Video Summarization Algorithm Based On Temporal Subspace Clustering

With the explosive growth of mobile phones and other directional camera devices,people are capturing and storing more and more video data.Due to the high degree of information redundancy caused by big data,people hope to more fully and intelligently utilize the large amount of information and the parts of interest in these data information.So we need a way to quickly browse and understand the content in the video.Automatic generation of video summarization is one of effective techniques to tackle these problems which extract succinct summaries to represent the original long videos.It involves two problems: video segmentation and summary generation.Most previous works just focused on addressing the second problem by exploiting a simple strategy like boundary detection to segment videos.However,this type of approach leads to suboptimal results because they not only lack of learning mechanism in video segmentation stage,but also separate the whole task into two independent stages.Considering that the video has time continuity,in order to get a valid segment of the video,we cannot ignore the important information of time at all stages of video processing.The classic video digest algorithm has gradually evolved from unsupervised to supervised machine learning,making full use of human self-cognition to guide video algorithm research.In this paper,we propose a novel structure-transfer-driven temporal subspace clustering segmentation(STSC)method for video summarization.This is a video summarization algorithm that uses a priori knowledge for supervised machine learning.We first learn the structure information from source videos and then transfer it to target videos.Reusing the structure matrix as the temporal subspace clustering to segment the video in time domain.By the Determinantal Point Process(DPP)algorithm,we select an informative subset of shots to create the final video summary.Experimental results on SumMe and TVSum datasets demonstrate the effection of our method,against state-of-the-art methods.