Phase unwrapping in digital holographic interferometry via physics-informed convolutional-Fourier neural network

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Image Processing

Background:

• Phase unwrapping is critical for optical imaging and metrology but is hindered by speckle noise.
• Conventional spatial-domain Convolutional Neural Network (CNN) methods struggle with frequency-domain feature extraction under speckle noise.

Purpose of the Study:

• To develop a deep learning-based phase unwrapping algorithm robust to high speckle noise.
• To improve phase unwrapping performance by integrating physics-informed neural networks and an uncertainty-aware speckle model.

Main Methods:

• A physics-informed convolutional-Fourier neural network was developed, incorporating an uncertainty-aware mathematical model of speckle.
• A parallel convolution-Fourier architecture was used to unify spatial and frequency-domain feature extraction.
• An integrated loss function considering speckle noise, phase gradient, and absolute phase was employed.

Main Results:

• The proposed method demonstrated superior performance compared to existing phase unwrapping algorithms in simulations.
• Experimental validation using digital holography confirmed the algorithm's effectiveness.
• The deep learning approach significantly enhanced phase unwrapping accuracy under high speckle noise.

Conclusions:

• The developed physics-informed deep learning architecture effectively addresses speckle noise in phase unwrapping.
• Unifying spatial and frequency-domain feature extraction offers a powerful approach for optical information processing.
• This method shows promise for broader applications in information processing tasks requiring robust phase recovery.