Active learning for plankton recognition

Abstract

Plankton is essential part of the ocean ecosystem as it is an important food source for many marine life, both large and small and is also oxygen producer. Therefore, its quantity and condition must be monitored and its disappearance cannot be allowed. Successful recognition of a plankton is labour-intensive and time-consuming. With relation to traditional machine learning algorithms, Deep Learning (DL) has an advantage in most application areas for pattern recognition. DL requires a lot of labeled data to complete model training, which can sometimes not only take a long time for annotators, but their expertise and labor could also be expensive. In this thesis Active Learning (AL) technique will be considered as a solution for minimizing human interaction for labeling. The general idea of AL method is to understand on how to label as few data and still achieve high accuracy in classification and recognition tasks. The approach is to use a combination of DL and AL so-called DeepAL and evaluate the performance of plankton recognition task. During the work, such a system utilising DeepAL was implemented and tested based on a Convolutional Neural Network (CNN) model with and without pre-training. Experiments with 5 common AL methods and random sampling as a baseline were evaluated and ranked. One of these methods, entropy sampling, showed the best results, outperforming random sampling and the other considered querying strategies. Finally, experiments suggested that integration of AL to plankton recognition task can improve accuracy of the CNN classification.