Phase unwrapping using deep learning in holographic tomography.

Area of Science:

• Optics and Photonics
• Image Processing
• Artificial Intelligence

Background:

• Holographic tomography (HT) generates phase images crucial for 3D reconstruction.
• Phase unwrapping is essential but challenging due to noise and irregularities in HT data.
• Conventional methods for phase unwrapping are often slow, unreliable, and lack automation.

Purpose of the Study:

• To develop a robust, automated, and efficient phase unwrapping method for holographic tomography.
• To address the limitations of conventional phase unwrapping algorithms in handling noisy and complex experimental data.

Main Methods:

• A two-step convolutional neural network (CNN) pipeline based on the U-Net architecture was proposed.
• The pipeline incorporates a denoising step followed by a phase unwrapping step.
• Attention Gates (AG) and Residual Blocks (RB) were integrated into the U-Net for enhanced unwrapping performance.

Main Results:

• The proposed deep learning pipeline successfully unwrapped highly irregular and noisy phase images from HT.
• The U-Net architecture, enhanced with AGs and RBs, demonstrated improved performance in phase unwrapping.
• This is the first deep learning solution trained exclusively on real-world HT experimental images.

Conclusions:

• The developed deep learning pipeline offers a significant advancement in holographic tomography phase processing.
• The method provides a noise-robust, reliable, and potentially automatable solution for phase unwrapping.
• The findings pave the way for more accurate and efficient 3D reconstructions using holographic tomography.