Three-stage training strategy phase unwrapping method for high speckle noises

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Deep learning methods are prevalent in phase unwrapping but struggle with noise, leading to inaccuracies.
• High noise levels in wrapped phase data impede accurate wrap count prediction and phase calculation.

Purpose of the Study:

• To develop a robust phase unwrapping method resilient to high noise levels.
• To enhance measurement accuracy in challenging, noisy environments.

Main Methods:

• A three-stage multi-task learning approach: wrapped phase denoising, wrap count prediction with residual compensation, and unwrapped phase error compensation.
• Integration of a novel convolution-based multi-scale spatial attention module to mitigate spatially inconsistent noise.
• Training strategy divided into three distinct stages for progressive refinement of phase retrieval.

Main Results:

• The proposed method demonstrates superior noise robustness compared to traditional and deep learning techniques like TIE, UNet, and DeepLabV3+.
• Achieved high phase retrieval accuracy even under significant noise interference.
• The multi-scale spatial attention module effectively reduced noise interference.

Conclusions:

• The three-stage multi-task phase unwrapping method offers a significant advancement for accurate phase retrieval in noisy conditions.
• The approach provides a robust solution for applications requiring high-precision phase measurements.
• The method outperforms existing techniques in noise resilience and accuracy.