Open-Set Plankton Recognition Using Similarity Learning

Abstract

Automatic plankton recognition provides new possibilities to study plankton populations and various environmental aspects related to them. Most of the existing recognition methods focus on individual datasets with a known set of classes limiting their wider applicability. Automated plankton imaging instruments capture images of unknown particles and the class (plankton species) composition varies between geographical regions and ecosystems. This calls for an open-set recognition method that is able to reject images from unknown classes and can be easily generalized to new classes. In this paper, we show that a flexible model capable of high classification accuracy can be obtained by utilizing similarity learning and a gallery set of known plankton species. The model is shown to generalize well for new plankton classes added in the gallery set without retraining the model. This provides a good basis for the wider utilization of plankton recognition methods in aquatic research.