Deep-Learning-Based Rapid Imaging Through Scattering Media Beyond the Memory Effect

Abstract

By incorporating the ptychography with the shower-curtain effect (PSE), a large field-of-view object hidden behind the scattering media can be reconstructed from multiple diffused patterns. However, the original PSE method is subjected to low speed data acquisition and time-consuming image reconstruction because of the mechanical scanning scheme and the iterative retrieval algorithm. Here, a deep-learning method based on the PSE with improved optical scheme is proposed to accelerate the data acquisition and image reconstruction speed. By replacing the mechanical translation stage with the digital micromirror device (DMD), it facilitates a large number of data collection for training the network. Single-shot pattern and sub-second reconstruction for the well-trained network model make the method appropriate for rapid imaging. Both qualitative presentation and quantitative analysis for binary resolution target and 2D biological slide specimens demonstrate the effectiveness and feasibility of the proposed method, thereby offering a prospective application in tissue imaging.