By Dr. Luca Calatroni, LABORATOIRE D'INFORMATIQUE, SIGNAUX ET SYSTÈMES DE SOPHIA-ANTIPOLIS, FRANCE
Fecha seminario: 2023-07-06
In this talk, I will focus on how to combine model-based techniques with data-driven learning approaches to define suitable regularization and data fitting terms for reconstructing high-resolution images from sequences of low-resolution and stochastically fluctuating blurred data. In particular, I will show how convergent plug & play gradient-step denoisers can be effectively used to define desirable prior functionals and explain how Generative Adversarial Networks (GANs) can be combined with physical knowledge to define a suitable data term comparing the (unknown) distribution of measured data with the one simulated within the adversarial process.