Research On Static Video Summarization Technology Based On Key Frame Extraction

With the popularization of the Internet and the gradual maturity of video capture technologies,the digital video has emerged significantly in numbers.In order to search for the important content from massive video quickly and accurately,video summarization technology is generated in this case,by perceiving main content from the video with the limited key frames.However,the existing methods of video summarization have little discussion on the measurements of key frame similarity.Besides,most of the image similarity calculation methods are mainly based on the traditional image features,paying little attention to the topological structure of image pixel space.In order to solve the above issues,this thesis mainly focus on the static video summarization technology,involving key frame extraction and key frame images similarity calculation.The main work can be summarized as follows:(1)Based on the analysis of optical flow motion,a key frame extraction method combining optical flow technology with the improved hill-climbing algorithm is proposed.Firstly,the motion curve of the video frame sequence is calculated using the optical flow method.Secondly,the initial search point is presupposed through the improved hill-climbing method,guiding the search from the local minimum value of the motion curve to a better solution space;then,a variable step search is adopted to speed up.the converge to the local optimal solution.Finally,the video frame corresponding to the local minimum is extracted as a key frame.This method extracts key frames according to the intensity of optical flow motion between successive frames.The key frames not only cover the video content more comprehensively,but also highlight the importance of content;at the same time,it can be applied to the fast browsing and retrieval of the video.(2)A key frame similarity calculation method based on superpixel segmentation is proposed.This method uses the superpixel segmentation algorithm to locally cluster the pixels of key frame images,which can elevate the pixel points to the image region with more semantic space.In this way,the regional topological relationship between pixels can be effectively utilized to achieve the accurate comparison of image blocks.This method is used to calculate the similarity of the extracted adjacent key frames and compress the similar redundant key frames.At the same time,the video main information that people are interested in will not be omitted,so as to obtain more effective and better performance static video summary results.(3)This thesis proposes a static video summarization method to complete experiments on two public benchmark datasets,namely the OVP dataset and the YouTube dataset,and compares it with several representative static video summarization methods.Through the subjective presentation and objective performance analysis,it is proved that the video summarization generate in this thesis is more consistent with the user summarization and shows better performance than the compared methods.The above results verifies the validity and advancement of the proposed method.