Research On Method Of Retrieval From Remote Sensing Image With Global And Local Features

With the rapid development of remote sensing technology, the remote sensing image data provides a important way for people to getting information. And it is playing a more and more important role in military intelligence, national security, resources detection, city planning, planet detection and such kind of fields. Content-Based Image Retrieval is a important solution of information getting and data sharing from tremendous remote sensing data, and now it has become a pressing issue.Based on lunar data, this thesis focuses on the method of content-based remote sensing image data retrieval. With analyzing and absorbing relative research results, this thesis undertakes a more thorough study to this problem, including features of remote sensing image, retrieval algorithms, feature matching, etc. The main works are as follows:1. Based on analysis of remote sensing image features, this thesis presents a new idea about solution to remote sensing image retrieval, it express a concise idea of the differences between remote sensing image and that of other fields. As a conclusion the existing retrieval methods can not satisfy the real needs. Finally, this thesis presents a new method of image retrieval with both global and local features.2. Based on the analysis of remote sensing image, research on the global feature description method of remote sensing image. Firstly, compared with color and texture description method, we find that the area-based shape description method is better for remote sensing image retrieval. Secondly, propose a standard of ideal feature descriptor. With this standard, the thesis researches on the feature description method based on Hu moments and its distance measurement.3. Research on a scale invariant Feature description. The scales diversification of remote sensing image is decided by the diversification of collections, sensor and missions, it is an important problem in content-based remote sensing image retrieval. SIFT feature is invariant to image scaling, this thesis apply it in to remote sensing image retrieval, and experiments have showed that this method can effectively improve the retrieval accuracy.4. With the idea above, this thesis propose a new method of remote sensing image retrieval with both global feature and local feature. The method firstly search the image database with Hu moment feature descriptor, and then use SIFT to search in a smaller range. Experiments have showed that this method is more accurate and faster than existing methods.Supporting for the research above, this thesis realizes a prototype system for content-based image retrieval. And the testing results show that the research of this thesis is efficiently.