Study On Video High-Level Feature Extraction

With the rapid development of information technology,multimedia data of huge amount is getting more and more involved into people's daily life,bringing us the challenging task of efficiently finding information among the specific dataset.Of the researches of multimedia information retrieval,content based video retrieval is most representative as its related content has the characteristics of large scale,complex relationship and multimodality.Our paper contributes to this field,and focuses on the technologies of video high level semantic feature extraction using a metric learning approach,as well as the according application system.In the context of content based video retrieval,K Nearest Neighbor algorithm is most effective for finding related data with high level semantic approximation by query examples,which requires for similarity measurements that could well describe the feature space,leading to the field of metric learning.Our paper performs a systematic survey on supervised metric learning algorithms,and optimizes the Large Margin Nearest Neighbor algorithm - which is fitted in the SVM like large margin framework and possesses the state-of-art performance - regarding the scale of video data,in order to achieve higher efficiency when integrated into retrieval tasks.On the other hand,one important goal of content based video retrieval is to reasonably organize and represent the useful information to the users.Our paper presents a design of the retrieval system that utilizes this concept and the related algorithms,and introduces hierarchical retrieval that could effectively interact with the user with regards to the structural characteristics of the video data.According experiments have been conducted with the context of TRECVID and VideOlympics video retrieval evaluations,and indicate promising results.