Generating synthetic plankton images improves rare species classification
Multimodal Taxonomic Conditioning for Generative Plankton Imagery
Computer Vision and Pattern RecognitionMachine Learning
Summary
Studying tiny ocean creatures called plankton is hard because some important types are very rare and there aren’t enough images to teach computers to recognize them well. The researchers made a method to create realistic fake images of plankton based on their categories by combining smart language and image tools. This helps train better computer programs that can identify rare plankton species. They checked that these fake images closely match real ones and help improve classification.
What this means in practice
- •For marine monitoring teams: Create synthetic plankton images of scarce species to improve automated ecological surveys with limited real data.
- •For environmental ai developers: Train more accurate plankton classifiers by augmenting datasets with taxonomy-conditioned synthetic images for ecological data projects.
Authors
Daniela Ivanova, Ozgu Goksu, Nicolas Pugeault
Abstract
Automated plankton imaging produces severely long-tailed datasets, where the rare taxa of greatest ecological interest have too few images to train or evaluate classifiers reliably. We generate synthetic plankton imagery conditioned on taxonomy: a CLIP encoder is adapted on a large plankton corpus with a ranked contrastive objective extended to deep, ragged taxonomies, then frozen to condition a parameter-efficient diffusion transformer. We evaluate synthetic sample quality on distributional fidelity and downstream classifier utility.