Research And Application Of Outlier Processing Method In Image Registration

Image registration is the basic task of image processing,in which two or more images of the same scene are registered to ensure that the image pairs are aligned on a common space axis.The images can be multi-view(obtained from different viewpoints),multi-temporal(obtained from different times)and multi-source(obtained from different sensors).Successful image registration is a prerequisite for many remote sensing applications,such as natural disaster damage assessment,resource census,environmental monitoring,ground target identification,and map updates.However,problems such as scaling,geometric distortion,and low overlap rates often occur during image capture,which will lead to the matching of point sets more and more difficult in image registration.The paper first proposed an iterative image registration algorithm to solve the problems in the above image capture process,named remote sensing image registration method based on dynamic threshold calculation strategy and multiple-feature distance fusion.First,the dynamic threshold calculation strategy is used to gradually screens reliable inliers to reduce the negative effect of outliers during the iterations.The multiple-feature distance fusion Gaussian mixture model is then applied to compensate for the defect of a single feature,and the result of the dynamic threshold calculation strategy acting as the prior probability combines with the deterministic annealing thought to achieve the optimal mapping of image pairs from local to global scale.Moreover,the structure constraint based on local applying force is added into the global structure constraint to control the alignment of feature points more accurately in the overlapping area,so as to guide the subsequent image transformation.Next,to further improve the accuracy and efficiency of the algorithm,the paper proposed a robust image registration method using two-layer cascade reciprocal pipeline and context-aware dissimilarity measure.Firstly,a two-layer pipeline is built to find the optimal correspondence and to achieve the image transformation.In the first layer,the neighborhood structures of point sets areexploited to find and register reliable feature point sets.In the second layer,a point-to-point neighborhood structure is created to recover the remaining potential inliers for image registration.On this basis,the context-aware dissimilarity measure with scaling,translation and rotation invariance is then added in the two-layer pipeline to evaluate the difference of feature points more accurately.For experiments under two different frameworks,feature matching and image registration experiments are designed respectively in the image registration algorithm of the iterative framework.In comparison with nine state-of-the-art methods,the results show that the proposed algorithm is superior to the others in most cases.In the image registration algorithm of non-iterative framework,in addition to the conceive of feature matching and image registration experiments,image retrieval experiments are also designed,which are compared with twelve state-of-the-art methods.In particular,when the inlier ratio is lower than 0.5,our method still achieves a better accuracy-efficiency tradeoff.