Research On Video Summarization Method Based On Convolutional Neural Network

With the development and progress of social technology,smart phones and camera tools have become more and more diversified,which has led to an explosive growth of short films taken by people on the Internet,and the number of videos uploaded by video websites every day is huge.In addition to video files uploaded on video networks or software,a large amount of video data is involved in daily life and online chat.In the era of the rapid development of the Internet,various videos can be retrieved and viewed,but this also brings many problems.On the one hand,the accumulation of a large number of videos makes people unable to quickly find the videos they want,and people's time and energy is wasted to the repeated scenes in the videos;on the other hand,the explosive growth of video data has also brought tremendous pressure on storage.To solve these problems,video abstracts came into being.In this paper,the existing video summarization model has the problem of inaccurate key frame selection and the problem of how to select the video frame feature.At the same time,the image memorability and image entropy of the video frame were extracted by Fei et al.to generate the video summarization model.Inspired by,this paper proposes a multi-feature-based video summarization generation model(DME-VSNet),The features of video frames are extracted by model,including importance score,image memory strength and image entropy.The model consists of three parts: The first part is aimed at the problem of inaccurate video shot segmentation,video shot segmentation algorithm is proposed based on the Trans Net network in this paper.The output result of the Trans Net network is processed by the algorithm to obtain the shot boundary and divide the original video into several short pieces.The second part first proposes that the importance score of the video frame is extracted by the D-IRV2 network,then the image memory strength of the video frame is obtained through the Mem Net network,and finally the image entropy of the video frame is obtained through the image entropy algorithm.The third part is to input the three characteristics obtained above into the MLP architecture proposed in this paper to obtain the video frame score,and the key frames are selected to generate the video summary through the score.In order to verify the feasibility and effectiveness of the video summarization model based on the convolutional neural network proposed in this article,and to verify the impact of different video shot detection technologies and different feature extraction on the video summarization model,first,video shot segmentation method based on Trans Net network is used to segment the video on the Sum Me dataset,and the effectiveness of the segmentation method is verified through comparative experiments.In addition,the overall model based on convolutional neural network is used to segment the video shots of the Sum Me dataset and focus on the key points.The evaluation results of the video summary generated by extracting the importance score of the video frame,the image memory strength score and the image entropy are better through experiments.