Hough-domain image registration by metaheuristics

Image registration is the process of registering two or more images, which may be acquired under different imaging conditions. The critical issues in image registration are robustness and speed of the algorithm, which most current algorithms are devoted to. In this paper, a robust and efficient algorithm is presented for registering images in Hough space using heuristic approaches. By Hough transform, the main structure of an image can be represented in the position-orientation space. This representation has almost all structural information of the original images, especially for images rich with line segments. This representation allows us to register images efficiently in Hough space rather than in the original image space. To account for differences between the images to be registered, the generalized partial Hausdorff distance is proposed and used to measure the image similarity. In the presented algorithm, the rotation parameter is computed simply by 1-D correlation, and other transformation parameters are determined by a new hypothesis-test method, where each hypothesis is generated by a heuristic approach, i.e. random local search. The proposed algorithm has very low computational complexity, and works well for most natural images rich with line segments resulting from man-made structures.