The Atmospheric Imaging Assembly (AIA) on board the Solar Dynamics Observatory (SDO) (launched in February 2010) provides uninterrupted full-disk solar images over 10 wavebands. In the case of violent solar flares, saturation would happen to SDO/AIA images in their core regions, which leads to signal loss, hindering us to understand physical mechanism behind solar flares. This paper introduces a deep learning based image restoration model which can recover signal of saturation region by referring to other normal/valid region within an image. The proposed model, namely PCGAN, combines partial convolution (PC) and conditional generative adversarial network (GAN). The PC module was originally designed for image inpainting, for repairing images with scratches and holes. In addition, a new comprehensive loss function consists of an adversarial loss, a pixel reconstruction loss, a gradient loss, a perceptual loss, a style loss and a total variation loss. Moreover, for validating the proposed model, a new dataset consisting of paired saturated and normal SDO/AIA images is established. Experimental results demonstrate that the proposed PCGAN can get appealing desaturated solar images with respect to both objective and subjective validations.

太阳动力学天文台（SDO）上的大气成像组件（AIA）（于2010年2月推出）在10个波段上提供不间断的全盘太阳图像。在强烈的太阳耀斑的情况下，SDO / AIA图像在其核心区域会发生饱和，这会导致信号丢失，从而使我们无法理解太阳耀斑背后的物理机制。本文介绍了一种基于深度学习的图像恢复模型，该模型可以通过参考图像中的其他正常/有效区域来恢复饱和区域的信号。提出的模型PCGAN结合了部分卷积（PC）和条件生成对抗网络（GAN）。PC模块最初设计用于图像修复，用于修复带有划痕和孔洞的图像。此外，新的综合损失功能包括对抗性损失，像素重建损失，梯度损失，感知损失，样式损失和总变化损失。此外，为验证所提出的模型，建立了一个新的数据集，该数据集由成对的饱和SDO / AIA图像和正常图像组成。实验结果表明，所提出的PCGAN可以在主观和主观验证方面获得吸引人的去饱和太阳图像。