An automatic filtering algorithm for SURF-based registration of remote sensing images

The registration of remote sensing images has been often a necessary step for further analyses of images taken at different times, different viewing geometry or with different sensors. For this task there exists many approaches. This paper focuses on the feature-based category of image registration methods. Particularly, we propose an improvement of the SURF algorithm on the point matching step. Indeed, in order to achieve a correct registration, a good matching of feature point is required. However The presence of outliers lead to a fail in the registration. Therefore, in this paper, we introduce an efficient method devoted to the detection and removal of such outliers. It's based on an automatic filtering of outliers based on both distance and orientation between feature points. Images from IKONOS and QuickBird satellites are used to evaluate this proposed method, which we compare to classical SURF as well as SURF followed by RANSAC filtering. The results show that our method outperforms the others regarding all assessment criteria.