The Research Of Image Registration Technology Based On Feature Points

Image registration is one of the key technologies in image processing. This dissertation addresses the problem of image registration based on feature points in images. Detection of feature points, Correspondence of feature points, the image registration method based on feature points and applications of image registration are systematically studied in this thesis.1,Detection of feature points and performance analysisThe detection algorithm for feature points in an image and its performance is studied. Based on wavelet transform, a rotation and scale invariant detection algorithm is presented. This method uses the multiscale properties of wavelet transform to solve the problem of scale invariant in feature points detection and employs the same way as Harris's method to obtain the rotation invariant. The experiment results demonstrate that our method has better scale invariant characteristic than SIFT method. A detailed performance comparison of different feature point detectors is given. The comparison is carried out on the same band of images and different bands of images.2,Correspondence of feature points and performance analysisCorrespondence of feature points involves two aspects: point similarity measures and search strategy of correspondence.Point similarity measures can measure similarity of individual image points. This thesis suggests two kinds of point similarity measures which are respectively based on Fourier-Mellin transform and trace transform of image. These two methods can deal with complex and unknown image transform. Remarkably, the point similarity measure based on Trace transform is very fast and robust and its efficiency is as high as the NCC method.Based on the similarity of feature points, search strategy of correspondence selects efficient correspondences of feature points and excludes false correspondences according some rule. Correspondence of feature points is an important step in image registration based on feature points; also a detailed performance analysis is given.Local entropy and local standard deviation of images are efficient feature transforms for different band images. This thesis improves on the local standard deviation and presents a method so called logarithmic local standard deviation. The correspondence of feature points for different band images in feature transform domain is given.3,Image registration based on feature pointsThis thesis analyzes image transform model, reviews the direct linear transform for image projective transform parameter computing and RANSAC for robust estimation of image transform model. We also present the flow chart of image registration based on feature points and the results of simulation experiments. For image registration with similar transform, a fast and robust image registration method based on local feature points group is proposed.4,Application of image registration based on feature pointsThere are many applications of image registration. We analyze the difficulties and characteristics of some typical applications, such as wide-baseline image registration, registration for different bands images, aerophotographical mosaic, scene matching and panorama for multiple images. The performance of our image registration technology is presented.