Frequency-domain learning-driven lightweight phase recovery method for in-line holography.

Area of Science:

• Optical Imaging
• Computational Optics
• Deep Learning Applications

Background:

• Phase retrieval from single intensity images is an ill-posed problem in optical imaging.
• Traditional methods struggle with dynamic scenes and consistent quality.
• Deep learning offers solutions but often requires complex architectures.

Purpose of the Study:

• To develop a lightweight, efficient phase recovery method for optical imaging.
• To address the computational complexity of existing deep learning approaches.
• To improve phase reconstruction in challenging scenarios like in-line holography.

Main Methods:

• Proposed a frequency-domain learning-driven lightweight phase recovery method (FNet) using complex-valued networks.
• Designed models with fewer parameters by analyzing optical diffraction in the frequency domain.
• Incorporated complex-valued total variation regularization for enhanced reconstruction.

Main Results:

• FNet achieved performance comparable to conventional and real-valued methods.
• Demonstrated significantly fewer parameters and reduced computational resource demands.
• Showcased improved reconstruction quality and artifact reduction on in-line holography datasets.

Conclusions:

• The proposed FNet offers an efficient and effective solution for phase recovery in resource-constrained environments.
• Aligning neural networks with physical models is crucial for operational efficiency and broader applicability.
• This work advances phase recovery techniques in optical imaging and holography.