Robust Kernel Spectral Methods For Registration Of SAR Images

Synthetic Aperture Radar(SAR)is an active microwave remote sensing sensor,which works on all-weather and all-time and produces high-resolution images.SAR image registration is the process of spatially aligning two or more images of the same scene taken at different times,from different viewpoints,or by different sensors.It has been urgently required in many applications such as disaster monitoring and automatic guidance.However,when the scenes in the referece and sensed images contain large geometric deformation,the registration of SAR images is still a challenging task.Kernel principal component analysis(KPCA)and graph spectral method are effective for contour feature extraction and feature matching in SAR images registration.They are refered to as kernel spectral method for simplicity in this thesis considering their equivalence.Nonetheless,the robustness of kernel spectral method is not satisfactory.On the one hand,it is sensitive to speckle noise and the extracted contour feature points contain large position jitter.On the other hand,it suffers from position jitter and outliers in feature matching and there are many mismatches.This thesis is mainly concerned with the the robustness of kernel spectral method for SAR images registration,and has presented six improved kernel spectral models for contour features extraction and feature matching.Simulated and real-world images registration experiments prove the robustness of the proposed methods.The main research contents and innovation points include:(1)To cope with the contour feature extraction problem in SAR image registration,the closeness degree cut(CDC)and local smoothing weighted graph cut(LSWGC)are proposed based on normalized cut.The two methods use regions clustering and smoothing term to suppress the influences of the speckle noise on SAR image segmentation,and avoid the storage of the similarity matrix in graph cut model.Simulated and real SAR image segmentation experiments show they have higher boundary location precision than normalized cut,weighted kernel k means and parametric kernel graph cut,where LSWGC performs best because it makes use of the equivalence between weighted graph cut and weighted kernel k means and the redundancy of the kernel matrix.(2)For the SAR images registration problem with variform objects,kernel subspace matching and nonnegative subspace matching methods are proposed based on subspace learning method.The two models enhance the robustness of graph spectral method by adding a regularization term considering the across-set feature similarity.Compared with the graph spectral matching methods such as KPCA,the spectral embeddings of the two methods combine the feature similarity and the smoothness of local neighborhood matching.Meantime,their complexities are same as graph spectral matching methods.Experiments on simulated points matching and the feature matching of real SAR images show their robustness to position jitter and outliers is better than graph spectral method.(3)Since feature matching methods are vulnerable to similar structures and intensity distributions,there are usually false matches.A collinearity degree criterion is designed to indicate mismatches because kernel correlation components of correct matches show colinear property.Based on the collinearity degree criterion,a robust kernel correlation analysis model and corresponding algorithm are proposed to remove mismatches.The influence of each match on the collineartiy degree is analyzed to indicate false matches in the algorithm.Experiments show that its robustness to position jitter and outliers is better than Random Sample Consensus(RANSAC)and Optimized Random Sampling Algorithm(ORSA).(4)Second-order graph matching methods are sensitive to the scale transform and it fails when there are large scale difference between reference and sensed images.To overcome the problem,the second-order probabilistic graph matching model is extended to the third-order probabilistic hypergraph matching model.It keeps the scale invariance property of the third order hypergraph matching.Different from the classic hypergraph matching methods,it can modify the third-order affinity tensor in the iteration procedure.Specifically,it increases the elements corresponding to the the correct matches and decreases the elements corresponding to the false matches.As a result,it enhances the robustness to position jitter and outliers compared with the classical hypergraph matching methods such as the tensor matching method.