A study of generative adversarial networks to improve classification of microscopic foraminifera

Abstract

Foraminifera are single-celled organisms with shells that live in the marine environment and can be found abundantly as fossils in e.g. sediment cores. The assemblages of different species and their numbers serves as an important source of data for marine, geological, climate and environmental research.

Steps towards automatic classification of foraminifera using deep learning (DL) models have been made (Johansen and Sørensen, 2020), and this thesis sets out to improve the accuracy of their proposed model. The recent advances of DL models such as generative adversarial networks (GANs) (Goodfellow et al., 2014), and their ability to model high-dimensional distributions such as real-world images, are used to achieve this objective.

GANs are studied and explored from a theoretical and empirical standpoint to uncover how they can be used to generate images of foraminifera. A multi-scale gradient GAN is implemented, tested and trained to learn the distributions of four high-level classes of a recent foraminifera dataset (Johansen and Sørensen, 2020), both conditionally and unconditionally. The conditional images are assessed by an expert and a deep learning classification model and is found to contain mostly valuable characteristics, although some artificial artifacts are introduced. The unconditional images measured a Fréchet Inception distance of 47.1.

From the conditionally learned distributions a total of 10 000 images are sampled from the four distributions. These images are used to augment the original foraminifera training set in an attempt to improve the classification accuracy of (Johansen and Sørensen, 2020). Due to limitations of computational resources, the experiments were carried out with images of resolution 128 × 128. The synthetic image augmentation lead to an improvement in mean accuracy from 97.3 ± 0.4 % to 97.4 ± 0.7 % and an improvement in best achieved accuracy from 97.7 % to 98.5 %.