PSNet: A Deep Learning Model-Based Single-Shot Digital Phase-Shifting Algorithm.

Area of Science:

• Optical Metrology
• Computational Imaging
• Artificial Intelligence

Background:

• Traditional phase-shifting (PS) algorithms require multiple fringe patterns, limiting efficiency.
• Digital PS algorithms generate multiple patterns from one, enhancing efficiency.
• Deep learning models have advanced phase retrieval for complex surfaces but often ignore temporal dependencies.

Purpose of the Study:

• To propose PSNet, a novel deep learning-based digital phase-shifting algorithm.
• To improve the accuracy of phase retrieval by incorporating global temporal information.
• To address the limitation of ignoring inter-frame dependency in existing deep learning PS methods.

Main Methods:

• Developed a deep learning model, PSNet, for digital phase-shifting.
• Constructed a loss function that integrates local and global temporal information from fringe patterns.
• Focused on learning inter-frame dependency between adjacent generated patterns.

Main Results:

• PSNet demonstrated improved accuracy in generating phase-shifting patterns.
• Enhanced accuracy was observed in the associated phase retrieval process.
• Both simulations and real-world experiments validated the algorithm's performance.

Conclusions:

• The proposed PSNet algorithm effectively utilizes temporal information for accurate phase retrieval.
• PSNet offers a significant improvement over state-of-the-art methods in digital phase-shifting.
• The integration of local and global temporal loss enhances the robustness of phase retrieval for complex surfaces.