Scientific organisers

• Julie Delon -- Université Paris Cité, France

• Audrey Repetti -- Heriot-Watt University, Edinburgh, UK

• Carola-Bibiane Schoenlieb -- University of Cambridge, UK

• Gabriele Steidl -- TU Berlin, Germany

About:

In the last few years, an important trend has emerged for using data-driven image models, in particular encoded by neural networks. Novel families of hybrid imaging methodologies, mixing data-driven and traditional mathematical approaches (such as optimisation or sampling methods) have flourished. For instance, generative or discriminative networks such as GANS, VAEs or normalising flows, can be either used in optimisation or sampling schemes as data-driven regularisers for solving inverse problems. Similarly, denoising networks or more generally regularising networks can be incorporated into optimisation or sampling schemes leading to Plug-and-Play methods. From another perspective, unrolled optimisation approaches have been investigated to provide robust network architectures as alternative to traditional black-box end-to-end networks. 

All these approaches have shown a remarkable versatility and efficiency to solve inverse imaging problems. They hence pave the way to restoration algorithms that exploit more powerful and accurate models for images than traditional hand-crafted models (e.g., sparsity-promoting regularisations).  Nevertheless, they also come with important mathematical challenges regarding the robustness of the delivered solution e.g. in terms of interpretability and uncertainty quantification on the solution. Further, the environmental impact of using data-driven approaches is gaining increasing attention from the imaging community, leading to the need for frugal (but efficient) architectures.

In this workshop, three main research areas will be explored, both from a theoretical and an application viewpoint:

1. Generative models for imaging inverse problems

2. Robustness and convergence of data-driven regularisers

3. Computational complexity challenges