Phys-HUSPU: a physics-constrained and heteroscedastic uncertainty-guided phase unwrapping framework.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Phase unwrapping is crucial for optical metrology and imaging, but current deep learning methods struggle with non-ideal data due to reliance on simulated datasets.
• Supervised deep learning models often fail to generalize to real-world conditions like phase aliasing and overexposure because they learn statistical patterns, not underlying physics.

Purpose of the Study:

• To develop a physics-constrained phase unwrapping framework that reduces reliance on ground-truth labels and improves generalization.
• To enhance the robustness and accuracy of phase unwrapping in optical metrology and imaging, particularly in non-ideal scenarios.

Main Methods:

• Proposed Phys-HUSPU framework, integrating physics-informed penalties and regularization terms to guide the solution space.
• Reconstructed phase unwrapping as a joint regression of wrap-count gradient fields and pixel-wise heteroscedastic uncertainty.
• Utilized predicted uncertainty as adaptive weights to balance data fidelity constraints and physical priors.

Main Results:

• Demonstrated significant improvement in network robustness against noise, phase aliasing, and signal truncation.
• Achieved excellent generalization capabilities in cross-domain applications, outperforming existing methods.
• Reduced dependence on ground-truth labels through physics-informed guidance.

Conclusions:

• The proposed physics-constrained and uncertainty-guided framework offers a more robust and generalizable solution for phase unwrapping.
• Phys-HUSPU effectively handles non-ideal conditions by adaptively integrating physical priors, leading to physically consistent phase solutions.
• This approach advances deep learning applications in optical metrology and imaging by incorporating physical constraints.