Research On Algorithm Of Accurate Registration Of SAR Image

Synthetic Aperture Radar(SAR)emits microwave to monitor the Earth's ground continuously,and is widely applied to military and civilian needs.Registration is of great importance for SAR image processing and applications by impacting the detection of surface variation,the acquisition of sudden disasters,target tracking,and motion analysis,having always been the hotspot in SAR remote sensing.This research mainly focuses on feature extraction and feature matching with the multi-temporal,multi-sensor,and multiband satellite SAR images,as follows:In order to solve the drawbacks in the Scale Invariant Feature Transform(SIFT)algorithm,i.e.,missing details,blurred edges,and incapacity in anti-speckle noise,we provide a multi-feature SAR image registration method based on multi-scale bilateral filtering.Instead of using the traditional Gaussian filtering,the proposed algorithm takes advantage of the multi-scale bilateral filtering and multi-scale detail enhancement to construct a multi-scale bilateral filtering nonlinear scale space,effectively suppressing the interference of speckle noise and preserving the edges and details of SAR images.The algorithm developed in this study also realizes the complementary of multiple features because of the extracting of Blob features and Laplace-Harris corners,increasing the extracted features and improving the unevenly distributed features.The registration results of SAR images show that the proposed Multiscale Bilateral Filtering-Scale Invariant Feature Transform(MBF-SIFT)algorithm increases the correct matching logarithm by 1.5 times and improves the registration accuracy by 0.2175 pixels,compared with the SIFT algorithm.This research also improves the SAR image registration method based on deredundancy and FFSC to decrease the influence of "pseudo-feature point" speckle noise and random sample iterations in Fast Sample Consensus(FSC).In order to reduce the feature extraction time,repeated gray values in the image are removed.Using the matching filtering algorithm to remove outliers,and thereby reducing the iterations in the FSC algorithm and enhancing the randomness of the sample dataset.Gaofen-3measurements are employed to test the performance of the proposed registration algorithm.The registration results indicate that the proposed algorithm effectively reduced the running time of 74 s by removing de-redundant features.Compared with the MBF-SIFT algorithm,the registration accuracy of the proposed algorithm is improved by0.1257 pixels.An improved multi-modal image registration method based on Phase Consistency(PC)and structural similarity is developed to resolve the difficulties,i.e.,insufficient extracted feature and constructing feature descriptors from difference in nonlinear intensity between multi-modal images.The improved method extracts the PC information from the multimodal images,and then uses weighted to construct a phase-consistent structure map of images.Dividing the structure map into different sub-regions relies on the PC information.Furthermore,generating Phase Gradient Consistency Structure(PGCS)descriptors with the directional gradient channel features extracted from the subregions.The multimodal images registration results report that the improved algorithm is able to effectively reduce the nonlinear intensity difference between multimodal images,and enhance the accuracy and significance for the structural feature descriptors.Compared with the directional channel gradient features,the matching accuracy of PGCS descriptor feature is improved by 8.2% and the registration accuracy being improved by0.3039 pixels.