Image registration is a research focus in many fields such as computer vision, pattern recognition, medical image analysis and remote sensing image processing, etc. Starting with a brief overview of existing registration techniques for multi-source remote sensing data, we focus on the study of automatic registration of multi-sensor satellite images especially for SAR and optical data. The related work and main contributions are summarized as follows: 1. Image filtering. Based on an analytical review of several kinds of image filters, a new method is proposed for SAR de-speckling. The filter kernel is a modified version of Frost one, which combines the statistical information from the pixel and its neighbor pixels within the filter window. Therefore, the method not only avoids over-smoothing of SAR image structural details, but also amends the deficiency of keeping "false edges" by the enhanced Frost filter. So it makes a good trade off between speckle filtering and details preserving. 2. Feature Extraction. With the assumption that water bodies hold weak radiometric randomness in both SAR and optical remote sensing images, based on information theory, we propose a method for water body extraction. By means of empirical entropy, the gray level image is mapped into entropy image where dark areas correspond to water ones. Water areas can therefore be extracted by means of threshold technique and some post-processing. In order to get a more accurate border of water bodies, we also present a template-based scheme for border detection by combining the template and edge information. 3. Multi-sensor satellite image registration. Based on multi-layer feature matching, a coarse-to-fine registration procedure is proposed. In the stage of coarse registration, area features are depicted by shape matrix, which provides a unified measure rule not only for shape similarity but also the orientation consistency of this shape. Further the local orientation consistency, which is described by both Principle Component Analysis (PCA) and the shape matrix, effectively avoids the ambiguity of shape matching. In the stage of registration refinement, the parameterized description of edge features by NURBS curves properly solves the "irregular edge "problem because of speckle noise in SAR images. The utilization of local controllability of NURBS curve overcomes the local deformation, which is caused by the terrain or environmental change due to the different acquisition conditions of images. By combining some constraints on two corresponding edges, such as distance constraint, orientation constraint, etc. we obtain the matched control points of NURBS curves, which are invariant under affine transformation. The registration refinement can therefore be realized. 4. For the registration of large area images, an adaptive propagati
修改评论