Deep plug-and-play and unrolling approaches for phase retrieval

Date

2025-8-28

Author

Külbay, Deniz

Metadata

Show full item record

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

Item Usage Stats

347
views

0
downloads

Cite This

Image reconstruction (IR) tasks, as encountered in imaging, require estimating an unknown image from available physical observations. For example, blurred and noisy observations of an unknown image are available for the image deconvolution task, and noisy magnitude measurements are available for the phase retrieval task. Existing IR methods are either purely analytical methods or incorporate learning-based models. While purely analytical methods simultaneously exploit the underlying physics of the problem and hand-crafted image priors for regularization, deep learning enables to learn the image priors directly from the data. An important such example is deep learning-based plug-and-play (PnP) regularization approach, which has gained popularity because of its superior performance. These approaches iteratively use analytical and learning-based steps to obtain a solution. Although they are able to leverage the strengths of analytical and learning-based methods, their performance is highly dependent on hyperparameter selection and pre-trained model used. In this study, we first focus on analyzing and improving the performance of some existing PnP methods for image deconvolution and phase retrieval through better hyperparameter selection and integration of deeper pre-trained network models. Based on the PnP method, we then develop an end-to-end trainable approach for phase retrieval using algorithm unrolling, which is a technique that transforms an iterative algorithm into a deep neural network by mapping each iteration to a network layer. This enables to learn all hyperparameters through end-to-end training, enhancing both performance and interpretability. Through simulations, the performance of the developed unrolling-based approach is comparatively evaluated against the PnP method both qualitatively and quantitatively.

Subject Keywords

Image reconstruction, Deep learning, Plug-and-play regularization, Phase retrieval, Algorithm unrolling

URI

https://hdl.handle.net/11511/116071

Collections

Graduate School of Natural and Applied Sciences, Thesis