Research On Several Techniques For The Retrieval Of Objects With Complicated Semantics

With the growing of media data, how to retrieve objects with complicated semantics has become a very important and challenging problem. In the literature, it has become an important trend to use machine learning techniques to improve the effect of the retrieval of objects with complicated semantics. This thesis focuses on this direction and makes several contributions summarized as follows:First, this thesis proposes a new bag generation strategy for text data named WPS and compares bag generation strategies for text and image data. Experimental results show that bag generation strategies have a significant effect on learning performance and the proposed strategy WPS achieved better performance in the text data.Second, this thesis proposes a new framework named TRUE. In this framework, the knowledge on the internet such as Wikipedia is used to help to understand high-level concepts, and then helps retrieve unannotated images with high-level concepts. Experimental results show that TRUE can improve the performance of retrieving images with high-level concepts significantly.Third, this thesis proposes an improved NRA algorithm for top k query problem. Experimental results show that the improved NRA algorithm is much better than simple top k query algorithm.