Research On Astronomical Image Restoration Based On Generative Adversarial Network

Astronomical observation is often disturbed by many factors,resulting in various forms of degradation of the collected images.Common degradation types include defocus blur,motion blur,atmospheric turbulence blur,etc.,as well as photoelectronic noise pollution,etc.In practice,complex noise and blur hybrid degradation often occur,which is difficult to recover.With the advanced development of astronomical observatory and supporting equipment,in the field of astronomical observation data volume has reached the PB magnitude classic image restoration method is difficult to deal with massive amounts of astronomical image data,In addition,traditional restoration methods are difficult to recover high quality images which can be used in astronomical research from mixed degraded astronomical images,and there are problems of low restoration efficiency and poor restoration effect.At present,machine learning methods shine brilliantly in the field of image restoration.It is of great application significance to use the accumulated image data and machine learning methods to guide the follow-up astronomical observations.In this paper,two algorithms of generating adversarial networks(GAN)for degraded images in astronomical observations are proposed.The following is a brief introduction to the work of this paper:In this paper,a GAN image restoration algorithm based on astronomical observation prior information is proposed.The fuzzy kernel is extracted from the actual defocused fuzzy image to expand the data set,and then the mapping relationship from the clear image to the defocused degraded image is automatically learned by the machine.Improved generation network architecture,using the leaky-Re LU and Re LU activation function to extract celestial details in depth,and optimized the countermeasure loss function,adding smoothing improvements on the basis of the original loss function to improve the stability of gradient transfer.In addition,aiming at the ill-posed problem in the restoration of mixed degraded images,that is,the restoration effect is not ideal when noise and blur are processed at the same time,it is proposed to use UNet++ to improve the feature extraction network and combine label smoothing method to improve the generalization performance of the restoration network.At the same time,the gradient cutting method is introduced to improve the training stability.Experiments show that the improved method has better image quality and is suitable for astronomical research with large amount of data.