Super-resolution imaging through the diffuser in the near-infrared via physically-based learning

The optical scattering system illuminated by the near-infrared (NIR) light suffers more severe resolution loss than systems in the visible light range due to the limitations of the detector or the diffraction at longer wavelengths. Conventional imaging methods cannot perform super-resolution (SR) imaging by using the single-frame low-resolution (LR) speckle pattern. Here, based on the mechanism of the signal degradation we propose a physically-based learning method to achieve SR reconstruction. We prove that the resolution lost in the detector or the inappropriate optical parameters of the NIR scattering system can be compensated, based on which our method successfully achieves SR reconstruction with only one frame of the LR speckle pattern. Specifically, we investigate the degradation process when the detector observes the signal and discuss the optical system parameters that affect the imaging resolution. Besides, the redundancy of the speckle pattern provides a physical-basis that allows deep-learning (DL) to mine features of the single-frame sub-speckle pattern and supplement the lost information. This method significantly reduces requirements for the detector and provides a reference for SR imaging in the NIR scattering system.