Automatic image registration by local descriptors in remote sensing

Abstract

Registration of multiresolution and multisensor images is a challenging problem in the research area of remote sensing. Conventional image registration algorithms have greatly suffered from control points selected manually. In this study, we propose a novel method for automatic affine image registration based on local descriptors, called automatic image registration by local descriptors (AIRLD). In this algorithm, we first apply the Harris-affine method to extract the interesting points of the given images, and then they are used to calculate the affine invariant feature descriptors. These descriptors are combined through the normalized mutual information (NMI) method to obtain robust matching control points in both reference and sensed images automatically. Finally, we show that subpixel registration accuracy can be reached by means of a least-square fitting of a paraboloid function to the surface generated by the NMIs of the control point neighbors. Experimental results show that our method can provide better accuracy than the conventional registration process.