Feature-based Image Registration Algorithm

Image registration is a classic problem in image processing, and it is extensively applied to the field of computer vision, pattern recognition, medical image processing and remote sensing information processing. Basically, the goal of image registration is to align images of same scene by removal of the potential geometrical distortion existed.In this paper, we propose a new hybrid registration algorithm combined with high level features (salient curve) and low level features (critical points), based on the analysis of related works. The hybrid algorithm improves the robustness effectively by using both the stableness of salient curves in image and accuracy of critical points. The curves that are stable under various distortions are extracted in Graph-cut Theory, which is fully using gray level, texture and spatial information of image. Through construction of a series of local geometry descriptors and statistical analysis of their distributions, we avoid the difficulties of direct point matching problem. When the coarse registration is finished, we further adopt Gabor filters to locate several critical points and establish the correspondence by fast normalized correlation coefficients. Then we solve the transform parameters in the means of least square error method. In order to reduce the computation, we build Gaussian pyramid structure. Curve matching is only applied in the top level while the point matching is used in the rest level so that we can solve the transform parameters in a coarse-to-fine procedure. The experiment shows that the algorithm we proposed is much effective than the usual correlation-based algorithm, even the images have complex scene and the distortion is great.