Classification of Marine Plankton Based on Few-shot Learning

The current computer vision usually requires abundant training samples to classify target images, while it requires only a small number of samples if it is the same task for humans. This article attempts to address the few-shot marine plankton image classification problem. The model proposed in this paper uses transfer learning to train a classifier on base classes and fine-tunes on new classes to train a new classifier. Considering the small number of training samples, effective representation of each image is very important. We use the soft max loss function and center loss function to jointly train the model in order to minimize the intra-class distance of the depth features and obtain more robust and discriminative depth features. We conduct experiments on multiple marine plankton image data sets, and the outcomes have proved the effectiveness of the model. Our model is competitive with many existing few-shot image classification models in performance.