Application and influencing factors analysis of Pix2pix network in scattering imaging

Abstract

The imaging accuracy of deep learning-based scattering imaging techniques depends largely on the network structure and the speckle data quality. Up to now, many schemes based on deep learning to achieve imaging through single-layer scattering medium have been proposed. However, the performance of these schemes is limited when the scattering medium is a thick multilayer or dynamic medium. At the same time, the influence of complex changes in scattering environment on the quality of speckle data is obscured. In this study, a scheme of Pix2pix network based on the Peak Signal-to-Noise Ratio (PSNR) loss function is proposed to reconstruct the images passing through dynamic and double-layer scattering media. The influence of physical factors such as light intensity, dynamic perturbations of scattering medium, and optical depth of scattering medium on network imaging are quantitatively analyzed. In order to analyze the influence of these factors on network imaging more objectively, a typical Dense-unet is also used to train. In the experiment, the imaging results of both networks exhibit the same varying trend with varying physical factors. In addition, the proposed Pix2pix network displays better performance compared to Dense-unet. This work is helpful to future imaging studies based on machine learning.