Research On Super Resolution Reconstruction Algorithms Of Astronomical Image

As the core index to measure the performance of the astronomical imaging system,image spatial resolution directly determines whether the all-weather,multi-target and high-definition observation needs can be achieved.However,the astro-nomical imaging process is affected by many factors,such as system aberration,atmospheric turbulence,high vacuum and extreme temperature,which inevitably lead to a significant reduction in the spatial resolution of the actual observed astro-nomical images.The effect of improving the spatial resolution of astronomical im-ages through"hardware approach"is extremely limited.Therefore,this paper stud-ies how to use super-resolution reconstruction technology to break through the the-oretical diffraction limit of the astronomical imaging system,meanwhile,restore the original appearance of the target astronomical scene and improve its clarity as much as possible.This paper firstly introduces the proposal,basic principle and research status at home and abroad of image super-resolution reconstruction algorithm,meanwhile,discusses the observation model,specific constraints,technical approaches and evaluation system of astronomical image super-resolution reconstruction task around the main degradation factors of astronomical imaging process,and finally studies and solves the sub problems of super-resolution reconstruction such as se-quence aliasing astronomical image registration,non-uniform sampling data fusion,deconvolution,joint estimation of degenerate function and sparse modeling.The main research contents and contributions of this paper are summarized as follows:Aiming at the characteristics of noise pollution,spatial aliasing and strong di-rectionality in observed astronomical images,a parameterized global image regis-tration algorithm based on spatial transformation model is proposed,which com-bines the sub-pixel motion parameter estimation method based on frequency domain and spatial domain to complete the registration process of sequence aliased astro-nomical images.The rotation angle is estimated by calculating the spectrum com-ponent function of the astronomical image to be registered and the maximum value of its cross-correlation function curve.A linear regression model is established ac-cording to the phase spectrum change matrix,and the overdetermined matrix equa-tion is solved by the least square method to reduce the impact of astronomical image noise on the translation distance estimation.Finally,the spatial domain registration algorithm based on M-pyramid structure is used to realize the fine estimation of small angle rotation.A super-resolution reconstruction algorithm based on improved normalized convolution frame is proposed to deal with the degradation factors such as noise or outliers in astronomical observation.The mixed function based on Gauss function and Laplace function is used as the certainty function to suppress noise and outliers simultaneously.The structure tensor is calculated according to the Gaussian smoothed image and the effective local structure information is extracted.Com-bined with the scale space parameters,the structure adaptive applicability function with anti noise ability is constructed,and the certainty function and applicability function are introduced into the facet model based on quadratic polynomial to com-plete the fusion restoration of non-uniform sampled data.Finally,the Tukey norm is used to replace the₂L norm or₁L norm as the data fidelity term to suppress the impact of various noises on the reconstruction results and improve the robustness of the deconvolution process.Aiming at the problem that the super-resolution reconstruction algorithm under the"divide and conquer"framework can not avoid the estimation error,and can not be corrected and adjusted in time,a hierarchical Bayesian"joint estimation"frame-work based on multi-layer prior distribution is proposed.Combined with Bayesian formula and variational inference method,the posterior distributions and their closed-form variational distributions of the original image and degradation func-tions are solved respectively.The optimal estimates of all probability distribution parameters are obtained by numerical iterations through the mutual restriction rela-tionship between the variables to be solved.In addition,the composite statistical prior model is used to locally describe the astronomical image,and the unique tex-ture features and geometric structures of the astronomical image are restored to the greatest extent.In order to solve the ill posed problem caused by the insufficient number of sequential astronomical images,a super-resolution reconstruction algorithm for sin-gle astronomical image based on adaptive sparsity prior model is proposed.The richness of geometric structures extracted by the gradient operator,such as edges and textures in the sample images,is improved and makes it has rotation invariance.Considering the common characteristics between different training samples,an over complete dual sub dictionary with special direction is constructed by combining the K-means clustering algorithm and K-SVD algorithm.At the same time,according to the high-frequency information of the image block,the principal component anal-ysis is used to complete the adaptive selection of the dual sub dictionary.Finally,the non local sparse decomposition mean is used to constrain the sparse coding error to improve the sparse decomposition accuracy of the input image block.To sum up,this paper makes a detailed analysis and specific improvement on the basic theory and algorithm design involved in astronomical image super-resolu-tion reconstruction at this stage,focusing on solving the related sub problems of interpolation,denoising,deblurring,anti aliasing and image registration in the com-plex astronomical imaging environment.However,in the actual implementation of large field of view dynamic observation or multi-color sky survey observation tasks,there are still some problems to be solved and studied,such as local motion,blind deblurring and the application of deep learning algorithm.