Convolutional Neural Networks-based Plankton Image Classification System

To the marine ecosystem, plankton are vitally significant for they produce more than half of the primary productivity on earth and play an irreplaceable role in the global carbon cycle. However, most of traditional plankton population measuring and monitoring methods become gradually impractical for they are time consuming and cannot scale to the granularity required in large-scale studies. Thus we introduced a machine learning model, convolutional neural networks, and utilized it to implement an image classification system and automate the plankton image identification process. In the experiment, our classification system’s logarithmic loss in test set was 0.7736, performing much better than systems using linear model and multi-layer perceptron model. In addition, the noise resistance ability of our system was examined by introducing Gaussian noise.