Fish Detection from Underwater Images Using YOLO and Its Challenges

Abstract

Object detection is one of the most popular areas of research on computer vision. Monitoring of marine ecosystems is now a demand for saving nature. Sending human beings for observing the marine environment for some of the tasks is more dangerous. Multiple works are going on the improvement of Autonomous Underwater Vehicle (AUV) to monitor the underwater environment. In aquaculture fisheries, fish monitoring by AUV has got great importance. It is required to collect information regarding different types of fish to maintain the marine ecosystem. The work concentrates on object detection by using real-time object detectors like YOLOv3 and YOLOv4. The work used the Roboflow fish detection dataset for validating our work. It is a dataset of 26 classes. YOLOv3 and YOLOv4 are rigorously tested here and the final result is shown using precision, recall, IOU, and mAP. The work achieved 50% mean average precision after 16,000 iterations. The work discussed the main challenges of fish detection. Also, it stated the different reasons for the getting not higher values of performance parameters.