Research On Edge-based Remote Sensing Image Retrieval Technology

In the application of content-based remote sensing image retrieval, how to retrieve images effectively through shape features has become an urgent problem. Edge is a carrier of spatial information. The methods to capture spatial information, contained in remote sensing images, from edge features for similarity matching are quite direct. This paper mainly researches some techniques in the field of edge-based remote sensing image retrieval. Aiming at the problems that are possibly encountered in the process of global-feature based image retrieval, some solutions are proposed, whose performances are theoretically analyzed and practically validated. The main work and innovations are as following:1. The generalized Fisher information measure is presented. The multi-scale response of shape features and texture features is analyzed systematically through the above tool. It is used as the theoretical basis of multi-scale representation of edge features.2. Two groups of retrieval methods, which are based on local structural pattern of edge maps, are improved. They are multi-scale edge map primitive moment and multi-scale edge map local binary pattern. The experiments confirm that the edge-map-based retrieval mode has such excellent characteristics as small time cost and low feature vector dimensions.3. The concept of directional field is introduced. Aiming at the insufficiency of edge angular histogram to describe spatial information, it is improved by multi-scale analysis technique. Processed by Fourier operator, feature vectors become rotation-invariant, whose dimensions are reduced at the same time. The improved histogram enhances the retrieval quality for object images remarkably.4. In the spatial domain, considering that the dimension of angle histogram is too large to be applied for similarity computation, some retrieval methods, which are based on parameters property of transform domain, are put forward. The theory of Contourlet transform is studied emphatically. A statistical analysis of the retrieval problem is expounded, and the modified energy characteristic is proposed. After comparing Contourlet transform with wavelet transform, it is found that the wavelet transform combining with generalized Gaussian model is fit for natural texture image retrieval, while the improved energy features of Contourlet's subbands are more suitable for object image retrieval.5. Partition and hierarchical structure schemes for image databases are proposed. Balancing the overlapping and storage index, images are partitioned to blocks by the Quin-tree method. In the application of image resolution transformation, an interpolation model, which is based on Gaussian function, is proposed. The experiments done on single remote sensing images confirm that the proposed schemes can locate objects or regions precisely and, at the same time, response flexibly to scale variations.6. Combining the features based on edge map with edge directions, the performance of retrieval is analyzed. It is indicated that, through assigning reasonable weights, the retrieval performance for object images can be greatly enhanced.