Research And Application Of NVST Image Reconstruction Algorithm Based On Weakly Supervised Learning

The 1m New Vacuum Sun Telescope is affected by atmospheric turbulence,and the observed images are often blurred or severely degraded,have more noise and lose local details.Although deep learning has been widely used in image reconstruction in recent years,they are usually only suitable for motion or dither blurring.The reconstruction of the solar images by them still has problems such as loss of high-frequency details,generation of artifacts,and over-smooth edge contours.Therefore,an improved reconstruction method based on Generative Adversarial Network(GAN)and Weakly Supervised Learning(WSL)is proposed.The main contributions of this thesis are as follows:(1)A deblurring model(DSFSP)based on coupling double-stage feature pyramid networks(FPN)with a single pipe-line is proposed to reconstruct high-resolution solar speckle images.In stage1,it uses one FPN to recover structure features;In stage2,it uses another FPN to enhance the structural contextualized,and uses the single pipe-line coupled with this FPN to extract gradient information.After fusing these to generate a reconstructed image,discriminators are used to make it closer to the reference.(2)A blurring classification method for solar speckle images is proposed,which can effectively separate and train image patches with different distributions.As different areas of the solar speckle image have different degrees of blurring,this is similar to the appearance of multiple different distributions in a single shot image.If multiple distribution data are directly applied to model training,it would lead to severe overfitting of the model.Based on our methods,this situation would be effectively reduced.(3)The supervised training method usually requires GT images,but the newly captured solar speckle images only have degradation images.Considering that,this thesis proposes a training method based on degradation and inverse degradation,which effectively transforms the supervised training of the solar speckle image into weakly supervised training.The method in this thesis uses the solar speckle datasets provided by the Yunnan Observatory of the Chinese Academy of Sciences for training and testing.The experimental results verify that the reconstruction method DSFSP proposed in this thesis can effectively enhance the gradient spatial and contextualized information,improve image clarity,restore high-frequency details and drop artifacts.And the weakly supervised learning method used in this thesis can improve the versatility and reduce training complexity.