Micro-video Event Detection Research Based On Multi-view Low-rank Representation Learning

With the rapid development of the Internet and social platforms,users are passing information on social platforms through various media every day.Micro-video,as a new form of social media,is becoming popular because it adapts to the fast pace of modern society.Different from traditional video,micro-videos are generally created by individual users with very limited time lengths.At the same time,with the increase of the number of micro-videos,it is essential to classify micro-video,which can help us make personalized recommendations for users among the huge number of short videos.In this paper,we focus on the complex event detection of micro-video.Event detection is a task that requires searching for specific events from a large number of videos.Compared with the recognition of objects and actions,it is more abstract and complex involving the interaction of many low-level semantic information,such as scene,task,target,etc.In addition,they cannot provide enough valuable information,which increases the difficulty of this type of task.Therefore,we need make full use of the existing information.Besides,the existing event detection database is for traditional video and there is not any event detection database for micro-video.To deal with the abovementioned problems,we first proposed a multi-view microvideo event detection model based on low-rank representation learning.The proposed model can not only maximize the complementarity and relevance between views,but also embeds the learning of lower-dimensional subspace structure.On this basis,we introduce a robust elastic regularization network to learn the potential label matrix to prevent overfitting.Secondly,in order to further solve the linear problems,we further construct a new feature matrix through kernel function,and extend the proposed model to a nonlinear form,so as to improve the ability of the model to deal with highdimensional nonlinear problems.In addition,to better verify the validity of the proposed model,we crawl a large number of micro-videos from Flickr and filter them to build a database for micro-video complex event detection.Experimental results on the database show the effectiveness of the proposed model.