Feature-based image registration using the shape context

Abstract

Image registration is a basic and important process for multi-sensor or multi-temporal remote sensing. In this article, a new feature-based method named shape context is proposed for airborne multi-sensor image matching. This method has been found to be robust in hand-written digit and object recognition, and it is now introduced into remote-sensing image matching after some adjustments. In the proposed method, control points (CPs) are extracted on the reference image, and edge features are extracted on the reference and the sensed image, respectively. The shape context exploits feature similarity between circular regions of the two images to find corresponding CPs on the sensed image. Finally, the sensed image is warped according to the CPs using thin-plate spline interpolation. This method is successfully applied to register airborne optical and multi-band synthetic aperture radar (SAR) images in two experiments, and the results demonstrate its robustness and accuracy.