Research On Full - Map Registration Method Of Wide Baseline Image Based On One - Dimensional Matching

Image matching is one of the cores of image processing. The method to achieve image matching is to estimate the spatial transformation from the target image to the reference image of one same scene. With the transformation model, the target image can be projected onto the coordinate system of the reference image in order to match the reference image accurately.This thesis based on efficient one-dimensional matching, studies the two-dimensional global matching to achieve the goal of wide baseline image matching. First, this thesis studies and proposes the principles, methods and techniques of the two-dimensional global matching which expands from local blobs matching to the global total overlapping registration, to achieve the goal of wide baseline image. This thesis by studying the one-dimensional matching technology proposes and verifies three feature extraction algorithms which include Edge point features, corner point features and Bubble point features.First, this thesis describes the conversion of image from Cartesian coordinates to log-polar coordinates, and then extends to the conversion of edge point feature image, corner point feature image and bubble point feature image. After transforming the feature image blocks into log-polar coordinates, the first-order differential method can be used to obtain edge point gradient image, corner point gradient image and bubble point gradient image respectively.On this basis, this thesis presents the mathematical forms of feature extraction algorithms which include edge point features, corner point features and bubble point features. According to the characteristics of the three feature extraction algorithms, this thesis validates the direction and gradient mapping and the responses of it respectively, to judge whether a single algorithm has the capabilities of feature extraction. Based on the characteristics of these three feature extraction algorithms and the responses of the three ideal feature gradient image groups, this thesis verifies whether these three feature extraction algorithms has obvious degree of distinction with each other. After this, image set of the data group "Rensselaer" was given to extract the three features. Based on this, the NCC algorithm is used to get different types of feature points of the matching pairs. This thesis also describes the generalized dual bootstrap iterative closest point algorithm(GDB-ICP), with this algorithm and NCC algorithm to obtain and increase the bootstrap region, and finally to achieve the two-dimensional image’s global registration.