A blind image-restoration technique for space-based solar telescopes that jointly estimates the instrument's point-spread function and the underlying true image directly from observations, without prior knowledge of either. A convolutional neural network transforms co-registered reference-channel images (e.g. SDO/AIA 304A) into ideal images that obey the optics' linear-convolution model, while the wavefront aberration of the system's generalized pupil function is optimized so the ideal image convolved with the derived PSF matches the observed data; the recovered PSF then drives Richardson-Lucy deconvolution to sharpen the imagery.
The Solar Disk Imager (SDI) aboard the ASO-S mission's Lyman-alpha Solar Telescope achieved an in-flight spatial resolution (~9.5 arcsec) far worse than its 1.2 arcsec design, owing to wavefront erro…