Evaluation of the Performance of Transfer Learning Techniques in Classifying Fish Species

Classification of fish is of great importance for the sustainability of ecosystems and protection of biodiversity. In this study, fish are classified according to their species using deep learning techniques. Convolutional Neural Networks and transfer learning approaches based on these network models are tested on a fish dataset consisting of 9 species and the performance of the models is compared. In addition to Convolutional Neural Networks, transfer learning models used are ResNet, MobileNet, VGG16, Inception and AlexNet models. The highest performance obtained from the experiments, where $80 \%$ of the dataset is divided for training and $20 \%$ for testing, is obtained from the ResNet architecture with $99.94 \%$ according to accuracy and F1-score metrics. When the performances obtained from other models are examined, there are classification results between $92.09 \%$ and $99.67 \%$.