Registration of image feature points using differential evolution

Abstract:

This paper introduces a robust global nonlinear optimizer—differential evolution (DE), which is a simple evolution algorithm to search for an optimal transformation that makes the best alignment of two sets of feature points. To map the problem of matching into the framework of DE, the objective function is proportional to the registration error which is measured by Hausdorff distance, while the parameters of transformation are encoded in floating-point as the functional variables. Three termination criteria are proposed for DE. A simulation of 2-dimensional point sets and a similarity transformation are presented to compare the robustness and convergence properties of DE with genetic algorthms (GA). And the registration of an object and its contour model have been demonstrated by using of DE to natural images.