Research On The Algorithm Of Solar Speckle Image Reconstruction Combining MCycleGAN And RFCNN

The observation of solar activity and the study of its laws are extremely important for human beings.Due to the influence of the earth's atmosphere,ground-based telescopes can only observe obscure solar speckle images.Only high-resolution solar images can play an important role in solar research.However,the traditional reconstruction algorithm has the disadvantages of single feature,more noise,blurred local details,too smooth edges,and easy loss of high-frequency information when reconstructing the solar blob map.Therefore,this article combines the deep learning method to reconstruct the solar speckle image and strengthens the improvement of the solar observation data analysis and processing capabilities.This thesis is mainly based on the image reconstruction technology of deep learning,and conducts in-depth research on the single-frame solar blob image reconstruction of the generative confrontation network and the multi-frame solar blob image fusion of the convolutional neural network.The main research contents are summarized as follows:(1)Considering the low contrast of the solar speckle,similar grain shape,and small differences between frames,in this thesis,the structural features of the input image and the reconstructed image are extracted using the VGG network and these structural features are added to Cycle GAN.Using the network as MCycle GAN,it does not get the high-frequency information of the input image properly using the feature,and it improves the network's ability to reconstruct the high-frequency information of the single frame image using feature loss calculations to create a reconstructed image.There is more detailed information and it is closer to the target image.(2)In view of the problem that atmospheric turbulence and disturbance will cause the blur,distortion and flicker of the image of the solar speckle,the high-frequency information of the single-frame solar speckle image is difficult to be effectively restored,resulting in the MCycle GAN reconstructed image with insufficient edges and relatively high contrast.Insufficiency of the difference,this thesis adds dense blocks and new fusion strategies to the Deep Fuse network and records it as RFCNN.It uses similar but not identical information between image frames to complement each other,and uses a small number of multi-frame images for reconstruction,which further improves the reconstructed image.The visual effect improves the authenticity and stability of the reconstructed image.This thesis uses the original data collected by the 1m New Vacuum Solar Telescope(NVST)from the Yunnan Observatory of the Chinese Academy of Sciences,and after a series of data preprocessing,a data set is constructed and experiments are carried out.The experimental results show that this thesis uses the RFCNN network to perform multi-frame solar speckle image fusion after MCycle GAN reconstructs a single-frame solar speckle image,which greatly improves the reconstruction quality of a single-frame solar speckle image.Under the subjective and objective evaluation criteria,the accuracy of the reconstruction result of the method in this thesis is not lower than that of the traditional algorithm,and the result of multi-frame fusion is better than the traditional algorithm in terms of visual effects and algorithm efficiency.