The User-Pleasant Video Skimming by Multi-Sources-Indices Analysis

Video skimming is the process to extract the most significant content of the video and represent it in a concise form. Previous research of video skimming generation could not make full use of the multi-modal information and thoroughly consider the user preferences. In this paper, we propose a novel approach of video skimming through exploiting the fusion of video temporal information and semantic information, supported by audio and text classification and keyframe extraction. By means of analysis on multi-modal video information including audio, text and visual indices, one video skim is generated for one original video. In addition, we introduce the brand-safe filtering and sentiment analysis in order to only reserve the user-pleasant content. Experiment conducted on Youtube-8M dataset has proved that our method highly outperforms the approaches which only consider partial information in video. Moreover, with brand-safe filtering, the unpleasant content has been successfully removed.