Research On Rigid Registration Algorithm Of Medical Images Based On Distribution Estimation Algorithm And Mutual Information

Image registration is an important subject in image processing, which also firstly needes to be solved for image fusion. Image registration technique is of great significance in many practical applications, for instance, medical image analysis, remote sensing image processing and target recognition problems. The essence of image registration is to calculate the similarity measure between images after a series of spatial transformation, and to find the corresponding optimal transform parameters within geometry transform space by using optimization algorithm. The rapidly progress of science and technology led to the fashion of different modal sensors, therefore the study in which method different modal images could be registrated efficiently also gradually become a research focus in the field of image registration. In order to improve the correct rate and accuracy of registration, researchers hope to reduce the error of data to the greatest extent. As the high-quality need and the real-time demand, the research about how to improve the precision and speed of multi-mode image registration is particularly important. This thesis bases on mutual information, gradient similarity and copula estimation of distribution algorithm to discuss the two important steps of image registration, and summarizes the advantages and disadvantages by comparing with the classical registration algorithms. After doing some research, an improved registration optimization algorithm is proposed.This thesis aims to found a rigid registration algorithm with higher accuracy and stability. First of all, this thesis made a detailed introduction for the measure function. Due to ignoring the image spatial information, the measure function based on mutual information gives rise to the decrease of the registrated precision. When added with the gradient information, the measure function contributes to improve the registrated precision, but the consideration about gradient similarity is not comprehensive, the registrated precision still is not satisfied. This thesis puts forward an improved measure function based on mutual information and gradient similarity which has higher precision having been proved by experimental data. The reason is that the relationship function between the images’ s rotation angle and the measure function is more smooth with relatively fewer local extremum, which partly contributes to avoid the optimization algorithm falling into the local extremums. And then the optimization algorithm are introduced in detail. The optimization of registration parameters is the other most important in image registration process, also directly related to image registration accuracy and efficiency. The weakness of the copula estimation distribution algorithm is analysed. When the optimization algorithm is used for high-dimensional function, the nonparametric estimation method with randomness can result in the parameter of copula function being estimated in error. The erroor eventually leads to the optimization result is not ideal. To solve the weakness, this thesis put forward a new improved optimization algorithm, and the related tests show that this proposed method has obtained an ideal effect.