影像套合是遙測領域中一門重要且充滿挑戰性的研究。由於影像的來源、時間不同，影像間的尺度、角度、輻射性質皆可能不同；又遮蔽、雲霧、地形變化、高差位移等的因素皆會導致套合成果不佳。 影像套合時需使用大量的共軛點對，因此特徵萃取以及影像匹配成了影像套合的重要課題。本研究使用尺度不變特徵轉換演算法進行影像匹配之研究，主要因素為本法在抗尺度與抗旋轉方面為最佳的描述元。但因其匹配成果仍有許多匹配錯誤之特徵點需要被移除，本研究使用隨機樣本一致性檢定與迭代資料搜評的粗差濾除方法來獲得更可靠的匹配點對。 本研究提出了一個有效且半自動的方法，與一般全域性的方法有所不同。在某些情況下，影像間的轉換參數並非單一，所以需要找出像對中擁有不同轉換參數的區域。為了要更佳的反應影像區域的複雜映射關係，首先需要將影像切割成若干對應之小區域，藉由尺度不變特徵轉換及粗差濾除後所得到的共軛點建立三角網。接著由仿射參數將網形中的三角形分群，不同群組對應到不同的轉換參數。這個區域性的策略將會對有複雜映射關係的影像有更適應的套合效果，本研究將其與其他的套合方式做比較。 Image registration is one of the important and challenging work in remote sensing. 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) and IDS (Iterated Data Snooping) are 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. In order to represent the image which has many different mapping regions, 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 k-means. Different clusters are assigned to different parameters. The local strategy is expected to more adapt to complex mapping. It is compared with global affine, and also compared with result without clustering. In the end, it shows good results which have complex mapping relationship.
