Remote Sensing Image Registration Based on ResNeXt

Deep learning is widely used in image registration, but as the depth of the network model increases, it leads to an increase in the number of network hyperparameters, which makes the network difficult to train and also makes it more difficult to design the network structure. In this paper, a remote sensing image registration method based on ResNeXt is proposed. The ResNeXt network pre-trained on ImageNet is used as the feature extraction network to ensure the improvement of model accuracy with the same number of parameters. The concept of regularization is introduced to prevent the network from overfitting, and the idea of transfer learning is used to continue training on the GoogleEarth dataset to extract features and improve the network convergence speed. Two-way matching is used in the feature matching stage to eliminate asymmetric matching. And the secondary affine transformation is proposed for the problem of inaccurate transformation parameters to accurately estimate the geometric transformation between two images. The experimental results show that the registered images obtained by this algorithm are better and the registration accuracy is improved, which is of practical significance.