Automated selection of corresponding point candidates for image registration with artificial neuralnet using rotation invariant moments as the input data

Abstract

Accuracy of image registration is severely affected by that of corresponding control point (CCP) selection in remote sensing or GIS (Geographic Information System) . In this paper, a new automated system for CCP candidate selection from target images is proposed. In the system, first, grayscale image within a quasicircular field of view (FOV) is transformed into binary one after intensity modification for the several extreme intensity pixels. Next, the binary image is transformed into a rotation invariant intermediate representation. Finally, the system determines whether the central pixel of the FOV is appropriate as the CCP by using well-trained 3-layer feedforward artificial neuralnet. Pseudo Zernike moments are employed as the intermediate representation. Consequently, without selection accuracy deterioration, we achieve quite fewer training patterns, shorter training time, and higher noise tolerance in comparison with conventional neuralnet-based systems.