Image registration based on Genetic Algorithm and weighted feature correspondences

Super resolution is a technique of enhancing image resolution by combining information from multiple images. It is widely applied in fields like camera surveillance, satellite imaging, pattern recognition, etc. One challenging problem of super resolution is its high demand on image registration accuracy. This paper introduces a high accuracy registration approach for the purpose of super resolution. It is invariant to translation, scaling, rotation, and noise, and can be used to automatically obtain the maximize a likelihood estimation (MLE) of image homography (registration result) using information only contained within the images themselves. An effective genetic algorithm based approach is used to filter out all the mismatches. Comparison with RANSAC and Keren's method will be given to prove the effectiveness of the proposed method.