Neural blind deconvolution to reconstruct high-resolution ground-based solar observations

Abstract

Ground-based solar observations enable unprecedented spatial, spectral, and temporal resolution of the lower solar atmosphere, yet Earths turbulent atmosphere imposes significant limitations, requiring advanced post-facto image reconstruction. State-of-the-art reconstruction methods are based on restoring a burst of short exposure frames to a single observation. Limitations of these techniques arise due to the sparse information about the atmospheric point spread function (PSF) that degrade the observations and consequently the quality of reconstructions. We develop a novel image reconstruction method to achieve unprecedented spatial resolution from short exposure image bursts. This can provide high-quality reconstructions and therefore advance the study of the smallest spatial scales from the solar photosphere to the chromosphere. In this study, we present a novel approach for high-resolution solar image reconstruction based on physics-informed neural networks. In the training process, the neural network maps coordinate points directly to their corresponding intensity values while simultaneously updating the PSF parameters. The method convolves the true object from the neural network with the estimated PSFs and optimizes the network by minimizing the loss between the synthesized and real short-exposure image burst. This approach enables the simultaneous estimation of both the degrading PSF and the real high-resolution intensity distribution. We demonstrate the method on synthetic intensity data derived from a radiative MHD simulation and apply it to high-resolution observations from GREGOR and DKIST. Our results demonstrate the ability to reconstruct small-scale solar features that exceed the reconstruction performance of state-of-the-art reconstruction methods. With this approach we lay the foundation for future spatially varying PSFs.

Figures (9)

• Figure 1: a) Comparison of a single frame of the burst from the degraded simulation (left), the NeuralBD reconstruction (center) and the real simulation (right). b) shows the histogram of the normalized intensity for the convolved frame (green), the NeuralBD reconstruction (red) and the real simulation (blue).
• Figure 2: Comparison of the degraded convolved frame (left) the NeuralBD reconstruction (center) and the real simulation (right) for quiet Sun and penumbra and umbra with penumbra. The bottom panel shows the corresponding power spectra density (PSD) for the convolved frame (green), the NeuralBD reconstruction (red) and the real simulation (black).
• Figure 3: Comparison of five PSFs initialized to degrade the synthetic MURaM simulation (top panel) and the PSFs estimation by our NeuralBD model (bottom panel).
• Figure 4: Comparison between the baseline method (Richardson-Lucy), the torchmfbd method and our NeuralBD reconstruction for MURaM simulation data. In panel a) we show a visual comparison of a single convolved frame, the MURaM simulation, the NeuralBD reconstruction and the Richardson-Lucy deconvolution and torchmfbd. Panel b) shows the difference maps of our NeuralBD method, the baseline method and torchmfbd. Panel c) shows the 2d histograms comparing NeuralBD with MURaM (left) and Richardson-Lucy with MURaM (center) and torchmfbd with MURaM (right). In panel d) the power spectral density (PSD) for all five observations in a) are shown.
• Figure 5: Comparison of the performance of the NeuralBD reconstuction method with the speckle reconstruction and torchmfbd for g-band at 430.7 nm (left) and blue continuum at 450.6 nm (right). a) Single frame from the original burst (first column), the speckle reconstruction (second column), the torchmfbd reconstruction (third column) and the NeuralBD reconstruction (fourth column), for the GREGOR observation on June 2, 2022. The red and blue rectangles indicate the regions shown as cropped views in the second and third rows, respectively. b) Corresponding azimuthal power spectra for both channels are shown. The single frame of the original burst (green), the speckle reconstruction (black), the torchmfbd reconstruction (brown) and the NeuralBD reconstruction (red).