Imaging through opaque scattering layers via transmission matrix assisted learning

Abstract

Deep learning (DL) has a wide application in imaging through scattering media, however, most DL approaches lack related physical principle priors. Aiming at the limitation of DL methods that require high completeness of training set, a two-stage network is proposed to complete the transmission matrix (TM) measurement and image reconstruction. Thanks to the appropriate structure of the network, the amount of data required in the Measurement Stage is greatly reduced. The self-closed-loop constraint in the Imaging Stage also enables the imaging network to break from the dependence on the completeness of the training set, and achieve a reconstruction with an SSIM of 0.84 using only 10 pairs of training data. Besides, both the Imaging Stage and the Measurement Stage can be used as a stand-alone method in combination with conventional phase retrieval algorithms. This method can drive the development of TM-based imaging and provide an enlightening reference for the practical application in optical imaging scenes.