applications Faced with numerous videos which are from 10 seconds long to more than one hour,movie has become one part of people's life,users need to find their favorite film in a short time,they want to know the contents of a two hours long movie by scanning a few of pictures or a short video clip,and then determine whether watch the entire movie or not Based on users' demand,video summarization algorithm is an effective way to address this demand.However,the outputs of most video summarization algorithms are sweeping statement of the long video,or key frames from start to finish,or the video compression,this kind of summarization doesn't highlight the interesting things.And there are few summarization methods for movie data,the summarization of movie data(movie trailer extraction)is on the stage of manual producing,and this procedure needs abundant energy and time.The purpose of this paper is to propose a video summarization algorithm to output movie trailer automatically without manual operation,and the summary is not the sweeping introduction,but the highlight of interesting content,which will attract viewers in a very short time.We first propose an integrated semi-supervised learning framework,which combines the semi-supervised clustering algorithms and the semi-supervised classification algorithms with an assumption of agreement.The classification methods and the clustering methods cover their shortages with each other,and make the predicted results be more reliable.Then we use a few films and the corresponding official trailer as the labeled data,and put the limited labeled data and abundant unlabeled data into the proposed semi-supervised learning framework.We can obtain the predicted label of unlabeled data,and combine the clips which are composed by the positive samples.We first perform experiments on our semi-supervised learning framework with abundant data,and demonstrate that this framework is able to get good classification performance and be robust.Then we do the summarization experiments on several movies,and compare the output trailers with the official trailers,the results illustrates that the generated trailers have high similarities with manual trailers.In fact,we not only propose a simple integrated semi-supervised learning framework,we first apply this on extracting movie trailers automatically,and get a good result.The proposing of this approach lays a good foundation for many video applications,EG video retrieval and video recommendation,and it is very meaningful for the future research of video summarization.