A Machine Learning Approach for Image Registration Accuracy Estimation

This work continues our previous paper devoted to the application of machine learning to predict the accuracy of intensity-based image registration algorithms. Now we consider the estimation of a more complex geometric transformation involving translation, scaling and rotation. The proposed approach includes the training data creation using the Monte Carlo simulation based on the image formation model typical for image-based navigation. Then, based on the generated data, we calculate seventeen features that we have determined to be relevant for this task. The analysis of the feature importances and different machine learning algorithms performances is carried out. As the source maps, we used real images obtained by the AVIRIS radiometer and taken from Yandex maps. The proposed approach is superior to the separately taken Cramér-Rao bound method or the bootstrap simulation which is not surprising since the use of machine learning allows you to combine different sources (algorithms) using their estimates as features.