A fully automatic method for feature-based image registration

Most applications in computer vision rely heavily on a common task: the registration of a reference image with a newly sensed one. We propose a feature-based method for reliably registering two images. This method uses points of interest as feature space, rigid transformations as search space and as a novelty, a search strategy based on extremal optimization. Feature extraction and feature matching are the main crucial steps for feature-based image registration and so far no general method exists to solve both of them. Due to the results of the feature extraction step, feature matching may be computationally intensive. The search strategy we propose is able to reduce the complexity of feature matching and to withstand the effect of outliers. It integrates the search in the transformation space with the search in the feature correspondence space and progressively eliminates bad matches which results in robustly estimating the registering transformation parameters. Experimental results on a wide range of image pairs are very encouraging and demonstrate the effectiveness of the proposed method.