BlindNet: An untrained learning approach toward computational imaging with model uncertainty

Abstract

The solution of an inverse problem in computational imaging (CI) often requires the knowledge of the physical model and/or the object. However, in many practical applications, the physical model may not be accurately characterized, leading to model uncertainty that affects the quality of the reconstructed image. Here, we propose a novel untrained learning approach towards CI with model uncertainty, and demonstrate it in phase retrieval, an important CI task that is widely encountered in biomedical imaging and industrial inspection.