Research And Application Of Image Classification Based On Transductive Support Vector Machines

With the rapid development of Web technology and popularity of using the Internet, search engines become an important way for many people to get information. Multimedia, especially image and video information become an important content of information retrieval, and image retrieval is more fundamental. In image retrieval technology, how to use content-based image retrieval is the key of making the difference between retrieval results and user's semantics as smaller as possible.In this thesis, we firstly review some kinds of retrieval algorithms, in which SVM is considered to be an efficient one. Basing on it, Transductive Support Vector Machines are more efficient than inductive SVMs, especially for very small training sets with large test sets. But they have disadvantages, such as high time complexity and the requirement of "num+".We promote a new algorithm in this thesis, which bases on TSVM. But we add some filter in it and improve the flow, which reduces the time complexity with little influence on the performance.Further more ,an experimental system was designed and implemented in this thesis, which used the improved algorithm, and the progressive character. It based on multi-layer semantic which contained semantic class, image class, self-define class, and had two search models which were sample based and semantic key words based.