Machine-learning-based size estimation of marine particles in holograms recorded by a submersible digital holographic camera

Liu, Zonghua; Giering, Sarah ORCID: https://orcid.org/0000-0002-3090-1876; Thevar, Thangavel; Burns, Nick; Ockwell, Mike; Watson, John. 2023 Machine-learning-based size estimation of marine particles in holograms recorded by a submersible digital holographic camera. In: Oceans 2023 - Limerick, Limerick, 05 - 08 June 2023. IEEE Xplore, 1-8.

Before downloading, please read NORA policies.

Preview

Text “© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.” Oceans Limerick_ZLiu.pdf - Accepted Version Download (2MB) | Preview

Abstract/Summary

Particle size estimation is key to understanding carbon fluxes and storage in the marine ecosystem. Images of particles provide much information about their size. A subsea digital holographic camera was used to image particles in vertical trajectory in South Georgia. The holograms were processed using a rapid hologram processing suite that extracted focused particle vignettes from these raw holograms. A machine-learning-based method has been developed to analyse the particle size information from these vignettes. To be specific, a structured-forest-based model trained on a group of synthetic holographic particle images is used to detect the particle edges in these vignettes. Following that, a set of pixel-wise morphology operators are used to extract particle regions (masks) from their edge images. Lastly, the size information of the recorded particles can be calculated based on these mask images. The proposed method has been evaluated on a group of synthetic holograms and real holograms, compared with the other ten methods, including four edge-based methods, four region-based methods, a thresholding-based method, and a Kmeans-based method. The results show that our method has the best performance regarding accuracy and processing time. It reaches ∼0.7 of mean IoU and ∼25 s of running time on the 1,000 test vignettes. In terms of qualitative analysis, the regions of the given examples extracted by the proposed method closely match the real particle regions. We also use this method to analyse the size distributions of two profiles, and some generic results are given in this paper.

Item Type:	Publication - Conference Item (Paper)
Digital Object Identifier (DOI):	https://doi.org/10.1109/OCEANSLimerick52467.2023.10244456
Date made live:	29 Nov 2023 12:50 +0 (UTC)
URI:	https://nora.nerc.ac.uk/id/eprint/536353

Actions (login required)

View Item

Document Downloads

More statistics for this item...