The combination of advanced machine/deep learning methods with model-based physical modelling has proven to be key in designing interpretable and stable variational image reconstruction methods. Upon tailored neural parameterisations of specific model/algorithmic components (e.g., image regularisers, algorithmic parameters and geometry, just to name a few) can be used to derive convergence and stability properties of the underlying reconstruction procedure, in contrast to end-to-end approaches. By further incorporating additional knowledge of both the statistical and geometrical properties of the problems at hand, as well as invariances in the desired solution, such model-aware approaches can also be applied in weakly-supervised or self-supervised settings. This is particularly valuable in real-world applications with limited training data.

Luca Calatroni DIBRIS, Università di Genova

Kostas Papafitsoros QMLU, UK
Samuel Vaiter CNRS, France