Research On The Non-rigid Point Set Registration Method Based On Gaussian Mixture Model

Non-rigid point set registration is the prerequisite and foundation of the non-rigid image registration.Non-rigid point set registration has been widely used in medical image detection,military target recognition,remote sensing images,and data analysis,etc.Due to some factors,such as complex deformation,changeable application scenarios,and the interference of noise or occlusion,it is a challenge task to achieve the registration of non-rigid point sets.This thesis proposed a non-rigid point set registration approach using the neighborhood structure and driving force criterion.Besides,a local distance calculation method is proposed on the basis of Coherent Point Drift algorithm.By combining it with the spatial distance,the matching accuracy is improved.In addition,this thesis improved the traditional shape context approach in the registration process,and constructed a new driving force criterion,which can increase the search speed during the initial registration process and reduce the registration error in the later stage.Finally,the Expectation Maximization algorithm is used to iteratively solve the correspondence and the spatial transformation function.Experimental results compared with the existing methods show that the proposed method is robust in the case of non-rigid deformation,noise,outliers and occlusion.Besides,the proposed algorithm can obtain ideal registration effect for real images.We proposed a new probabilistic model,called Asymmetric Gaussian model,to capture the spatial asymmetric distribution.First,a mixture of asymmetric Gaussian model is used to represent each point set,and the soft assignment technique is used to recover the correspondences.And then,the correlation-based method is utilized to estimate the transformation parameters between the two point sets.At last,the problem of point set registration is transformed to the optimization problem,which is solved using the regularization theory.In the proposed algorithm,the computational complexity is reduced by establishing the kernel with the low-rank kernel matrix approximation.The registration results on the classical 2D,3D non-rigid point sets and real images demonstrate that the proposed algorithm can obtain satisfactory results even for the degenerated data.In addition,we proposed a robust point set registration algorithm based on D divergence.After modeling the whole point set using a Gaussian mixture model(GMM),the registration problem is transformed into solving the minimum problem of the D divergence between the two GMMs.The problem can be solved using EM algorithm.The simulation results on different datasets demonstrate the effectiveness of the proposed method.Finally,the work in the thesis is summarized,and the study direction in the future is also expected.