AIRI denoiser shelves for PnP algorithms in high-dynamic range astronomical imaging

The dataset consists of non-expansive denoiser DNNs, underpinning the Plug-and-Play (PnP) algorithm AIRI for high-dynamic range astronomical imaging.

Two denoiser shelves are available, each composed of 8 DNNs, trained for the removal of random white Gaussian noise, with mean zero and a fixed standard deviation in the list {3.2E-4, 1.6E-4, 8E-5, 4E-5, 2E-5, 1E-5, 5E-6, 2.5E-6}:

1) airi_astro-based_oaid_shelf: trained from optical astronomical images (NOIRLab/NSF/AURA/H.Schweiker/WIYN/T.A.Rector (University of Alaska Anchorage));
2) airi_mri-based_mrid_shelf: trained from medical images (NYU fastMRI Initiative database; Knoll et al. 2020).

Two sets of denoisers are provided for uncertainty quantification of AIRI, each composed of 14 realisations of DNN denoisers trained for the removal of random white Gaussian noise, with mean zero and standard deviation 1.6E-4:

1) airi_astro-based_oaid_uncertainty_quantification: trained from optical astronomical images;
2) airi_astro-based_mrid_uncertainty_quantification: trained from medical images.

All DNNs are available in ONNX format. They can be deployed for arbitrary image sizes, subject to hardware memory constraints.