Real-Time Fish Detection in Underwater Videos Using YOLOv8n and YOLOv8m Architectures

Abstract

Fish detection in the underwater setting is a difficult problem considering that visibility is low, there is refraction of light, turbidity and intricate surroundings in water. This paper provides a comparative analysis of the YOLO8n and YOLO8m to underwater fish detection in real-time. A wide range of underwater data with illumination variations, occlusion, and an orientation of the fish was used to train and test the models. The experimental results prove that YOLO8n can obtain excellent overall performance with Precision of 0.96, Recall of 0.96, mAP@50 of 0.97, and mAP@50-95 of 0.76 and high inference speed of 110 FPS. Though YOLO8m has a little higher accuracy 0.95, it has good recall and much more complex computation. According to the findings, lightweight architecture can provide competitive and robust performance in detection without consuming a lot of resources. Consequently, the proposed model is selected as YOLO8n because it has the best balance between accuracy, efficiency, and the ability to process in real- time. The framework proposed has great potential in use in the marine ecosystems monitoring, aquaculture control and smart underwater surveillance system.