Robust Image Registration Based On Geometric Constraints

Image registration is the process of transforming two or more images of the same scene into one coordinate system. The images may be taken by different sensors, at different times, under different lighting conditions or from different viewpoints. It is basic work in all image processing and it is widely used in computer vision, medical imaging, military automatic target recognition and remote sensing, and so on.As a widely use registration method, spectral matching has a limitation on computational efficiency and scale invariance. It requires eigen-decomposing a large affinity matrix when the number of candidate correspondences is large. It lacks scale invariance because the pair-wise constraints cannot hold when large scale variation occurs.To solve the problem of high computational complexity, this thesis proposes a image registration method based on voting and summation-ranking. In the proposed method, each candidate correspondence has a function of voter, and it gives voting to the other candidates and votes itself too. The voting scores is then summed and ranked to get the maximum values which are the optimal correspondences of all candidates.To make the proposed method scale invariant, this thesis generalizes the pair-wise constraint matching method to tri-angle constraint, and then accomplishes image registration by a mechanism of voting and summation-ranking. The proposed make use of three candidates to form a novel triple-wisely geometric constraint which is robust to scale variation. This kind of constraint is measured by the similarity of the pair of triangles constituted by the three correspondences. The information of triple- wise constraints are encoded in a 3-dimensional matrix from which the optimal correspondence can be obtained by simple summation and ranking operations.Experimental results on real-data show the high effectiveness of the first proposed method and efficiency of the second proposed method for different-scale image registration.In this thesis, a tensor-based algorithm for high-order graph matching is discussed because it has much in common with the method we have proposed. This tensor power iteration method generalizes of the spectral matching method to higher-order potentials by taking in tensor theory so that it can be competent for both the same scale and different-scale image registration task. The tensor-based method is compared with our weighted voting method and the triple-wise constraint method. Experimental results on real-data show that the effect is almost the same, but our methods have much higher efficiency when computing the matrix.