Research On Remote Sensing Image Registration Based On ResNet

Remote sensing images refer to images with certain target characteristics obtained through various imaging systems(such as aviation and satellites).Remote sensing image registration technology refers to the process of geometrically calibrating two or more remote sensing images acquired in the same scene with the same or different remote sensing equipment in different shooting periods,different shooting angles,and is indispensable in remote sensing image processing.It plays a vital role in large-scale remote sensing detection systems such as urban change detection,geological disaster monitoring,environmental monitoring,geographic mapping,and detection guidance.Due to the inability of traditional remote sensing image registration technology to accurately capture the complex geometric features of remote sensing images,and the high labor cost and lack of selflearning ability,the performance achieved in remote sensing image registration tasks is far inferior to natural image processing tasks.This thesis proposes two kinds of registration algorithms based on deep neural networks to improve the accuracy of remote sensing image registration and enhance the robustness of registration models in various fields,mainly including the following two aspects:1.In view of the lack of key information in feature extraction and low registration accuracy caused by the complex shooting environment of remote sensing images,this thesis proposes a remote sensing image registration algorithm based on asymmetric convolution dense residual network.In this thesis,the 34-layer ResNet structure is used for feature extraction,and asymmetric convolution and dense connection are used to improve it,while paying attention to the local and global information of the image.In the feature matching stage,two-way matching is used to solve one-way matching asymmetrical problem.This thesis also proposes a quadratic affine transformation to solve the problem of inaccurate estimation of transformation parameters,so as to obtain parameters that can better represent the actual transformation between two images,and improve the registration accuracy.2.In view of the problems of low registration accuracy caused by complex feature descriptors and difficulty in learning the invariant mapping function in the feature matching stage in the existing image registration framework,this thesis proposes a remote sensing image registration algorithm based on three-stage matching of image pathces,to solve the image distortion that may be caused by image downsampling.First of all,the image is cropped with the feature points extracted by the Shi＿Tomasi algorithm as the center.And a three-stage matching algorithm is proposed in the image patches matching stage,which performs one-to-many matching,one-to-one matching and image patch distortion compensation on the image patches respectively to find exact matching between image patches and eliminate geometric deformation between two image patches.Finally,the greedy algorithm is used to homogenize the feature point distribution and estimate the global affine transformation parameters to complete the registration.In the experiment,multiple remote sensing datasets were applied to conduct comparative analysis with several current popular methods,including classic traditional algorithms and deep learning algorithms.The experimental results show that the registration result obtained by the algorithm in this thesis has a better effect,and the registration time and registration accuracy have been improved.And it solves the problems of low registration accuracy caused by complex feature descriptors and difficulty in learning the invariant mapping function in the feature matching stage.