<![CDATA[The Advancements in precision aquaculture demand robust visual monitoring systems capable of accurate, real-time fish detection in complex underwater environments characterized by turbidity, occlusion, and dynamic illumination. While YOLO (You Only Look Once) architectures have demonstrated high efficiency in object detection tasks, their comparative performance for underwater fish detection remains underexplored, particularly across recent variants such as YOLOv5, YOLOv8, and YOLOv11. This study presents a systematic evaluation of three state-of-the-art YOLO models using a curated GlowFish dataset consisting of 533 annotated images across three fluorescent species. Data were acquired under controlled but visually diverse conditions using multi-angle imaging and standardized illumination. A uniform training pipeline, consistent annotation using the COCO format, and identical hyperparameters were applied across models to ensure fair benchmarking. Key evaluation metrics include precision, recall, mAP@0.5, and mAP@0.5:0.95. Experimental results reveal that YOLOv5 achieved the highest precision (0.963) and mAP@0.5 (0.967), while YOLOv8 delivered superior recall (0.930) and more balanced detection across species classes. YOLOv11 demonstrated architectural potential but showed greater sensitivity to class imbalance and reduced confidence stability. Visual analysis and confusion matrices further confirmed model-specific trade-offs in classification reliability and localization precision. This work contributes critical empirical insights into the selection of YOLO architectures for intelligent aquaculture systems, offering practical guidance for real-time aquatic monitoring deployments. Future research will extend this framework to multi-species, multi-environment datasets, integrate spatiotemporal behavioral tracking, and investigate deployment on resource-constrained edge-AI platforms, advancing the field toward interpretable and autonomous aquatic monitoring solutions.]]>
Copyrights © 2026
