Automatic Multi-Source Remote Sensing Image Registration Based On Densenet

With the development of commercial space and the continuous progress of remote sensing satellite imaging technology,a large number of remote sensing data produced abundant information for many earth observation applications,such as change detection,image mosaic,map update and disaster detection.Remote sensing image registration is the basis and prerequisite of these applications.It aligns the spatial coordinates to ensure that images are in the same coordinate system,which provides a guarantee for downstream tasks.However,traditional remote sensing image registration methods often have the problems of low efficiency and poor ability to extract feature points,which affects the registration accuracy;in addition,the imaging mechanism of multi-source images is different,and the multi-source image pairs used for training are insufficient,which makes the task of image registration based on multisource remote sensing fusion extremely challenging.Therefore,from the perspective of feature point extraction and multi-source remote sensing image fusion,this dissertation proposes a Densenet model that integrates attention and combines the generative adversarial idea to carry out research on optical remote sensing images,and optical-SAR images automatic registration,the summarize of this dissertation are as follows:First,an optical image registration method based on the attention mechanism and Densenet is studied.For the problem of low accuracy in image registration,this dissertation proposes a densely connected network(IC-Densenet)feature point extraction strategy fused with attention mechanism.The proposed strategy uses Inception to achieve multi-scale feature extraction,and combines the attention mechanism to obtain more effective image features from the two dimensions of channel and space.Then the features are propagated through the densely connected network,and feature reuse is strengthened.So the problem of low registration accuracy caused by poor feature point extraction ability can be solved.Secondly,a method of multi-source image registration based on CycleGAN is studied.For the problem of insufficient pairs of multi-source images,different imaging mechanisms,and difficulty in effective training,this dissertation uses a generative adversarial idea,and converts multi-source images into homologous images to register.This dissertation uses CycleGAN to optimize the generator from three aspects: sample distribution,pixel mapping,and image reconstruction.By transferring the style of multi-source images,it realizes the mutual conversion between different images,which can not only reduce the impact of imaging differences,but also solve the problem of fewer SAR images than optical images.Finally,this dissertation verifies the effectiveness of the proposed method through optical images registration,and optical-SAR images registration.This dissertation uses root mean square error,peak signal-to-noise ratio,correct matching rate and relative error as evaluation indicators.The experimental results show that the proposed method can not only effectively improve the network’s ability to fit multi-source image features,but also improve the accuracy of automatic registration of remote sensing images.