Research On Analysis Of Video Semantics And Recommendation

With the rapid development of Internet,more and more video appear on the Internet.In the field of scientific research and business,there is an increasingly demand for video information.For most of Internet user,it's not easy to rapidly and accurately find out the favorite video from the mass video in database.Therefore,video recommendation has become a hot research direction.Besides,more and more users like to upload videos shooting by themselves,and great video websites pay more attention on these original videos.However,traditional collaborative filter recommendation don`t work well on these videos with no captions or other text descriptions.Therefore,this thesis is intended to mine video semantics and use semantics to recommend videos.In this thesis,the main work is as follows:Firstly,this thesis define the structure of video semantics(we mainly discuss sports video in this thesis)and introduce their importance on video indexing and video recommendation.Then,this thesis introduces two method to extract video semantics.First one is a supervised learning method,it use three dimensional convolution neural network to extract the frame semantics and use connectionist temporal classification to integrate the frame semantics into video semantics.Experimental results show that this method improves the accuracy of three dimensional convolution neural network in video semantic extraction.The other method is an unsupervised learning method,it also use three dimensional convolution convolution neural network to extract the frame semantics.Faced with problem tha C3D-CTC doesn't work well on no label videos,we use deep auto-encoder to integrate frame semantics into video semantics.Experiments show that this method is better than the method of key frame extraction in clustering of video semantic features.Based on two method of extraction of video semantics mentioned above,this thesis introduce a method of recommendation based on similarity of video semantics to tackle the "cold start-up".Then,this thesis introduce the importance of video descriptions on tackling "cold start-up".And integrate two method into a multisource based recommendation.It is proved by experiments that this method has a certain improvement on the recommended accuracy compared to traditional video recommendation method.