A learning-based approach for rigid image registration accuracy estimation

This work presents the machine learning approach to estimate the accuracy of pixel (or intensity) based image registration algorithms. The considered method is applied to the problem of image-based navigation but can be used in a variety of registration tasks. Training was conducted on the basis of twenty one features that extracted from the pair of registered images. For creating the train data a Monte Carlo simulation under four different models of subpixel shifts is considered. The RMSE (root mean square error) of registration estimates (along two axes) is used as target. The fragments of real terrain images were used in experiments. The bias-variance analysis of registration errors is conducted. Proposed method was compared with one theoretical (Cramer-Rao bound) and one simulation (bootstrap) approaches. Comparison showed the superiority of the proposed method to the existing state-of-the-art-methods.