Dual-constrained physics-enhanced untrained neural network for lensless imaging.

Area of Science:

• Computational imaging
• Applied physics
• Machine learning

Background:

• Lensless imaging offers advantages but requires accurate phase retrieval.
• Untrained neural networks (UNNs) enable imaging from intensity data without large datasets.
• UNNs face challenges with overfitting, hindering stable and robust reconstruction.

Purpose of the Study:

• To develop a robust phase retrieval method for lensless imaging using UNNs.
• To address the overfitting problem in UNNs for improved reconstruction stability.
• To introduce a dual-constrained network framework for intensity-to-phase conversion.

Main Methods:

• A dual-constrained untrained network was modeled for phase retrieval.
• A phase-amplitude alternating optimization framework was designed.
• Phase optimization incorporated deep image and total variation priors.
• Amplitude optimization utilized a total variation denoising-based Wirtinger gradient descent method.

Main Results:

• The proposed method effectively splits the intensity-to-phase problem into distinct optimization tasks.
• Iterative optimization of phase and amplitude led to high-performance wavefield reconstruction.
• Experimental results validated the superiority of the dual-constrained UNN approach.

Conclusions:

• The dual-constrained UNN framework successfully mitigates overfitting in lensless imaging.
• This approach provides a stable and robust solution for phase retrieval from single-frame intensity data.
• The method demonstrates significant potential for advancing lensless imaging technologies.