Remote Sensing Image Registration Based on Grid Prediction of Pixel Deformation

Aiming at the problem of weak texture blurring in remote sensing image registration under complex conditions, a remote sensing image registration algorithm using grid generator to predict pixel deformation is proposed. We first improve the pixel-level deep fully convolutional encoder-decoder network using overlapping attention in the feature extraction stage to obtain the features of the two input images. The feature matching stage then utilizes a mesh generator to predict and parameterize pixel deformations. The network is then trained using the grid loss. The final parametric regression stage uses bidirectional parameters to optimize the network loss.