R2D2 takes a hybrid structure between a Plug-and-Play (PnP) algorithm and a learned version of the well-know "Matching Pursuit" algorithm. Its reconstruction is formed as a series of residual images, iteratively estimated as outputs of Deep Neural Networks (DNNs) taking the previous iteration’s image estimate and associated data residual as inputs. R2D2's primary application is to solve large-scale high-resolution high-dynamic range inverse problems for RI in radio astronomy, more specifically 2D planar monochromatic intensity imaging. R2D2 provides significant acceleration of the image reconstruction process over highly iterative PnP and proximal (optimisation) algorithms, while maintaining imaging precision. It thus opens the door to ultra-fast precision imaging.
As of early 2026, several active research directions are expanding the scope and capabilities of R2D2 at BASP. These include scaling R2D2 to handle increasingly large and complex datasets, such as large-scale and 3D wideband radio interferometric imaging, and transferring the methodology to challenging medical imaging modalities including MRI and ultrasound. Ongoing developments also extend R2D2 to spherical imaging (S-R2D2) and explore its generalisation into a cross-domain foundation model for inverse problems across multiple sensing modalities. Stay tuned!
