Image Registration Research Based On Improved SURF And Delaunay Triangulation Combination

Image registration is one of the key technologies in the field of image processing, which used to match and superimpose two or more images taken, for example, at different times, from different viewpoints or from different sensors, its purpose is to eliminate or reduce the geometric distortion between images. Currently,image registration technology has been widely used and researched in the field of information fusion, image stitching, change detection, object recognition, stereo vision, environment surveillance, map update, weather forecast etc.The core problem of image registration is to improve the registration speed, accuracy and the robustness of algorithms. Therefore there are still many more challenges we have been facing when everyone studies a great registration algorithm that including better robustness, high precision, stable performance and strong adaptability.SURF algorithm is a new method which has been used to extract many feature points of images in recent years. Then it has a lot of advantages on the aspects of robustness, repeatability, uniqueness,etc., which compared with the conventional proposed method. Furthermore, this algorithms introduces integral image and box filter to improve computation efficiency. So the SURF algorithm has been an absolute advantage method.From the registration efficiency and accuracy, this paper proposes a method which based on improved SURF and Delaunay triangulation combination. This method, which based on the original SURF algorithm, introduces color invariant model and constraint conditions which including Delaunay triangulation, triangle similarity function and photography invariant. The specific implementation process are as follows. First, use color invariant model to process both remote sensing images as the input images. Second, use SURF algorithm to extract feature points of the input images and apply plane Delaunay triangulation on them, so it can obtain two images(including reference image and image to be registered) of the Delaunay triangulation. Then use the triangle similarity function to calculate the similarity between both images and select pairs of triangles which the similarity greater than 0.75 as many candidate triangles. Furthermore, use the photography invariant to process these candidate triangles in order to eliminate error registration feature points and extract precise feature points. Finally, the spatial transformation model is used to transform these precise feature points so as to realize precise registration process of remote sensing and achieve the final result of registration.In this article, based on Matlab2012 a platform finished the experimental results show that this method not only effectively retains the images’ color information and reduce error feature rate of registration, but it has high speed, many more feature points which is well-distributed and high registration rate, in order that this method has been proved to be reliable and effective.