R2D2: a deep neural network series paradigm for robust ultra-fast precision imaging in radio astronomy

Mapping the radio sky with the new modern radio-interferometric (RI) arrays requires solving challenging inverse problems for the formation of high-resolution high-dynamic range images from large volumes of visibility data. A new generation of image reconstruction algorithms grounded in optimisation theory have demonstrated remarkable capability for imaging precision, well beyond the capability offered by CLEAN. These range from advanced proximal algorithms propelled by handcrafted regularisation operators, such as the SARA family, to hybrid plug-and-play (PnP) algorithms propelled by learned regularisation denoisers, such as AIRI. While already capable of scaling to large image and data dimensions, these techniques are highly iterative, which still hinders their ability to handle the extreme data volumes expected from future instruments. To address this scalability challenge, a novel deep learning approach was recently proposed, dubbed “Residual-to-Residual DNN series for high-Dynamic range imaging”. R2D2’s reconstruction is formed as a series of residual images, iteratively estimated as outputs of DNNs taking the image estimate from the previous iteration and the associated data residual as inputs. R2D2 thus features a hybrid structure between a PnP algorithm and a learned version of the matching pursuit algorithm underpinning CLEAN. We will dive into the R2D2 algorithmic structure and discuss its validation in simulation and on real data, showing that it opens the door to robust ultra-fast precision RI imaging.