Point Feature-based Image Registration Algorithm

Image Registration is a technique that relates or aligns different images taken from different viewpoints, at different times, or by different image sensing devices. Image mosaic technology uses the result of image registration to build large view panorama-like image from small adjacent image series, which could enlarge the user's view scope and increase image resolution as well.Till now, image registration and mosaic technology is widely used in the field of virtual reality(VR),video compression, image super-resolution, intelligent surveillance system, etc, which came to be an active research area in computer vision during recent years.Feature-based, direct pixel difference optimization based, and Fourier based method are three typical ways of image registration, which have its own appropriate application area separately. The key problems focus on how to increase the algorithm speed, to increase the registration precision and to enhance the robustness of these registration methods.This paper did research on overlayed-image registration and mosaicing technique and presented some new effective algorithms based on point-matching.Firstly, the author did some research on frequency based and corner-matching based image registration technology used in short baseline correspondence condition. Then a new fast and robust image auto registration algorithm is presented. In this method, there is a new twice guided-matching scheme that first use phase correlation to estimate translational parameters to roughly guide local corner matching, and second use an improved RANSAC algorithm to estimate 8 parameters of perspective transformation to guide local corner matching again, with more precision. So much more valid corner matches could be quickly and stably generated. SVDLS (Singular Value Decomposition Least Square) is used to estimate transform parameters and Levenberg-Marquardt optimization is used to decrease registration error. The thought of guided matching mechanics is experimented to show better results than existing methods on matching speed, registration error, and the whole robustness of the algorithm.Secondly, a simple, practical image auto stitching method based on brightness and white-balance automatic tuning is presented. Both of the geometric and photometric...