Neural Network Architectures for Assessing the Accuracy of Image Registration

In this paper, we study the problem of assessing the accuracy of image registration. Image registration is considered concerning the problem of remote sensing of unimodal images of the Earth. A simulation model based on twostep registration (SIFT, normalized cross-correlation) has been created to form a dataset for estimating the probability of correct registration and the root mean square error for registering small patches of real remote sensing maps. The paper analyzes the use of well-known approaches for assessing registration accuracy (for example, based on bootstrap), as well as machine learning methods: using handcrafted feature extraction, convolutional neural network, and applying the attention mechanism. It is shown that machine learning solutions (in particular, neural networks) can improve the accuracy of assessing registration quality.