Unsupervised feature learning and clustering of particles imaged in raw holograms using an autoencoder

Area of Science:

• Optics and Photonics
• Machine Learning
• Microscopy

Background:

• Digital holography offers high-resolution particle shape information for identification.
• Hologram reconstruction is computationally intensive, limiting real-time applications on low-power platforms.

Purpose of the Study:

• To investigate unsupervised object clustering directly from raw digital holograms, bypassing computationally expensive reconstruction.
• To enable real-time particle analysis on low-power sensor platforms.

Main Methods:

• Utilized deep-learning autoencoder and self-organizing mapping networks for unsupervised clustering of raw interference patterns.
• Demonstrated the method on synthetic and real holograms of microscopic particles.
• Employed transfer learning with combined synthetic and real datasets to improve accuracy.

Main Results:

• Clustering raw holograms achieved 94.4% accuracy on synthetic data, comparable to 97.4% with reconstructed holograms.
• Processing time reduced by three orders of magnitude (<0.1s per image on a low-power CPU).
• Clustering accuracy on real holograms increased from 47.1% to 75.9% using synthetic data training and transfer learning.

Conclusions:

• Direct clustering of raw digital holograms is a viable and efficient alternative to reconstruction-based methods.
• The developed deep-learning approach enables real-time particle identification on resource-constrained platforms.
• Training with synthetic data significantly enhances clustering performance on real-world holographic data.