Segmenting Freshwater Fish Images with Convolutional Neural Networks

The acquisition of important information from fisheries for studies in Ichthyology and Fisheries currently takes place manually. Data on the length and morphology of freshwater fish are valuable for researchers to determine indicators in their studies. Technological development in this area aims to increase the speed and effectiveness of information gathering, assisting researchers, students, and fishermen in this goal. Aligned with this objective, this work presents a comparison between two models, Mask R-CNN and YOLOv8, used to segment fish images and generate their masks, ultimately feeding them into an automatic measurement algorithm, which is the broader purpose of this project. The results show that the models can segment various types of fish in different positions and environments, with Mask R-CNN achieving 80.16% using the IOU (Intersection over Union) metric and YOLOv8 achieving 86.15%.