A Fast and Accurate Image-Registration Algorithm Using Prior Knowledge

We propose an iterative image-registration algorithm to estimate the homography between two plane surfaces based on prior information. The difference to established methods using algorithms like SIFT and SURF lies in the kind of features used for the calculation. Instead of patch-based features, the algorithm uses prior information to search for features along 1-D signals and their expected counterparts. This reduces the computational complexity by one dimension and decreases the search space for feature correspondences enormously. The reduction in dimension and the comparison of whole image strips yields better accuracy and lower computational effort, while being able to handle any perspective transformation. Based on a broad variety of test images, an estimation of sufficient prior knowledge and a comparison with the established methods SIFT, SURF and ORB is given. We show that the iterative nature and the strip-based approach makes the algorithm very flexible, thereby achieving high accuracy at low computational effort.