Research On Content-based Remote Sensing Image Retrieval Technology

With the rapid growth of remote sensing image data amount, the scale of remote sensing image database increases rapidly. How to quickly browse and efficiently retrieve the target images from the database of large-scale remote sensing images become the bottleneck in remote sensing image information retrieval and sharing. It has very important theoretical and practical value to use the content-based image retrieval technology in the remote sensing image retrieval field. In view of this, this thesis researches the technique of content-based remote sensing image retrieval. The content of the study will be shown as follow:1) By analyzing the characteristics of remote sensing Image Retrieval, the feature extraction and similarity measurement technology are studied through a lot of experiments. A new retrieval method which integrates the dominant color feature and segment color moments is proposed, and it can enhance the retrieval efficiency. After analyzing the shortage of the texture feature based on classic co-occurrence matrix, the improved gray-smooth co-occurrence matrix is proposed to use in the analysis of remote sensing images. Furthermore, a group of Gabor filters are used in texture-based remote sensing image retrieval. It is proved that the method has a good applicability. This thesis also analyzes the necessity of the shape feature extraction in remote sensing image retrieval.And the extraction method of the moment invariant is presented. The experiments of two groups indicate that the method is efficient in remote sensing image retrieval.2) The techniques of feature fusion, including similarity measurement and normalization of features and so on, are studied. And these techniques are applied to remote sensing image retrieval.The experiments indicate that the technique of combining kinds of features in image retrieval can achieve good effects.3) Relevance feedback techniques which are related to the image retrieval are studied. A relevance feedback algorithm based on weight adjustment in the vector space model, which is aimed to the features-integrated search is proposed. After experimental analysis, it is proved that this method can improve the system performance with high feedback efficiency and accuracy.4) Based on the above, this thesis constructs a prototype remote sensing image retrieval system, which takes advantage of many research results in this thesis and has better scalability.