Pufferfish detection with computer vision and deep learning methods

This thesis focuses on the training of an object detection model aimed at identifying pufferfish known as Lagocephalus sceleratus. Harmful species like Lagocephalus sceleratus pose threats to economic losses and human health. This type of fish is a species known for being non-native, widely distributed, and poisonous, lacking natural predators. The study aims to train a model for pufferfish detection using computer vision and deep learning techniques. Considering the existing literature and technology, the research focuses on the most recent You Only Look Once (YOLO) algorithms to automatically detect such fish. The data obtained were sourced from diving schools and instructors in the Mediterranean. Frames extracted from videos were appropriately labeled in YOLO format, resulting in a dataset comprising 2473 images. YOLOv8, the latest version of YOLO, deep learning models were trained with this dataset, and experiments were conducted using the trained models. During the training, it was observed that pufferfish could be better detected from their heads and side angles. This stems from the fact that the model attempts to capture the fish's eyes and surroundings and can identify the fish's shape from a side angle. Due to challenges in the manual labeling process, the tails and fins of the fish were not fully considered, affecting the model's focus. Additionally, using images of fish taken from different angles and lighting conditions enhanced the model's generalization capability. This model presents a more effective and sustainable approach to solving the pufferfish problem, laying the foundation for future research endeavors.