Revisiting Cygnus A with the R2D2 deep neural network series

Reconstruction of results of R2D2, a novel AI model for radio interferometric imaging, applied to real S-band observations of the celebrated radio galaxy Cygnus A with the Very Large Array. The dataset consists of monochromatic images (estimated model images and residual dirty images) obtained by two proposed incarnations of the R2D2 model, and benchmark algorithms. R2D2 models: (1) R2D2 series, underpinned by U-Net blocks, (2) R3D3 series, underpinned by novel R2D2-Net blocks. Benchmark algorithms: (1) U-Net, end-to-end DNN (also first term of R2D2 series), (2) R2D2-Net: proposed by the authors, end-to-end DNN (also first term of R3D3 series), (3) AIRI, PnP algorithm for RI (Terris et al. 2023), (4) uSARA, sparsity-based algorithm for RI (Repetti & Wiaux 2020, Terris et al. 2023), (5) CLEAN, multiscale (Offringa et al. 2014). Cygnus A images obtained with a joint calibration and imaging framework (Repetti et al, 2017) for direction dependent calibration and using AIRI and uSARA as imaging modules, are provided for reference: (6) AIRI_DDE-CALIM calibration, (7) uSARA_DDE-CALIM calibration.