Research And Application Of Several Key Technologies For Image Registration

The image registration directly affects the practical application of image processing.The main steps of image registration are image feature extraction,image interpolation,transformation model estimation and feature point matching.However,there are some problems at present.First,incomplete feature extraction will cause image matching to fail.Second,the limitations of the interpolation algorithm will result in poor image matching quality.Furthermore,the insufficient transformation model estimation relationship will result in blurred image registration edges.In order to solve the above problems,this thesis carries out in-depth systematic research on the related algorithms of image registration.The main research contents are as follows.(1)An initialization algorithm for convolutional neural networks based on the long-short-term memory network and the multi-layer Maxout network is proposed to realize deep extraction of image features.It effectively overcomes the gradient disappearance and gradient explosion problems in the training of deep learning algorithms.The long-short-term memory network and the selective Dropout algorithm are combined,which effectively overcomes the over-fitting problem in the training process of the convolutional layer.The random sampling consistency algorithm is used to purify the feature point matching to improve the image matching accuracy.This lays a good foundation for subsequent image registration and image recognition to achieve better results.(2)A data-driven image fast decomposition algorithm is proposed.The reconstruction method is estimated to replace the surface interpolation.The variable neighborhood window method replaces the fixed neighborhood window method,and an isotropic window has been taken as a structural element.The isotropic nature of this kind of window is more consistent with the window nature of the image,which is more conducive to image adaptive decomposition.The window decomposition size takes the average of the maximum and minimum values of adjacent windows to make it more in line with the extreme value distribution characteristics of the image information.It overcomes the shortcoming of the source image features and the details of the modal function components obtained by the original bi-dimensional empirical mode decomposition.The algorithm is applied to synthetic texture images and has achieved good results.It lays a good foundation for improving the efficiency and accuracy of image registration and image fusion.(3)A bi-dimensional empirical mode decomposition image interpolation algorithm based on iterative function system is proposed.The problem of low interpolation accuracy in the bi-dimensional empirical mode decomposition and image matching process is solved.Firstly,a piecewise linear iterated function system is constructed based on the theory of iterated function system.It is used to interpolate discrete data in bi-dimensional empirical mode decomposition.The feature points extracted after the image is decomposed are recombined to obtain the original feature points of the image.Then,the original feature points are still interpolated with the constructed iterative function system to achieve adaptive interpolation of image matching.Experimental results show that the interpolation algorithm proposed in this thesis has high precision and high real-time performance,which lays a good foundation for improving the accuracy of image registration.(4)An image registration algorithm based on second-order smooth support vector regression is proposed.Firstly,a second-order smooth support vector regression function is constructed based on the smoothness of support vector machine.Then,the statistical analysis and traction fruit fly optimization algorithm is used to send the error function of the second-order smooth support vector regression as an adaptive parameter to the traction fruit fly algorithm for parameter optimization.Finally,the optimized algorithm is used to estimate the features between the registration image and the reference image,and the mapping relationship is established.The image registration is realized through the mapping relationship.The algorithm is applied to image fusion and image restoration,and has achieved good results.It lays a good foundation for improving the efficiency of image registration.This thesis proposes corresponding solutions to the problems in feature extraction,interpolation processing and feature point matching involved in image registration.At the same time,it is combined with the practical application of image processing.A large number of image processing examples prove that the proposed algorithm has a good application effect.