Research On Registration Methods Based On Feature And Frequency-domain Similarity Metric For Multi-source Remote Sensing Imagery

In recent years,remote sensing images have been widely used in the fields of resources,military,meteorology,agriculture,environment,and transportation.The collaborative processing of multi-source remote sensing images can make full use of the spatial-spectral information of different remote sensing images,which is significant for improving the performance of various remote sensing application systems in the military and civilian fields.Multi-source remote sensing image registration is a necessary prerequisite and key technology for collaborative processing of multi-source remote sensing images,and it is one of the research hotspots of remote sensing technology.Remote sensing image registration methods are generally divided into two categories: feature-based registration methods and region-based registration methods.Feature-based registration methods are robust to scale and rotation differences,but are susceptible to differences in non-linear grayscale and texture.Region-based registration methods are highly dependent on the accuracy and robustness of metrics.Usually,the construction of metrics is based on gray-scale statistical characteristics to measure the similarity of overlapping regions,and lack of the use of spatial structure information,leading to poor registration accuracy.In response to the above problems,combining with the robustness of feature-based registration methods for scale-rotation differences and the high accuracy of registration methods based on frequency domain similarity metrics,this paper mainly focuses on image registration based on features and frequency domain similarity metrics.The main research work of this article is summarized as follows:(1)Aiming at the problem that the high mismatch ratio in the feature-based remote sensing image registration method leads to inaccurate estimation of the registration parameters,a registration method based on the similarity of the global feature triangles is proposed.The method evaluates the global geometric similarity of the matching feature pairs by constructing a mathematical model,which uses the similarity principle of characteristic triangles,formed by any three correct matching feature pairs between the reference image and the sensed image,to remove outliers.Furthermore,the finally preserved correct matching feature pairs are used to calculate the registration parameters.Simulation and real data experiments validate the accuracy of the proposed outlier elimination method in this chapter.(2)Aiming at the problem that the feature-based remote sensing image registration methods cannot reduce the inherent error of the matching feature pairs,which leads to the insufficient accuracy of the registration parameter estimation,a registration method based on outlier elimination and feature point position adjustment is proposed.This method eliminates outliers with very large matching errors by relatively looser outlier elimination strategy,and uses the feature point position adjustment method to further reduce the inherent error of the remaining matching feature pairs.At the same time,a similarity metric based on phase structure consistency is proposed as an measurement for outlier removal and feature point position adjustment.Simulation and real data experiments validate the accuracy of the similarity metric in this chapter and the effectiveness of the registration strategy.(3)Aiming at the high dependence of feature-based remote sensing image registration methods on initial matching features,which leads to the limitation of registration accuracy,a registration method based on features and extended phase correlation is proposed.This method uses the feature-based registration method to estimate the initial registration parameters,which are used to rectify the sensed image to obtain rectified sensed image.Then,the fine registration parameters between the rectified sensed image and the reference image are calculated by the extended phase correlation method.Final registration parameters are obtained by fusing these two registration results.Simulation and real data experiments validate the efficiency of the registration framework that combines features and extended phase correlation methods.(4)Aiming at the problem of insufficient information used in the similarity metric construction in the region-based method,resulting in limited registration accuracy and robustness,a registration method based on the global hybrid structure similarity metric is proposed.a hybrid structural similarity metric is constructed by combining the structural similarity in the gray domain and the phase consistency in the frequency domain.Thus,the proposed metric can ensure a certain degree of robustness to grayscale differences and also make full use of sensitivity of the phase information to geometric deformation.Simulation and real data experiments validate the accuracy and robustness of the similarity metric constructed in this chapter.