Meiofauna Images Generation Using StyleGAN2: A Case Study of Copepoda
Type | Proceedings paper | ||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Date | 2023-06-07 | ||||||||||||
Language | English | ||||||||||||
Author(s) | Martinel Anthonin1, Benzinou Abdesslam1, Nasreddine Kamal1, Foulon Valentin1, Borremans Catherine2, Zeppilli Daniela2 | ||||||||||||
Affiliation(s) | 1 : ENIB, UMR CNRS 6285 LabSTICC, Brest, France 2 : Ifremer, Plouzané, France |
||||||||||||
Meeting | 2023 5th International Conference on Bio-engineering for Smart Technologies (BioSMART) | ||||||||||||
Source | 2023 5th International Conference on Bio-engineering for Smart Technologies (BioSMART), Paris, France, 2023, Electronic ISBN:979-8-3503-3849-2, Electronic ISSN: 2831-4352 , pp. 1-4 | ||||||||||||
DOI | 10.1109/BioSMART58455.2023.10162057 | ||||||||||||
Keyword(s) | Machine Learning, Generative Adversarial Net-works, StyleGAN2, Meiofauna | ||||||||||||
Abstract | In this work, we propose two StyleGAN2 hierarchical transfer learning approaches in order to generate images of animals belonging to the Copepoda group. Copepods are one of the most represented groups of the aquatic environment, yet only few publicly available images are available. These animals, like other groups of meiofauna, are formidable bio-indicators of environmental changes or pollution of an habitat. The used Copepoda dataset is composed of animals belonging to four orders namely Calanoida, Cyclopoida, Harpacticoida and Siphonostomatoida. Our approaches consists in first training with the available data of all the orders or with the most represented order images before training again with the images of the specimens we wish to synthesise. We evaluate the results using the FID and KID metrics. The synthetic images are promising, showing typical morphological features of Copepods, and could be used by future taxonomists. Generated images could represent a new research object for the creation of meiofauna classifiers, models that require a large number of images for training. |
||||||||||||
Full Text |
|