Astronomical Image Restoration Based On Machine Learning

The astronomical image is affected by the atmospheric turbulence and the electric noise of CCD in the process of during observations.The quality of observation images degraded,which results to the loss of observation information and affects the observation efficiency.The image quality can be improved by restoration.Most of researches on image degradation adopts deconvolution method.PSFs in the whole field of view is considered to be uniform.In fact,for the wide field of view short exposure imaging system,the PSF of the image changes with time,field of view.The deconvolution method with uniform PSF can not achieve good restoration result.Based on the above situation,we propose to use partition image restoration method for the wide field astronomical images,based on the statistical properties of PSFs and an image deconvolution neural network based on deep learning in this dissertation.In segment deconvolution algorithm,skylight background of image is estimated firstly.Then coordinate positions of stars in the image are extracted,and the PSFs of the stars are extracted.In order to eliminate the noise and reduce the computational load,we use the principal component analysis method and selforganized map to cluster images.The clustering results are labelled by selected evaluation indexes.The clustering results are combined with star coordinates to segment images,so that the segmentation results include the distribution characteristics of PSFs in the field of view.Finally,the PSFs of clustering are used to restore the sub-images after segmentation,and then the restoration results are spliced to get the segmentation deconvolution restored images.Compared with the traditional zonal restoration method,the restoration results in this paper are more stable in terms of signal-to-noise enhancement ratio,which verifies the effectiveness of the proposed algorithm in restoring wide field of view astronomical images.Applied restoration method in this dissertation to Quadchannel telescope image co-adding,the quality of results improved.For the deconvolutional neural network algorithm,image restoration is discussed according to the mathematical model of Wiener filter.And the principle of the convolutional neural network is explained by the decomposability of PSFs.Then,the network framework is built,and the model is trained with simulated data.Finally,the restoration effect is evaluated based on PSNR and SSIM,which proves the feasibility of the deep deconvolutional neural network in this dissertation.