Semantic classifier based on compressed sensing for image and video annotation

A new semantic classification approach for image and video annotation is proposed, which fits a semantic classification task into theory of a compressed sensing framework. The proposed approach first utilises training samples to create a dictionary matrix and then uses a matching pursuit algorithm to find the sparse vector. The final annotations are determined according to the reconstruction value from the positive samples and the sparse vector. A systematic performance study on TRECVID 2008 video dataset and Corel image dataset shows the proposed approach is more effective than the traditional support vector machine scheme.