Research On SAR Image Registration Technology Based On Point Cloud Feature

Synthetic Aperture Radar(SAR)is an active microwave remote sensing imaging system,which is widely used in civil and military fields such as disaster rescue,environmental monitoring and target detection because of its significant advantages of all-day and all-weather reconnaissance.In most applications of SAR,image registration is a very important step,and the accuracy of SAR image registration directly affects the effectiveness of subsequent applications.At present,the feature-based registration method has become the main research direction of image registration because of its advantages such as high adaptability.However,the performance of the registration is affected by the existence of speckle noise and scattering differences in SAR images.Point cloud is a collection of massive points reflecting the surface characteristics of an object,which can accurately reflect the spatial distribution of 2D and 3D targets.This thesis researches SAR image registration methods based on the point cloud features of SAR images,and the main work of the thesis is as follows:A structural point cloud-based SAR image registration framework is proposed.The framework consists of several steps of structured point cloud extraction,point cloud matching and image transformation.To improve the registration accuracy,a structural point cloud extraction network(SPCE-Net)is trained to extract the point clouds reflecting the geometric structural features of SAR images.Then,the structural point clouds are matched to estimate the transformation relationships between images.The experiments on the measured dataset show that the proposed SPCE-RPM-Net method outperforms the traditional registration methods in terms of various evaluation metrics such as mutual information(MI),structural similarity(SSIM),entropy correlation coefficient(ECC)and visual comparison.A feature recurrent super-resolution network(FRSR-Net)is proposed for image quality enhancement in response to the problem that low-resolution SAR images lead to difficult extraction of structural point clouds or inaccurate extracted point clouds.Since SAR image resolution degradation is affected by several factors,a single step of prediction may not be able to recover image details.For this reason,FRSR-Net is constructed by feature recurrent structure and residual dense block(RDB)to achieve fine image reconstruction by using high-level features to refine low-level features.It is shown experimentally that FRSR-Net has better super-resolution performance compared with the remaining two comparison methods under peak signal-to-noise ratio(PSNR)and SSIM evaluation metrics.A super-resolution structural point cloud matching method(S~2-PCM)is proposed by integrating FRSR-Net into the structural point cloud registration framework.The method first performs super-resolution processing of SAR images using FRSR-Net,and then uses the SPCE-RPM-Net method proposed in this thesis for registration.Experiments on simulated and measured data show that S~2-PCM has higher alignment performance compared with the traditional registration method under various evaluation indexes such as MI,SSIM,and ECC.The ablation experiments show that the super-resolution processing significantly improves the quality of structural point cloud extraction and the accuracy of SAR image registration.The research results of this thesis can provide technical support for applications such as moving target detection and shadow tracking.