Research On Image Registration Method Of Common Aperture Multispectral Camera

The co-aperture multi-spectral camera can acquire different spectral information of objects in the same scene by utilizing the characteristics of the same field of view.This paper mainly studies the co-aperture multi-spectral camera based on array splitting,which has the advantage of instantaneous acquisition.However,due to the design of multiple cameras,it is easy to produce geometric error on the image in the case of assembly and transportation.Find the spatial transformation relationship between multiple images of two images:compared with ordinary images,multi-spectral images have the problem of large difference in gray scale,plus the influence of different scenes,which increases the difficulty of registration and the accuracy of registration.The effect is not conducive to the subsequent panorama stitching or the extraction of spectral information.In this paper,the imaging principle and data characteristics of the co-aperture multi-spectral camera are analyzed firstly.For multi-spectral image registration,the characteristics-based and transform-domain based methods are mainly studied to test the existing feature extraction algorithms and analyze their advantages and disadvantages.And adaptability;and for the problem of large gray scale difference of different spectral segments,a method of level-by-level registration is proposed.The main idea is to use the data itself to calculate a set of registration routes,and then perform two-two registration according to the registration route.The calculation of the registration matrix is performed on any two different spectral images before,which improves the robustness and accuracy of the overall algorithm.Aiming at the problem that the traditional feature extraction algorithm is difficult to maintain the local precision and edge detail of the image,a multi-spectral image level-by-level registration method based on A-KAZE feature extraction algorithm is proposed.The method uses the fast display diffusion(FED)numerical analysis framework to solve the nonlinear diffusion filter equation in the two-two registration process,constructs the nonlinear scale space,and obtains the feature points by calculating the Hessian matrix of each pixel,using binary description.The(M-LDB)constructs the image feature vector with scale and rotation invariance.In this paper,the pyramid layer number and descriptor are optimized for the data characteristics of the co-aperture multi-spectral camera,which improves the speed of the registration algorithm.The eigenvectors are KNN matched by Hamming distance,and the matching pairs are refined by RANSAC algorithm.Finally,the transformation matrix is calculated based on the projection transformation model.For multi-spectral images from different channels,multi-spectral images are used to calculate multiple images.Optimal registration strategy,image registration through registration strategy and transformation matrix.In order to compensate for the shortcomings of the feature algorithm,a Fourier-Merlin Transform(FMT)-based sequential registration method is proposed.This method divides the entire image in the two-two registration process and performs frequency domain transformation on the divided sub-regions.Correspondence of the position of the center point is performed to form a more accurate matching point,and the parameter of the transformation matrix is estimated by the RANSAC for the matching point to obtain a transformation matrix.In some scenarios.FMT is used instead of AKAZE in the calculation of the transformation matrix between the advanced number channels,which can obtain more accurate transformation parameters.It can be called AKAZE feature extraction algorithm and FMT joint level-by-level registration algorithm:using AKAZE to calculate the transformation matrix of low-order numbers betw een the channels,using FMT to calculate the transformation matrix between the advanced number channels,and finally get all the matrices.In this paper,the proposed algorithm is tested,and the registration accuracy of sub-pixels can be obtained,and it has strong robustness.