Senior Computer Vision Engineer at GoPro
June 22nd, 2023
Neural networks allow solving many ill-posed inverse problems with unprecedented performance. Physics informed approaches already progressively replace carefully hand-crafted reconstruction algorithms in real applications. However, these networks suffer from a major flaw: when trained on a given forward operator, they do not generalize well to a different one. The aim of this presentation is threefold. First, we provide an overview of the different physics-informed reconstruction approaches that exist. Then, we show through various applications that training the network with a family of forward operators allows solving the adaptivity problem without compromising the reconstruction quality significantly. Finally, we illustrate that this training procedure allows tackling challenging blind inverse problems. The experiments include partial Fourier sampling problems arising in magnetic resonance imaging (MRI), computerized tomography (CT) and image deblurring. This presentation is based on a work done during my PhD at the Toulouse Mathematics Institute.