Data-driven phase retrieval using deep generative models

Date

2024-6-25

Author

Kaya, Mehmet Onurcan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

27
views

31
downloads

Cite This

This thesis addresses the nonlinear inverse problem of phase retrieval, which is the process of recovering a signal from the magnitude of its Fourier transform, a fundamental challenge in fields such as electron microscopy, crystallography, astronomy, and optical imaging. Classical phase retrieval techniques face limitations in robustness, noise sensitivity, and computational efficiency. To overcome these limitations, this work develops novel data-driven phase retrieval methods by exploiting advanced deep generative models. Firstly, we present a phase retrieval approach leveraging Langevin dynamics within diffusion models. This approach utilizes two different deep learning pipelines, namely prNet-Small and prNet-Large, and carefully balances the perceptual quality-distortion tradeoff. While we favor minimal distortion, we also aim to create high-perceptual quality images. Secondly, we use the Inversion by Direct Denoising (InDI) framework to solve the Fourier phase retrieval problem. The developed method also employs advanced initialization strategies and ensembling techniques, resulting in improved training efficiency and better image quality compared to traditional methods. Thirdly, we extend the Denoising Diffusion Restoration Models (DDRM) for phase retrieval by combining with the Hybrid Input-Output (HIO) method. This approach utilizes pretrained unconditional diffusion models. Overall, this thesis demonstrates that exploiting the score/diffusion-based framework significantly improves the solution of the phase retrieval problem by enabling unprecedented image quality, better noise robustness, and higher computational speed and efficiency. These advancements have a broad impact on computational imaging and various scientific and engineering applications.

Subject Keywords

Phase Retrieval, Diffusion, Generative Models, Deep Learning, Image Reconstruction

URI

https://hdl.handle.net/11511/110069

Collections

Graduate School of Natural and Applied Sciences, Thesis