Two-step registration of near-space remote sensing images via deep neural networks

Abstract

Near-space remote sensing image registration is an important foundation of near-space image processing. For large image jitter distortion, geometric and atmospheric distortion of its image, we propose a two-step method based on deep neural networks, which includes a coarse-to-fine registration process. We construct a near-space image registration dataset which is captured from Google Maps and hot air balloon platforms, etc. For obtaining candidates, the coarse alignment stage applies classical geometric validation methods to a corresponding set of pre-trained deep features. The fine alignment network is based on pyramidal feature extraction and optical flow estimation to realize local flow field inference from coarse to fine. We construct a regularization layer for each level to ensure smoothness. Applying our method to our synthetic dataset, the experimental result shows that it has a competitive result that is evaluated based on the root mean square error, peak signal to noise ratio and structural similarity.