AlgaeClass_Net: Optimizing Few-Shot Marine Microalgae Classification With Multi-Scale Feature Enhancement Network

Abstract

As the eutrophication of the water body becomes more and more serious, the algae in the water body grow in large quantities and eventually form harmful algal blooms, causing great harm to the marine ecosystem. Therefore, how to quickly and accurately identify algae and make precautions becomes the key to solving this problem. Currently, more than tens of thousands of microalgae species are known around the world, but publicly available data are sparse, many of the species are characterized similarly to each other, and it is currently challenging to train an effective classification model with limited data. Existing few-shot learning classification algorithms that utilize meta-learning for metrics can be a good solution to this problem. In this paper, an AlgaeClass_Net algorithm that combines an improved multi-scale feature fusion with a feature enhancement module is proposed for the fine-grained features of microalgae. Furthermore, it utilizes a metric learning approach to classify few-shot microalgae by calculating the distances between the feature vectors of samples in the query set and the feature vectors of samples in the support set. The experimental results showed that the method achieves 78.55% and 91.20% classification accuracies under different tasks of 5-way 1-shot and 5-way 5-shot, respectively, with an improvement of 3.51% and 4.97% on the suboptimal model, respectively. It provides new research ideas for the identification of marine microalgae and the development and utilization of marine renewable energy.

Keywords

• Marine microalgae
• few-shot learning
• multi-scale feature fusion
• feature enhancement
• metric learning