Research Of The Socialization Video Episode Detection And Fine-grained Emotion Analysis Algorithms

With the popularity of social video,many new video research methods have been proposed.Bullet data is applied to social video in recent years.Compared with traditional comment data,it has more accurate time information about the attributes of video time synchronization.Starting from the text and video image of the bullet screen,this thesis first divides the video episode,and then combines the two modal data to analyze the video emotion.The research content of this topic includes two aspects.First,a video episode boundary detection algorithm based on probabilistic graph model is proposed.The aim is to automatically segment video episodes and extract video episodes.Video episode boundary detection algorithm takes bullet text as input,gets the boundary information of video episode,and outputs the subject content of text for the segmented episode.On the basis of video episode boundary detection,the second research point of this topic is multi-modal video emotion analysis under episode granularity.Compared with the coarse-grained emotional analysis of the whole video,the proposed model can give the emotional polarity analysis results for each scenario.At the same time,multi-modal video emotional analysis of film and television plays is carried out by using the text modality of the barrage and the image modality of the time slice,in order to solve the problem of inaccurate emotional analysis under single modality.Combining episode boundary detection algorithm with multi-modal sentiment analysis constitutes the main content of this research.The main video analyzed in this thesis is film and TV series,and the model algorithm is validated according to different types of videos.The experimental results show that the video episode boundary detection algorithm based on the improved probability graph model can segment a video reasonably and give the subject content text for the time slice after segmentation.In addition,the experiment also analyzed the multi-modal affective analysis algorithm under plot granularity.Compared with the single-modal affective analysis algorithm,the multi-modal affective analysis algorithm proposed in this thesis combines the bullet text and the video face image,which can more accurately classify the emotions of video clips.