Research And Application About Registration Algorithm Based On Mixture Framework And Variational Inference

We propose a new point set registration method based on mixture framework and variational inference.In the MFVI algorithm,mixture framework is a coarse to fine registration strategy to automatically process the point sets registration under different conditions,which includes three steps:(1)linear phase,this is a pre-match process;(2)regression process,Gaussian variational mixture model is used to weaken the influence of outliers;(3)nonlinear phase,it is an accurate point set registration process,and the transformation of the point set is restricted to a non-rigid form.The variational inference is used to process parametric optimization.Under the variational inference framework,we designed an isotropic and anisotropic Gaussian Variational Mixture Model(GVMM)to reduce the influence of outliers,and the Dirichlet distribution is used to control the mixing ratio of Gaussian components in order to distinguish the missing points.In order to improve the robustness of MFVI algorithm,we designed fuzzy shape context(FSC)feature and local vector similarity constraint(LVSC).The nonlinear phase in MFVI algorithm,the correspondence of complementary of features is evaluated,firstly.At the same time,a fuzzy shape context(FSC)feature is defined and the Gaussian mixturemodel based on the fuzzy shape context distance and the global feature distance is designed.The spatial transformation of complementary of constraint is updated,secondly.Meanwhile,a local vector feature is defined and the local spatial vector similarity constraint based on local vector feature is built.In addition,we design SIFT algorithm with adjustable threshold strategy to extract the feature points of the image,and we use MFVI algorithm to process feature point registration and image registration.In the experimental part,we firstly test the retrieval rate of FSC features.Then,we use the open dataset to test the performance of the algorithm in point set registration.In comparison with popular algorithms,this algorithm gives accurate registration results,and in most experiments,the accuracy is higher than the currentpopular algorithm.Finally,we test the registration performance of this algorithm in Oxford dataset,retina image and remote sensing image,and compare it with the most advanced methods.In most cases,our algorithm shows good performance.