On the use of deep learning for phase recovery (2308.00942v1)

Abstract: Phase recovery (PR) refers to calculating the phase of the light field from its intensity measurements. As exemplified from quantitative phase imaging and coherent diffraction imaging to adaptive optics, PR is essential for reconstructing the refractive index distribution or topography of an object and correcting the aberration of an imaging system. In recent years, deep learning (DL), often implemented through deep neural networks, has provided unprecedented support for computational imaging, leading to more efficient solutions for various PR problems. In this review, we first briefly introduce conventional methods for PR. Then, we review how DL provides support for PR from the following three stages, namely, pre-processing, in-processing, and post-processing. We also review how DL is used in phase image processing. Finally, we summarize the work in DL for PR and outlook on how to better use DL to improve the reliability and efficiency in PR. Furthermore, we present a live-updating resource (https://github.com/kqwang/phase-recovery) for readers to learn more about PR.

• The paper demonstrates that deep learning efficiently recovers optical phase images, outperforming conventional methods in speed and accuracy.
• It details the use of DL in pre-, in-, and post-processing stages to enhance hologram resolution and reduce noise with CNNs and GANs.
• The paper suggests that physics-informed neural networks could further revolutionize real-time quantitative phase imaging by lowering computational demands.

Deep Learning for Phase Recovery: A Comprehensive Review

Phase recovery (PR) is a critical aspect of optical imaging pivotal for reconstructing the refractive index distribution or topography of objects. Conventional methods such as interferometry, ptychography, and wavefront sensing have been the mainstay for PR solutions. However, these methods often involve intricate experimental setups or iterative algorithms which can be cumbersome and time-consuming. This paper explores the innovative use of deep learning (DL), particularly deep neural networks (DNNs), as a tool to enhance efficiency and reliability in PR tasks.

Conventional Phase Recovery Methods

Conventional PR methods include holography and transport of intensity equations (TIE). Holography can be performed using inline and off-axis configurations, often requiring several holograms and extensive computational processes like phase-unwrapping to obtain high-resolution and accurate phase images. While powerful, these conventional techniques can face challenges with noise, aliasing, and resolution limits.

The Role of Deep Learning in Phase Recovery

Emerging as a transformative technology, deep learning augments computational imaging by tackling these challenges in PR across three stages: pre-processing, in-processing, and post-processing.

1. DL-Pre-Processing for PR: Pre-processing with DL involves tasks like pixel super-resolution, noise reduction, and hologram generation. These tasks aim to enhance holograms by improving resolution and reducing noise before conventional phase recovery processes begin. For instance, convolutional neural networks (CNNs) like U-Net have been employed for hologram super-resolution, significantly boosting resolution without increased computational time.
2. DL-In-Processing for PR: In the core phase recovery process, DL can be employed solely or in conjunction with physics-based models. The network-only strategy involves training neural networks to directly infer phases from intensity measurements, driven by vast datasets that encode the intensity-to-phase mapping. Alternatively, merging physics principles with network training enhances the robustness and adaptability of DNNs, combating the inherent ill-posed nature of PR, especially under low-light conditions.
3. DL-Post-Processing for PR: Post-processing using DL addresses residual noise, refines spatial resolution, and corrects aberrations in the phase-derived images. Techniques like using GANs for noise reduction or applying supervised learning for resolution enhancement are prominent here.

Strong Numerical Results and Claims

The paper presents robust evidence of DL significantly outperforming traditional methods in efficiency and versatility. Performance is evaluated through metrics like computation time savings and accuracy improvements in phase retrieval tasks. Notably, DL approaches have shown exemplary results in applications requiring real-time processing and high-resolution outputs.

Implications and Future Directions

The integration of DL into the domain of PR not only streamlines existing processes but also opens avenues for new applications in quantitative phase imaging, adaptive optics, and beyond. Practically, the reduction in computational complexity and resource requirements offers pathways for DL models to be used in real-time imaging systems. Theoretically, the paper suggests that continued exploration into physics-informed neural networks and unsupervised learning paradigms could yield further advancements.

Looking forward, the embedding of deep learning within optical systems, potentially leveraging hardware acceleration through optical neural networks, could transform phase imaging paradigmatically. More research into informed dataset creation, model interpretability, and training efficiency will further solidify DL's role in advancing optical sciences.

This comprehensive review highlights deep learning's potential to redefine phase recovery, fostering future innovations rooted in analytic and data-driven methodologies alike.