Corner Feature-based Image Registration Technology Research

Digital image registration technology matches two or more images of certain scene from different viewpoints or sensors, at different times, or under diverse conditions. The quality of image registration would affect the following processing steps directly.There are three main methods in this field: transform-domain-based, feature-based and gray-based. Although they have their own advantages, none of them can deal with all the problems. In practice, a proper method is chosen at the balance of speed, accuracy and robustness. The dissertation describes transform-domain-based and feature-based methods deeply, and a novel image registration algorithm is proposed.Transform-domain-based method often exists in image registration, for it is easy to compute and has a certain degree of robust against noise. Firstly, we discuss the estimation of rotation and translation with Fourier phase correlation method. Secondly, polar coordinate method is used to estimate the rotation which often failed to be computed in Cartesian coordinate. But the above method requires that the image must rotate around the image centre, which seriously narrows the spectrum of its utilization.The dissertation describes the concepts, theories and classic algorithms of feature-based registration methods. Then, a robust registration method is proposed. Multi-scale Harris corner detector is applied in feature extraction which leads to stable and accurate corners; a voting-strategy-based algorithm is employed to reduce the unstable and mismatched points, in order to obtain a coarse feature set; RANSAC is used to estimate the transform matrix. The experimental results show that the method has high accuracy with lower computation load.The most important use of image registration technology is image stitching. In chapter 5, we discuss the stitching techniques, such as image interpolations and gap-reducing. At last, a conclusion is made and future work is illustrated.