Self-supervised model-driven deep learning for two-step phase-shifting interferometry.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Metrology

Background:

• Deep learning in interferometry often requires high-quality training data, limiting accuracy and generalizability.
• Existing data-driven methods face challenges with dataset dependency for precise phase retrieval.

Purpose of the Study:

• To develop a model-driven deep learning framework for two-step phase-shifting interferometry.
• To enable accurate and robust phase retrieval without ground truth phase maps.

Main Methods:

• A pre-trained normalization network (PNNet) normalizes interferograms with arbitrary phase shifts.
• An untrained model-driven network (UMNet) performs self-supervised learning using a physics-based model.
• Phase maps and phase shifts are generated from normalized interferograms.

Main Results:

• The model-driven approach achieves accurate and robust phase retrieval.
• Errors are reduced by over 30% compared to traditional data-driven methods.
• Demonstrates the effectiveness of self-supervised learning in interferometry.

Conclusions:

• Model-driven deep learning offers a promising alternative to data-driven methods in phase-shifting interferometry.
• Self-supervised learning significantly improves accuracy and reduces reliance on extensive training datasets.
• This approach advances high-accuracy phase retrieval techniques.