Exploring the potential of deep learning models for fish classification

Abstract

In this thesis we have studied and applied one of the recently proposed deep learning architecture, Vision transformer (ViT). We have observed the performance of ViT model under conditions like with and without transfer learning, with and without image augmentation under three different publicly available datasets. We have also observed the performance of other two popular deep neural network models like VGG16 and Inception V3 under same conditions and same three datasets. In overall comparisons, ViT showed excellent performance and can be proposed for fish image classification.