An S-CNN-based phase conjugation method in imaging through random media

Abstract

Imaging through random media is a pervasive issue in many areas. Recently, deep learning (DL) has shown strong capability in speckle reconstruction, especially in extracting the correlation properties of speckle patterns. However, in practical imaging scenarios, most DL methods need numerous training data to adapt to various imaging condition changes. Here, we develop an S-CNN-based phase conjugation method in imaging through random media. Specifically, we propose the amplitude-phase distortion model to characterize the light scattering process. Then, we further develop a scattering convolutional neural network (S-CNN) based on our model to map speckle patterns to scattered light fields in the changed imaging conditions. Finally, we reconstruct the object through phase conjugation. We quantitatively validated the imaging performance of S-CNN-based phase conjugation imaging, meeting the possible changes of practical imaging conditions, encompassing unseen types of different sparse objects, different diffusers, different detection positions, and different apertures on the imaging path. We found that the Jaccard index (JI), Pearson correlation coefficient (PCC), and structural similarity measure (SSIM) in different physical scenarios were increased by more than 21%, 22%, and 8%, compared with digital CNN. The proposed method provides a new optical imaging system that can be adapted to accommodate various condition changes in practical imaging scenarios and paves the way for an all-optical imaging system in imaging through random media.