Generative Adversarial Networks for Annotating Images of Otoliths

Abstract

This thesis explores the use of generative adversarial networks (GANs) for annotating images of otoliths to determine the age of fish. The proposed solution not only provides accurate age determinations, but also visual representations of the otolith images with growth rings marked with dots, making it applicable as explainable artificial intelligence. The convolutional neural network models I propose are based on Pix2Pix GANs and Wasserstein GANs, with the latter showing the success in my experiments. The successful models achieve an accuracy of 82.8% and 81.5% in age determination, including an offset of plus-minus 2 from the real ages of the dataset.