Classification of Holograms with 3D-CNN.

Area of Science:

• Digital holography
• Machine learning
• Volumetric data analysis

Background:

• Holograms capture complete volumetric information via interference patterns.
• 2D Convolutional Neural Networks (CNNs) struggle with global volumetric data due to local feature extraction.
• Reconstructing holographic images requires advanced techniques to access depth information.

Purpose of the Study:

• To propose and evaluate a 3D CNN architecture for hologram volumetric data analysis.
• To demonstrate the limitations of 2D CNNs in processing holographic volumetric information.
• To compare the performance of 3D CNNs against 2D CNNs on a holographic classification task.

Main Methods:

• Hologram volumetric data extraction using wavefield propagation algorithms.
• Implementation of a 3D CNN architecture for volumetric data processing.
• Comparison with a 2D CNN using equivalent data.
• Introduction of hologram defocus augmentation for improved robustness.

Main Results:

• The 3D CNN method significantly outperforms the 2D CNN in holographic image classification accuracy.
• The 3D CNN demonstrates superior robustness to defocused input holograms.
• Hologram defocus augmentation enhances performance for both 2D and 3D methods on defocused inputs.

Conclusions:

• A 3D CNN approach is more effective for extracting and utilizing global volumetric information from holograms.
• 2D CNNs are inherently limited in decoding the full volumetric data encoded in holograms.
• The proposed 3D CNN method offers a more robust and accurate solution for holographic data analysis and classification.