Accurate Registration Of Polarimetric SAR Images And Rapid Implementation Technology

Synthetic Aperture Radar(SAR)can work in all-day and all-weather environments,and has a certain penetration ability,so it is widely used in image fusion,ground change detection and other fields.Compared with traditional SAR systems,Polarimetric Synthetic Aperture Radar(Pol SAR)can measure the reflection of ground objects in multiple polarization modes and can obtain the polarization information contained in the echoes of ground objects.Remote sensing image registration aims to align the images spatially by extracting and matching features or searching for the maximum similarity measure.High-precision registration of Pol SAR images is a prerequisite for subsequent image processing,but the registration accuracy of Pol SAR images still needs to be further improved.How to make full use of polarization information to achieve high-precision registration of Pol SAR images and how to improve the real-time performance of the registration algorithm will be the focus of this thesis.This thesis has carried out related research on the accurate registration of Pol SAR images and its parallel optimization realization,and the main work is as follows:(1)Aiming at the problem that the polarization information is not fully used,a structural feature extraction method that makes full use of the polarization information is introduced.Regarding the polarization coherence matrix as a multi-channel image,the GR gradient operator is used to calculate the partial derivative of each channel image in the horizontal and vertical directions,and then the Di Zenzo structural gradient tensor is used for information integration.The structural feature extraction method given in this thesis can get more detailed features and can better reduce the influence of noise.(2)Aiming at the registration problem of polarization SAR images with large noise,Pol SAR-SIFT registration algorithm based on fusion of polarization features and corner points is proposed.On the basis of the SAR-SIFT algorithm,the structural feature extraction method proposed in this thesis is applied to the construction of multi-scale feature space;in the feature point extracting stage,an adaptive threshold is used instead of an empirical threshold to obtain more stable feature points;in the feature matching stage,a more robust estimator called fast sample consensus method can be used to obtain more robust results in fewer iterations.The registration algorithm can obtain more accurate and stable registration results,and can still complete the registration task well in scenes with large changes.(3)Aiming at the problem of accurate registration of Pol SAR images,a coarse-to-fine accurate registration algorithm for Pol SAR images is proposed.In the coarse registration process,the Pol SAR-SIFT algorithm proposed in this thesis is used,and the estimated transformation model parameters are used as the initial parameters of the fine registration process;in the fine registration process,a mutual information measure based on polarization information is used,and the Powell algorithm is used to perform search optimization and complete the accurate registration of Pol SAR images.(4)Aiming at the time-consuming problem of the Pol SAR-SIFT algorithm,a parallel optimization implementation of the Pol SAR-SIFT algorithm is given based on CUDA.First,the parallelism of each step of the algorithm is analyzed.Then,a large number of two-dimensional convolution operations in the process of constructing the multi-scale feature space are decomposed into one-dimensional convolution operations,and the image blocks are loaded into the shared memory in the manner of merging memory access.As a result,a large speed increase is achieved.In the feature matching stage,replacing Euclidean distance with cosine similarity can reduce the amount of calculation.By reasonably allocating thread tasks,reducing global memory access and reasonably coordinating the specific tasks of the host and GPU devices,the utilization of GPU resources is improved.Through parallel optimization,a speedup of nearly 30 times has been achieved.