Study On Feature-Based Multi-Source Remote Sensing Image Registration Techniques

The main purpose of remote sensing image registration is to realize the process of geometrically overlaying two or more images of the same scene taken at different times, from different viewpoints or by different sensors. Remote sensing image registration is a crucial step in many image analysis tasks which include change detection, information fusion, environment surveillance, image mosaic, object recognition, weather forecast, map update and so on. On the basis of comprehensively summarizing and analyzing the current image registration techniques, this dissertation emphasizes the feature-based multi-source remote sensing image registration.In this dissertation, the effect of feature spatial relation, feature attribute and feature uncertainty on image registration is firstly analyzed across-the-board and depicted quantificationally, and then some feature-based remote sensing image registration methods are developed. Finally, the proposed methods are vertified by a lot of real remote sensing image experiments. The main achievements and contributions of the dissertation are listed as following:(1)A mathematical model which includes two key elements, namely the similarity measure and the optimal solution, is developed to unify the remote sensing image registration methods based on spatial relations in theory. The implementations of various existing classical methods based on spatial relations in the mathematical model are anatomized in detail. To overcome their shortcomings, this dissertation defines a new similarity measure. It has a more improvement than similarity measures of the existing classical methods in smoothness, continuity and prominence of its optimal solution. Moreover, the exclusive parameter of the proposed similarity measure has little influence on the prominence of the optimal solution and so the method is robust. A two step iterative local optimal method is devised to solve the optimization problem of the new similarity measure.(2)In the remote sensing image registration methods based on feature similarity, two new methods are proposed by the dissertation. To overcome the disadvantages of the conventional methods based on feature similarity and ones improved by Flusser, a new image registration method based on invariant descriptors is developed. In the new algorithm, the criterion of the nearest distant of feature invariant descriptors and row and column matching likelihood coefficients are used synchronously. Therefore, the new algorithm is suitable to register two images with some very similar shapes. In general, most of existing point-based image registration methods based on feature similarity is sensitive to image intensity. Aiming at the problem, another new method based on virtual triangle is proposed to solve image registration with rigid geometric distortion. The virtual triangle, which is composed of any three corners, is used as registration element and all matched corners are obtained according to the criterion that virtual triangles are congruent.(3)Feature spatial relations consistency and feature attribute similarity are two main rules to solve feature matching problem of image registration. The existing methods generally only use one of them or combine them sequentially, and all have some limitations.It must be stressed that spatial relations and feature similarity are two indispensable aspects of reliable image registration, and both are of equivalent importance. Therefore, a new method is developed. It is realized by introducing a function whose independent variable is the match matrix, which describes the correspondence of the features, to combine spatial relations and feature similarity organically and its global maximum is assumed to be reached if the sensed image is completely aligned with the reference image. Thus, the image registration problem can be transformed into the optimal one. Two approaches are devised to solve the optimization problem. One is based on the branch-and-bound strategy and the other is the two step iterative algorithm that combines graduated assignment and variable metric methods.(4)Most of the existing image registration methods hardly study feature uncertainty problem. According to experience on feature uncertainty analysis in other fields, feature uncertainty in image registration is studied deep in the dissertation and two main achievements are obtained: firstly, the uncertainty of corner and straight line pair, which is composed by any two straight lines, is described quantificationally from the views of statistic and shape analysis; secondly, feature uncertainty is considered in the process of feature selection and matching. Comparing to the methods not considering feature uncertainty, the proposed method has an improvement in computation complexity, robustness and registration accuracy.