Litcius/Paper detail

High-quality ghost imaging through highly complex scattering media with physics-enhanced untrained neural networks

Tianshun Zhang, Yang Peng, Wen Chen

2025Optics Express8 citationsDOIOpen Access PDF

Abstract

Optical imaging through complex media remains a challenge when illumination and detection paths are simultaneously disturbed. In this paper, we report an untrained neural network (UNN) enhanced by a physical model of ghost imaging (GI) to address complex-scattering-induced beam distortions and achieve high-quality object reconstruction. The experimental configuration consists of rotating ground glass (RGG) diffusers placed in front of and behind an object, coupled with a turbidity-varying liquid turbulence chamber in the optical path. Our analysis reveals that a series of dynamic scaling factors critically degrade the performance of GI. To overcome this challenge, speckle patterns induced by complex and dynamic scattering are recorded via the design of a reference beam arm, and a series of single-pixel intensities are collected in the object beam arm. A physics-enhanced UNN is designed and implemented to estimate a series of scaling factors, and a GI formation model is integrated into UNN to ensure the validity of corrected measurements and enable robust reconstruction. Experimental results demonstrate that the proposed method can achieve robust and high-quality object reconstruction through complex scattering media where illumination and detection paths are simultaneously disturbed. The proposed method can open an avenue for overcoming optical scattering challenges in complex scenarios.

Topics & Concepts

OpticsGhost imagingScatteringImage qualityQuality (philosophy)PhysicsLight scatteringArtificial neural networkForward scatterComputer scienceArtificial intelligenceImage (mathematics)Quantum mechanicsRandom lasers and scattering mediaAdvanced Optical Imaging TechnologiesOrbital Angular Momentum in Optics
High-quality ghost imaging through highly complex scattering media with physics-enhanced untrained neural networks | Litcius