Research On Video Summarization Method Based On KeyFrame

In recent years,development of digital technology has made video a valuable information resource.With the rapid popularization of multimedia equipment with video shooting function,the amount of video data has shown exponential growth in education,entertainment and multimedia applications.However,the huge amount of data makes the task of video retrieval especially difficult and faces huge storage pressure.People's attention has focused on how to quickly browse and find the video resources they need.The research on relevant technologies that can automatically,effectively and accurately manipulate these video data has gradually attracted people's attention.The video abstract based on key frames can meet the above requirements.Its basic idea is to extract some representative video frames from the original video,which reflect the main content of the video,so that people can browse the sequence of video frames to obtain key information and improve the efficiency of video management.In the keyframe-based video summarization,how to select distinctive frame image features is the basis to improve the quality of video summarization.How to select representative keyframes from the original video is the key to improve the performance of video summarization.The main work of this dissertation is video feature extraction and keyframes extraction,the main research contents are as follows:(1)Feature extraction: at present,most researchers choose manually pre-specified feature extraction rules to describe video frame information.However,due to the diversity of video content,specific extraction rules cannot be applied to all videos.So it is necessary to extract more robust features.Based on this,this dissertation proposes a feature extraction method based on stack autoencoder.According to the property that the autoencoder can learn the features in data unsupervised,this method constructs a depth neural network model and trains the stack autoencoder with greedy training method layer by layer to form a feature extraction model,thus performing feature extraction on frame images.The method can autonomously learn the feature space to describe the information of the frame image with the change of the video content.Experimental results show that the features extracted by this method have better discrimination for more video content,which makes it possible to generate high-quality video summaries.(2)Keyframe extraction: There are some problems that the number and center of clustering need to be predetermined and the selected keyframe is not representative of keyframe extraction based on clustering algorithm.To solve these problems,a method of keyframe extraction based on hierarchical clustering and TextRank algorithm is proposed.The hierarchical clustering algorithm is used to cluster video frames,the TextRank algorithm is used to select candidate key frame sets,and the optimization function is solved to select the final keyframe to generate video summaries.Verification is performed on two benchmark data sets with different video types.Experimental results show that the video summarization composed of keyframes extracted by this method can express video information comprehensively and accurately with low redundancy.