Research On Optical Image Registration Techniques From Multiple View Angles

With the constant development of sensors and algorithms,there are more and more different kinds of image.And the application fields of images have also become very widespread,mainly used in virtual reality and augmented reality,medical pathological analysis,satellite remote sensing and other aspects.The technology of image splicing,image fusion,and target recognition develops rapidly.As the basic work of many image application,the performance of image registration directly affects the subsequent work.Therefore,the research of image registration technology has important significance.This thesis performs registration algorithm research based on the actual application of optical images which obtained from different angles of view.The main work of the thesis is as follows:(1)Harris corner algorithm,SIFT(Scale-Invariant Feature Transform)algorithm and SURF(Speeded Up Robust Features)algorithm are implemented respectively.The performance of these three algorithms is compared in detail and the differences in their performance are analyzed.(2)For visible multispectral satellite imagery with many points,the thesis uses the Harris corner algorithm combined with the SIFT operator to match.Because of mismatching problem,TAR(Triangle Area Representation)which has affine invariance is used to remove false matching points.Compared with RASAC(Random Sample Consensus),it is proved that the algorithm presented in this thesis has better performance in accuracy.(3)Targets in ordinary optical images usually have obvious line features,so this thesis studies a straight line segment matching method.Firstly,extract line features of images using LSD(Line Segment Detector)algorithm.Then,combine the visual saliency algorithm to eliminate line segments whose length is less than the threshold.Finally,use GMM(Gaussian Mixture Model)for parameter optimization,at the same time finish the work of registration.(4)For ordinary optical images with a certain viewing angle difference,in order to make up for the shortcomings of traditional point registration,this thesis uses point-line combination method to register.Firstly,the transformation model parameters obtained by line registration are used to transform the registration image.Then,Extract Harris point features on the reference image and the transformed image to be registered.And complete the registration.The algorithm of this thesis has better registration accuracy than the traditional point registration algorithm.