Semi-Automatic Clustering for a Registration of Radar Images

摘要

Image registration is one of the important and challenging work in remote sensing. But there are some factors making image matching harder such as occlusion, clouds, terrain deformation and relief displacement. For image registration, image matching is a important step. In order to match the images, it has to extract the interesting points and find the most possible conjugate point pairs in the images. SIFT (Scale Invariant Feature Transform) is demonstrated as a better algorithm compared with the other approaches in past literatures. It is the most popular scale invariant feature descriptor because of its resistance to image deformations. But there are many blunders in the conjugate point pairs matched by SIFT, they must be removed. RANSAC (RANdom SAmple Consensus) is selected to get reliable pairs. This presentation proposed a useful and semi-automatic method which is different from the global strategy. In some cases, registration between images does not use just one set of mapping parameters. So we have to find out different mapping regions of an image pair. The first step is to split up the whole image to small parts which should also have the exact corresponding relationship. So we triangulate with the conjugate pairs and then cluster triangles by £-means. Different clusters are assigned to different parameters. This paper uses the method of clustering registration to RADAR satellite images, and compares it with global affine, and also compared with result without clustering. In the end, it shows good results which have complex mapping relationship.