Automated Feature-Based Registration Techniques for Satellite Imagery

We investigate and validate feature-based registration techniques for remotely sensed satellite images. Feature-based registration algorithms seek to detect image features such as boundaries, corners, segment intersections which are used for matching. We implemented some of the state-of-the-art feature detection, extraction and matching techniques, which are BRISK, FAST, HARRIS, Minimum eigenvalues, and SURF algorithms to register 19 satellite image pairs in our data set. We generated ground truth data by manually selecting control points in each image pair in our data set that we also registered using automated feature-based techniques. Finally, we computed the pixel discrepancy error of each method against the ground truth and compared the performance of each method. Our results show that feature-based methods may achieve subpixel accuracy for the specific image data set.