Research And Application Of Non-rigid Medical Image Registration Algorithm Based On Feature Points

Non-rigid image registration is a very basic and important research contents in the field of computer vision and pattern recognition.Non-rigid image registration method based on feature points application in medical image,military automatic target recognition,satel ite images and data editing and analysis have played a big role.It is also the focus of the research in this field.For different application scenarios,there will be differences between images,so it is necessary to propose a universal and robust point set registration algorithm.In this way,for different application scenarios,we only need to extract the feature points by using the prior knowledge,then we can use this algorithm for image registration.In this paper,a registration algorithm based on Gauss mixture model(GMM)with strong robustness is proposed to deal with the registration problem of data degradation.In order to solve the problem of non-rigid point registration,the key is to solve the problem of finding the corresponding relation and to calculate the space transformation function.As for data degradation,outlier and noise increasing,there is not a good method to solve this problem perfectly.This requires us to focus on not only the global information of the image but also the local information of the image in the image registration process.In order to solve the problem of finding correspondences,in this paper,the registration problem is considered as a probabilistic estimation problem,and the Gauss mixture model is used to model the target feature point sets.We will transform the problem of finding point correspondences into the problem of estimating the mixed density,so that the Gauss centroid of one feature point set is consistent with another feature point set.We add KL information entropy regularization and using shape descriptors to initialize the parameters of the model,using the method of deterministic annealing to optimize our model,to ensure that our method can correctly complete the task while existing outliers and noise.In order to calculate the space transformation function,in this paper,we use the non-rigid space mapping of the Thin Plate Spline plane to parameterize the space transformation of our solution.The space mapping is described as a combination of affine transformation and non affine transformation.The method of alternately updating correspondence and space transformation function is used to solve the problem until convergence.We designed four sets of experiments with different emphases,given the registration results and quantitative evaluation method,we proved the universality and robustness of our method,and compared with several advanced methods have been published.Experimental results show that the proposed algorithm can obtain satisfactory registration results under the condition of data degradation.Finally,the registration experiments of brain MRI images are carried out,and it is proved that our algorithm can be applied to the direction of medical image registration.