Untrained networks for compressive lensless photography

Compressive lensless imagers enable novel applications in an extremely compact device, requiring only a phase or amplitude mask placed close to the sensor. They have been demonstrated for 2D and 3D microscopy, single-shot video, hyperspectral imaging; each of these cases, compressive-sensing-based inverse problem is solved order recover data-cube from measurement. Typically, this accomplished using convex optimization hand-picked priors. Alternatively, deep learning-based reconstruction methods offer promise better priors, but require many thousands ground truth training pairs, which can be difficult impossible acquire. In work, we propose use untrained networks compressive image recovery. Our approach does not any labeled data, instead uses measurement itself update network weights. We demonstrate our on imaging as well high-speed video recovery camera's rolling shutter, imaging. provide simulation experimental verification, showing that method results improved quality over existing methods.