Enhancing Computational Image Reconstruction and Denoising: A Deep Learning Framework for Advancing Ghost Imaging and Non-Line-of-Sight Imaging

Author

• Shima Tabakhi, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2026

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Sayeh, Mohammad

Abstract

Ghost imaging has attracted considerable attention as a non-conventional optical sensing paradigm capable of reconstructing spatially resolved images from single-pixel intensity correlations, without requiring a lens or a spatially resolving detector in the object arm. Its extension to Non-Line-of-Sight (NLOS) imaging, where hidden objects are reconstructed from diffuse reflections off relay surfaces, further expands the operational envelope of the technique to scenarios inaccessible to conventional direct-view systems. Despite these theoretical advantages, the practical deployment of Ghost-NLOS imaging remains severely constrained by two interrelated challenges: the degradation of second-order coherence under multi-layer scattering, which drives the correlation signal toward noise floor, and the absence of reconstruction algorithms capable of maintaining fidelity under photon-starved, high-speckle-contrast conditions. Purely data-driven approaches to this inverse problem are fundamentally limited by their inability to exploit the known wave-optical structure of the measurement process, resulting in performance collapse under the most severe degradation regimes.This dissertation presents two complementary contributions to computational Ghost-NLOS reconstruction. The first is the Enhanced Attention U-Net (EAU-Net), a data-driven encoder-decoder architecture incorporating multi-scale spatial attention mechanisms, sinusoidal activations, and a composite reconstruction loss, designed to suppress structured speckle noise and recover fine spatial detail under moderate scattering conditions. The second, and primary, contribution is the Physics-Guided U-Net (PGU-Net), in which Fresnel diffraction, Beer--Lambert scattering, mirror reflection, and thin-lens transformation are implemented as differentiable, jointly optimised layers embedded within the reconstruction backbone. A physics consistency loss penalises reconstructions that, when re-degraded by the learned forward model, fail to reproduce the observed measurement, introducing a cycle-consistency constraint rooted in wave optics rather than image statistics. Both architectures are evaluated under a three-stage sequential transfer learning protocol across three experimentally acquired noise levels of increasing severity (L1, L2, L3), spanning input conditions from mild speckle to severe multi-layer scattering below 15 dB input PSNR.Experimental results demonstrate that the PGU-Net achieves a strictly monotonically decreasing performance profile across the noise hierarchy (PSNR: 32.82 to 32.55 to 31.38 dB; SSIM: 0.9283 to 0.9183 to 0.9001), in contrast to the non-monotonic profile of the EAU-Net baseline, which is identified as a consequence of dataset composition artefacts and cold-start training instability. At the most severe noise condition (L3), the PGU-Net outperforms the baseline by 3.98 dB in PSNR and 3.72 dB in contrast-to-noise ratio, with no instability events across any stage of training. These results establish that embedding a differentiable optical forward model into a reconstruction architecture constitutes a principled and practically superior approach to Ghost-NLOS imaging under severe scattering, one that generalizes the physics-informed reconstruction paradigm demonstrated across optical coherence tomography, fiber endoscopy, and low-dose medical imaging to the correlation-based sensing domain.