Log - Polar Image Registration Techniques In The Field

Digital image registration is a branch of computer vision that deals with the geometric alignment of a set of images. In recent years, with the rapid development of image registration technique, it has been widely applied in many areas, such as remote sensing, medical imaging, computer graphics and computer vision. In particular, image registration algorithms for applications such as stereopsis, motion object detection, motion estimation, mosaicing and vehicle navigation have received considerable attention. Despite comprehensive research spanning over thirty years, most registration techniques will fail in the presence of large deformations. So the robust technique to register images concerning large deformations is a big problem that requires urgent solution.This primary objective of this thesis is to investigate robust image registration algorithms under similarity or affine model. First this thesis introduces the use of Log-Polar techniques to invert the deformation among image pairs, and then it achieves sub-pixel registration accuracy through the use of Log-Polar mappings and least square optimizations. Second, for the problem that most similarity metrics used in image registration will fail in the presence of large image outliers, this thesis propose a new similarity: robust correlation coefficient, we use it to reduce the outliers' influence in the application of image registration of Log-Polar domain.This thesis is organized as follows. In Chapter One, we will review all the angles of image registration problem in the space domain, including image feature space, image search space and image similarity metrics used in the image registration problem. Then concerning that most algorithms are not very useful in the presence of large deformation image registration problem, we draw in the Log-Polar transform and introduce its characteristics. In Chapter Two, we apply the Log-Polar transform and investigate the global image registration algorithm under similarity model. We compare this algorithm with the famous Fourier-Meillin method in the aspect of registration accuracy. In Chapter Three, we introduce the concept of radical line correspondence according to affine transform characteristics, and then we propose the image registration algorithm in the Log-Polar domain under affine models. In Chapter Four, we bring up the new maximum likelihood estimation method for the correlation coefficient and propose the robust correlation coefficient metric to counter the problem that most common similarity metrics can not handle the outliers in images. Finally we apply our new metrics in the Log-Polar domain image registration problem. The last chapter -Chapter Five is this thesis's overall conclusion and prospects for future image registration algorithm in log-polar domain.