University of Crete (UOC)
Deep learning as a tool to decode optical transmission through complex scattering media
Abstract
dc:descriptionImaging and light control through complex scattering media remain fundamental challenges in modern optics, with significant implications for biomedical imaging, optical communications, and sensing technologies. When light propagates through disordered environments such as biological tissues or multimode optical fibers, multiple scattering events scramble the phase and amplitude of the optical field, giving rise to seemingly random speckle patterns. Although these patterns appear stochastic, they retain deterministic information about both the incident wavefront and the properties of the medium. Extracting this information, however, constitutes a highly complex inverse problem that is often ill-posed and difficult to solve using conventional techniques. In recent years, deep learning has emerged as a powerful tool for addressing such inverse problems, offering the ability to learn complex, nonlinear mappings directly from data without requiring explicit physical models. This thesis explores the application of deep neural networks, particularly convolutional neural networks, for imaging and light reconstruction in complex optical environments. The primary objective is to investigate how information encoded in scattered light can be exploited to infer otherwise inaccessible transmission properties, and how data-driven approaches can be combined with traditional optical methods to enhance imaging performance. The theoretical part of the thesis is organized to support these experimental and numerical studies. Chapter 2 establishes the wave-optical basis of the work, including scalar and vector propagation, diffraction, coherence, Fourier optics, random-media propagation, speckle formation, digital holography, optical phase conjugation, and numerical propagation methods. Chapter 3 then specializes this framework to guided-wave systems, emphasizing modal confinement, modal interference, multimode-fiber transfer operators, fiber-endoscopic architectures, wavefront shaping, transmission-matrix calibration, wavelength-diversity reconstruction, and non-Hermitian gain-loss waveguides with -symmetry. Chapter 4 introduces deep neural networks as statistical estimators for computational imaging, with attention to loss functions, U-Net-type architectures, physics-informed reconstruction, and the generalisation limits that arise when the optical transfer operator changes. The core contributions of this thesis are threefold. First, a data-driven framework is developed to predict transmission speckle patterns from reflection measurements in scattering media. Through experimental data acquisition and numerical modelling, the framework is evaluated across a range of scattering-medium thicknesses, demonstrating that neural networks can learn the statistical relationship between reflected and transmitted light over varying propagation regimes, including regimes where conventional approaches based on the optical memory effect become ineffective. These results establish the feasibility of non-invasive imaging approaches that infer transmitted optical information solely from reflection measurements.Second, a hybrid imaging methodology is proposed that combines wavelength-scanning ptychography with deep-learning-based fiber characterization and distortion compensation for multimode-fiber endoscopy. Wavelength scanning provides the measurement diversity required for ptychographic phase retrieval while eliminating the need for mechanical scanning at the distal end of the endoscopic system. Deep learning is incorporated to model and compensate for the complex optical transformations induced by multimode-fiber propagation, enabling recovery of the distorted optical information required for subsequent ptychographic reconstruction. By combining the physical constraints of wavelength-scanned ptychography with data-driven modeling of fiber-induced distortions, the approach provides a robust framework for high-fidelity endoscopic imaging through multimode fibers under noise and system imperfections. Third, the thesis extends deep learning to non-Hermitian photonic systems based on parity-time (PT)-symmetric waveguide structures operating near exceptional points, where small refractive-index perturbations produce strongly enhanced optical responses. Neural networks are trained to infer both the magnitude and spatial position of refractive-index perturbations from one-dimensional output intensity measurements, demonstrating robust reconstruction even in the presence of noise. Importantly, this capability extends the application beyond perturbation sensing toward optical imaging, enabling the reconstruction of two-dimensional binary perturbation maps from one-dimensional intensity measurements. The enhanced response of the PT-symmetric system near the exceptional point provides a clear advantage over its Hermitian counterpart, highlighting the potential of non-Hermitian photonic architectures for deep-learning-assisted optical sensing and imaging. Overall, this thesis demonstrates that deep learning provides a versatile and powerful framework for addressing complex imaging problems in optics. By enabling the extraction of meaningful information from highly scattered light and by enhancing traditional reconstruction techniques, the proposed approaches open new pathways for non-invasive imaging, fiber-based endoscopy, and optical sensing in challenging environments. Future work may focus on improving model generalisation, reducing data requirements, and extending these methods to dynamic and real-time imaging scenarios.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Skarsoulis, Kyriakos
- Σκαρσουλής, Κυριακός
Subjects
dc:subject × 23- Οπτική απεικόνιση
- Νευρωνικά δίκτυα
- Υπολογιστική απεικόνιση
- Οπτική ίνα
- Ενδοσκόπηση
- Μη ερμιτιανά συστήματα
- PT-συμμετρία
- Optical imaging
- Artificial neural networks
- Computational imaging
- Endoscopy
- Non-hermitian systems
- PT-symmetry
- Φυσική
- Φυσικές Επιστήμες
- Οπτική
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Τεχνητή νοημοσύνη
- Physical Sciences
- Natural Sciences
- Optics
- Computer and Information Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62503