EESANet: edge-enhanced self-attention network for two-dimensional phase unwrapping.

Area of Science:

• Image Processing
• Computer Vision
• Signal Processing

Background:

• Phase unwrapping is crucial for many applications, but existing methods struggle with accuracy and robustness.
• Wrapped phase maps contain inherent prior information that is not fully utilized by current algorithms.

Purpose of the Study:

• To develop a quantitative indicator for measuring prior information in wrapped phase maps.
• To propose an advanced deep learning model for accurate and robust two-dimensional phase unwrapping.

Main Methods:

• A novel quantitative indicator for prior information in wrapped phase maps was developed.
• An Edge-Enhanced Self-Attention Network (EESANet) with a symmetrical encoder-decoder architecture was designed.
• Key components include Serried Residual Blocks, Atrous Spatial Pyramid Pooling, Positional Self-Attention, and an Edge-Enhanced Block.
• A weighted cross-entropy loss function was employed to address category imbalance.

Main Results:

• The proposed method achieved higher precision compared to state-of-the-art techniques.
• The EESANet demonstrated superior robustness in phase unwrapping tasks.
• Experiments confirmed better generalization capabilities of the proposed approach.

Conclusions:

• The developed quantitative indicator effectively measures prior information in wrapped phase maps.
• EESANet significantly improves the accuracy, robustness, and generalization of 2D phase unwrapping.
• The proposed method offers a promising solution for complex phase unwrapping challenges.