Application Research On Remote Sensing Image Registration Based On Convolutional Neural Network

Remote sensing image registration technology has become a hot research topic because of its practicability and wide application range.Remote sensing image registration technology has been widely used in disaster survey.In order to obtain more effective and comprehensive post-disaster remote sensing information,in this paper,we propose a remote sensing image registration combined with spatial transformation network and gray projection,and a region constrained moving least squares remote sensing image registration method combined with convolutional neural network.The main research is as follows:1.In order to enable the image to be adaptively transformed with reference to the transformation model and reduce the registration time,a remote sensing image registration method combining spatial transformation network and gray projection is proposed.The spatial transformation network model is used to extract the image features and train to obtain the affine transformation coefficients,so that the image to be registered can be adaptive affine transformation according to the affine transformation coefficients to achieve the initial registration purpose.In order to obtain a more accurate registration effect,the gray projection algorithm is used for secondary calibration,and finally the accurate registration of the image to be registered is realized.2.In order to make the image localization of the moving least squares method not change due to the overall deformation,the paper generates feature descriptions by VGG16 convolutional neural network to obtain key feature points.The feature boundary pairs of the image are matched by the method of region boundary constraint,and the best transformation model is calculated by using the moving least squares method for the image in the region.Finally,the point-by-point affine transformation is performed to complete the registration.The experimental results show that the method proposed in this paper maintains the registration accuracy while accelerating the registration speed and has better robustness.In the registration process for local deformation regions,regional constraints are added.So that the global image does not change due to local changes,which improves the invariance of the image structure.