Benchmark of image classification using several large plankton datasets: Convolutional Neural Networks improve detection of rare taxa - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Benchmark of image classification using several large plankton datasets: Convolutional Neural Networks improve detection of rare taxa

Thelma Panaiotis
Rainer Kiko
Lars Stemmann

Résumé

Plankton imaging instruments generate an ever increasing volume of data which is mostly processed through machine learning algorithms. However, classifying plankton images is a challenging computer science task in its own right: datasets are strongly unbalanced; the dominant classes are often not biologically interesting (artefacts, bubbles) and/or very heterogeneous looking (marine snow); and images span a large size range. Despite a wealth of reports on the performance of automatic plankton images classifiers, we still do not have a definitive idea regarding how methods compare with each other and where they can systematically be trusted. This is mostly because those reports rely on rather small unpublished datasets, not necessarily representative of real-life biological samples in terms of size, number of categories and proportions. Here we report the performance of a classic classification method (Random Forest on handcrafted image features) and a more recent one (a Convoluted Neural Network) on large publicly released datasets, from five widely used plankton imaging instruments. We show that CNN improve classification performance but only noticeably on poorly represented (a few hundred images) classes. Finally, we showcase the difference between the predictions of the two classifiers and a human-checked truth on several real-world datasets, to give insights regarding which ecological questions can or cannot be studied from computer-generated classifications only.
Fichier non déposé

Dates et versions

hal-04026436 , version 1 (13-03-2023)

Identifiants

  • HAL Id : hal-04026436 , version 1

Citer

Thelma Panaiotis, Guillaume Boniface-Chang, Gabriel Dulac-Arnold, Benjamin Blanc, Tristan Biard, et al.. Benchmark of image classification using several large plankton datasets: Convolutional Neural Networks improve detection of rare taxa. ASLO 2021 Aquatic Sciences Meeting, Jun 2021, Online, Unknown Region. ⟨hal-04026436⟩
26 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More