Image Registration Using Structural Corners

Image registration is an important task for many applications of image analysis and computer vision. Robustness and efficiency are two critical issues to be addressed in registering images. In this paper, a commonly used framework is adopted, where local features are first detected and then image registration is carried out by establishing the feature correspondences between two images. To ensure the robustness and repeatability, a novel approach is proposed to detect structural corners rather than the textural ones, because the structural features of images cannot be represented by textural corners. To establish the feature correspondences robustly and efficiently, modified Random Sample Consensus approach(RANSAC) is presented, where the information provided by the structural corners is utilized. Finally, the solution is obtained using the voting method. Experimental results show the effectiveness of the algorithm, especially when the images have been taken at different times and even by different sensors.