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

Liu, Zonghua; Thevar, Thangavel; Takahashi, Tomoko; Burns, Nicholas; Yamada, Takaki; Sangekar, Mehul; Lindsay, Dhugal; Watson, John; Thornton, Blair

Abstract

Digital holography is a useful tool to image microscopic particles.Reconstructed holograms give high-resolution shape information that can beused to identify the types of particles. However, the process ofreconstructing holograms is computationally intensive and cannot easilykeep up with the rate of data acquisition on low-power sensor platforms.In this work, we explore the possibility of performing object clusteringon holograms that have not been reconstructed, i.e., images of rawinterference patterns, using the latent representations of a deep-learningautoencoder and a self-organizing mapping network in a fully unsupervisedmanner. We demonstrate this concept on synthetically generated hologramsof different shapes, where clustering of raw holograms achieves anaccuracy of 94.4\%. This is comparable to the 97.4\% accuracy achieved usingthe reconstructed holograms of the same targets. Directly clustering rawholograms takes less than 0.1 s per image using a low-power CPU board.This represents a three-order of magnitude reduction in processing timecompared to clustering of reconstructed holograms and makes it possible tointerpret targets in real time on low-power sensor platforms. Experimentson real holograms demonstrate significant gains in clustering accuracythrough the use of synthetic holograms to train models. Clusteringaccuracy increased from 47.1\% when the models were trained only on thereal raw holograms, to 64.1\% when the models were entirely trained on thesynthetic raw holograms, and further increased to 75.9\% when models weretrained on the both synthetic and real datasets using transfer learning.These results are broadly comparable to those achieved when reconstructedholograms are used, where the highest accuracy of 70\% achieved whenclustering raw holograms outperforms the highest accuracy achieved whenclustering reconstructed holograms by a significant margin for ourdatasets.

Citation

LIU, Z., THEVAR, T., TAKAHASHI, T., BURNS, N., YAMADA, T., SANGEKAR, M., LINDSAY, D., WATSON, J. and THORNTON, B. 2021. Unsupervised feature learning and clustering of particles imaged in raw holograms using an autoencoder. Journal of the Optical Society of America A [online], 38(10), pages 1570-1580. Available from: https://doi.org/10.1364/JOSAA.424271

• Journal Article Type: Article
• Acceptance Date: Sep 1, 2021
• Online Publication Date: Sep 27, 2021
• Publication Date: Oct 1, 2021
• Deposit Date: Jun 11, 2024
• Publicly Available Date: Jun 11, 2024
• Journal: Journal of the Optical Society of America A
• Print ISSN: 1084-7529
• Electronic ISSN: 1520-8532
• Publisher: Optical Society of America
• Peer Reviewed: Peer Reviewed
• Volume: 38
• Issue: 10
• Pages: 1570-1580
• DOI: https://doi.org/10.1364/JOSAA.424271
• Keywords: Digital holographic imaging; Fast Fourier transforms; Field programmable gate arrays; Holographic microscopy; Parallel processing; Real time holography
• Public URL: https://rgu-repository.worktribe.com/output/2114605
• Publisher URL: https://opg.optica.org/josaa/abstract.cfm?URI=josaa-38-10-1570