Analysis Of Laws In Representing Semantics And Discovery Of Topics For Web Images

With the great development of Internet and Web-based multimedia sharing communities especially Flickr and YouTube, the volume of multimedia content is explosively increasing. Therefore, the need for efficient organizing and retrieval of multimedia resources has become more and more important. The basis for these applications is reasonable presentation of images' semantics.This paper aims at analyzing and studying on the statistical properties and topic discovering strategy for web community images, which is based on the research of semantically representation and way of organizing for web images, and targeting at the application of image annotation. Thus, we propose the theoretical and application framework of precise semantic representation and topic discovering for web-community images. Besides, several groups of experiments on large scale data set is designed to validate the correctness and effectiveness of our framework.This paper firstly focuses on the statistical properties of annotation and representation for web multimedia community. Compared with traditional research fields of natural language processing and information retrieval, three assumptions are obtained: Sparse distribution assumption, locally convergence assumption and global convergent conjecture. These assumptions provide basis for further studies on the selectiveness of topics and keywords which produces the complete set framework and bipartite LSA algorithm. The complete set framework qualitatively analyzes and evaluates the keywords and topic selection procedures via heterogeneous data analysis and minimum error energy criterion. Thirdly, we propose the visual topic model framework based on our previous theoretical research– the three assumptions and complete set framework. The visual topic model is based on unsupervised learning approaches and maximum visual diversification criterion, aims at discovering the visual topic distribution under large scale data, and is applied to web-based image annotation. Experiment results show credible evidence for our theory and framework. At the end of this paper, we conclude our work and make future research prediction.The whole research bases on the statistical work on large scale data and theory analysis and proposes a complete framework for web image semantic representation and topic discovering, which is believed to be valuable for further study.