Deep Pyramidal Residual Networks for Plankton Image Classification

In this paper, we explore a new deep learning algorithm called Deep Pyramidal Residual Networks (PyramidNet) to solve plankton image classification problem. It is well known that plankton image classification is playing an increasingly important role in the research of plankton organisms. Nowadays, scientists tend to use image-based technologies to study marine plankton. However, manual classification of the plankton images is a time-consuming and labor-intensive work. So it is urgent to find methods that can do automatic classification using few samples labeled by experts and have good performance. In recent years, Deep Convolutional Neural Networks (DCNNs) have shown remarkable performance in image classification tasks. The recently proposed algorithm PyramidNet even performed better than all previous state-of-the-art DCNNs in image classification experiments using some image classification benchmark datasets. So we explore the PyramidNet to solve our plankton image classification problem and it shows superior generalization ability, gaining higher accuracies. Another meaningful point of our experiment is the improved F1 score compared to other benchmark methods. We use the WHOI-Plankton dataset in our experiment. Although some DCNNs have achieved high accuracies on this dataset, it doesn't mean that the classifiers are really accurate about the minority classes because of the low F1 score which is also another important measurement for image classification. In this paper, we choose the accuracy and F1 score as our evaluating indicators and it ultimately shows that the PyramidNet outperforms the benchmark methods not only on the accuracy but also on the F1 score.