Research On Dynamic Summarization Generation Algorithm Based On Video Content

With the explosive growth of all kinds of video data on the Internet,how to quickly understand the main content of video and shorten the browsing time is an urgent problem.Video summarization generates the short video that can represent the main content of the original video by extracting the important frames or fragments in the video,can provide a way for people to quickly understand the content of the video,so it has become one of the research hotspots.Based on the analysis of the video content,this paper studies the dynamic video summarization generation algorithm and obtains the summary video that can describe the overall content of the video.Dividing the video into several segments,estimating the importance of all video frames to measure the importance of each segment,and selecting the video segments with high importance to generate summaries has become one of the mainstream research directions of video summarization algorithm.In the whole process of this method,the quantitative evaluation of video frame importance is the key to solve the problem.In view of the problem that only video frame image features are extracted while motion information in video is ignored in the feature extraction process of video summarization algorithm at present,the optical flow image is obtained by calculating the optical flow information between adj acent frames of video,and the 3D convolutional neural network is used to extract the optical flow image features and introduce motion information.The two-stream feature fusion module is constructed to effectively integrate optical flow features and image features to better represent the video content.In order to obtain the timing information and consider the different correlations between the current frame and other frames,this paper combines the Bi-LSTM and the Self-Attention mechanism to estimate the importance of video frames,so that the video summarization model can measure the importance of video content much accurately.In the video summarization algorithm,the difference of importance scores between frames is low when the importance of video frames is estimated,which makes it easy to fail to select the truly important video content when generating summarization.Aiming at this problem,this paper introduced variance to measure the score difference between frame,to join the constraints that contains the interframe score variance in the loss function,make the model on the estimate to the importance of the video content can take into account the distinction between the score,and increase the score difference,the selection of the important content of the video in the paper.Finally,experiments and analysis are performed on SumMe and TvSum standard datasets,which fully verify the effectiveness of the proposed algorithm.