Efficient and robust phase unwrapping method based on SFNet.

Area of Science:

• Optical Metrology
• Computer Vision
• Deep Learning

Background:

• Phase unwrapping is essential for physical information retrieval in optical metrology.
• Existing deep learning methods often have complex models, limited interpretability, and are trained on simplistic noise types.
• These limitations lead to unsatisfactory performance in real-world applications.

Purpose of the Study:

• To propose SFNet, a highly efficient and robust spatial phase unwrapping method.
• To leverage the Transformer's self-attention mechanism for improved global phase relationship capture.
• To enhance accuracy and reduce errors in phase unwrapping, particularly in the presence of noise and discontinuities.

Main Methods:

• Developed SFNet, an improved SegFormer network featuring a hierarchical encoder and a lightweight MLP decoder.
• Utilized the self-attention mechanism to capture global phase dependencies.
• Trained the network on a diverse simulated dataset including various noise types and phase discontinuities.

Main Results:

• SFNet demonstrates superior performance compared to state-of-the-art deep learning and traditional methods.
• Achieved high structural stability, robustness to various noise types, and strong generalization capabilities.
• The method exhibits a lower parameter count, leading to accelerated phase unwrapping processes.

Conclusions:

• SFNet offers a significant advancement in spatial phase unwrapping, providing efficiency and robustness.
• The proposed architecture effectively addresses limitations of previous deep learning approaches.
• This method holds promise for practical applications in optical metrology requiring accurate phase retrieval.