Novel image registration method using multiple Gaussian mixture models

Traditional feature based image registration methods work like point set registration with treating feature points from a whole image as one feature point set. However, unlike point set registration problem where only one meaningful structure is present, remote sensing images are usually present with lots of details. This paper uses the spatial information of feature points to divide them into several meaningful interest areas. In this way it is more appropriate to use point set registration methods. The image registration method presented in this paper takes advantage of the robustness and efficiency of Gaussian Mixture Models (GMM) based point set registration methods and makes several improvements considering the particularities of image registration. Experiments performed on real remote sensing images proved our method´s superiority over traditional direct implementation.