Pose estimation and action recognition in fish with image prosessing techniques

The high nutritional value of fish, raise in the demand for fish meat due to population growth and the fact that the increase will continue even more in the future reveal the place of fish in human life. However, despite this importance, factors such as the lack of sufficient data sets make it difficult to conduct studies on fish. In other words, the number of studies on fish in the literature is not adequate. It can be said that studies to be done in this field will, on the one hand, contribute to the literature, on the other hand, increase the level of awareness on this issue. In this study, the data set shared openly in the study conducted by Xu et al in 2017 was used as data. There are 9 different classes of fish movements in the dataset consisting of 16528 images. In this study, as different from the literature, it is proposed to classify the motion of the fish by transferring the motion of the fish to a two-dimensional plane with the pose estimation method and using transfer learning methods over the image of the motion. In the study, image processing methods were used to transfer the motion to the plane. In this study, following models were used: YOLOv8 model for key point detection and Xception, ResNet50, ResNet101, Inception v3, NasNetLarge, DenseNet169, VGG 16, EfficientNet-B4 models for classification. As a result, among the models, EfficientNet-B4 had the highest model performance with 89.43% accuracy, 91.85% precision, 88.80% recall and 89.17% f1-score values.