Topic Preserving Video Compact Representation Research

With the development and popularization of the Internet,people's daily network communication tends to be more intuitive and expressive,such as voice and video.When the 5G coming,explosive growth will occur in network video,vehicle video,security and sports video.The process of video understanding and processing will also be extremely tough.This dissertation proposes a compact video presentation for complex videos,so as to promote other related works.There are ineffective video frames or clips in the video,but the traditional methods do not take this difference into account.The goal of video compact representation is to keep the important part of video content and the same time keep the video category unchanged.In order to achieve this goal,those important video topics that play a key role in maintaining video categories are needed.There are many challenges: how to define semanticlets? How to verify the significance our definition? Whether there are saliency differences between the semanticlets;how to mine the important video topics from the semanticlets,and how to verify the efficiency of the video compact representation? The research of this dissertation based on semanticlets and verify the significance semanticlets,and then propose an algorithm to mine the salient ones.Considering that video topics have the ability of keep video category,a new algorithm based on frequent items discovery to mine discriminative topics of the same video category is proposed;In order to further improve the effectiveness of video compact representation,an adaptive video semantic boundary generation algorithm is proposed.Specifically,the main contributions of this dissertation are divided into the following four aspects:Firstly,this dissertation proposes to define video semanticlets according to the basic concepts of actions and scene formed by videos in the local time region and propose to research the video content based on this definition.The proposed actor identification method based on semanticlets and actor appearance improves the accuracy of actor identification,which demonstrates the significance of the related video based researches.Secondly,this dissertation applies a variety of experiments to validate the saliency of video semanticlets.A salient semanticlets mining method based on sparse coding is proposed to extract the most salient semanticlets within the video.The performance ofvideo representation can be enhanced by assign each semantic weights according to its saliency.Thirdly,this dissertation mines the salient semanticlet with the ability of keeping video category,among videos of the same category.In order to obtain the video topics,serval latent pattern discovery algorithms based on deep visual word encoding are proposed by incorporating with the Chinese frequent words mining algorithms.Experiments demonstrate the effectiveness of the video compact representation method.The proposed method has brought significant performance improvements for event detection and video content analysis.Finally,in order to obtain more accurate and effective video topics to improve the performance of video compact representation,this dissertation proposes a new temporal distribution network for adaptively semantic boundaries generation.Experiments show that the proposed TDN method can generate high-quality proposal with the fastest speed for action detection.The performance of the video compact representation is further improved by applying the proposed method.Through the above research,this dissertation deeply explores video compact representation in each level,and provides practical and effective solutions for the key problems of video compact representation.It can be concluded that: video semantic is the basic concept formed by video in local time region;There are significant differences in video semantics,which are influenced by many factors such as video segment segmentation strategy,representativeness and so on;Weights assignment based on semantic saliency can enhance the performance of video representation;The latent pattern discovery algorithms based on deep visual words encoding can obtain semantic patterns consistent with hunman cognitive;The topic preserving video compact representation method achieves higher performance for event analysis and video understanding;Semantic segmentation method based on temporal action proposal generation can further improve the performance of video compact representation.