Feature-based Image Registration And Image Fusion Algorithm

Image registration and image fusion play important roles on image processingtechnologies, the former is computing the parameters of relative displacement androtation between images, the latter is the fusion of the two images based on theregistration parameters, so that a better understanding can be gotten.Generally, Image registration algorithm can be classified into three,Grayscale-based, Transform-based as well as feature-based registration algorithm,because the last has high accuracy, adaptability, eliminating light impact, it is appliedwidely. In this paper, we use SIFT algorithm on image registration, which has a verypowerful performance for general object detection/recognition. However, how toachieve an ideal matching result is the most importance that we study in our work.The original SIFT algorithm is famous for its abundant feature points, but the finalkeypoints are so excessive that the matching speed is very slow at the next step ofsearching for homonymy point-pairs. We analyze the performance of SIFT andconquer its deficiencies applying exhaustive searching and Kd-tree searchingarithmetic, after that eliminate mismatching points based on RANSAC arithmetic, atlast, computing the registration parameters with the Least Squares Method.Experiments on real-world scenes demonstrate that the method can reach a betteraccuracy for computing parameters of translation and rotation, which outperformspreviously proposed schemes with regard to precision and robustness. Then the newmatching strategy and localization strategy are investigated to ensure the stability andprecision of localization. Compared with conventional localization algorithm, thisalgorithm makes the precision more stable, which reaches0.01pixel. Finally, fusion the two images based on the registration parameters, the result is perfect as well.