Real-time detection of fish feeding behavior with PM-YOLO utilizing YOLOv8

Feeding inefficiencies in pond aquaculture not only elevate operational costs and environmental impact but also compromise fish growth and welfare, driving an urgent need for precision feeding technologies. However, accurate, real-time detection of feeding events in complex, cluttered pond environments remains challenging due to dense, small targets, variable lighting, and background disturbances. To address these issues, we propose PM-YOLO, an enhanced YOLOv8n framework that integrates Parallelized Patch-Aware Attention modules for multi-scale feature retention and a Minimum Point Distance IoU loss term for superior bounding-box regression of overlapping fish. We also curate FeedFishDatas, a 7,400-image dataset capturing diverse feeding scenarios, and demonstrate that PM-YOLO achieves 87.7% precision and 83.5% mAP-50—outperforming standard YOLOv8n and YOLOv11n—while maintaining real-time inference at 8.1 ms per frame. Our ablation studies isolate the contributions of each innovation, confirming their synergistic impact on detection performance. Looking ahead, we envisage lightweight PM-YOLO variants via model pruning and quantization for edge deployment, temporal smoothing for stability, and domain adaptation to extend applicability across species and water conditions, paving the way for fully automated, closed-loop feeding systems in sustainable aquaculture. PM-YOLO improves fish feeding behavior detection with higher accuracy in complex pond environments. The model uses advanced techniques like multi-scale feature fusion and optimized IoU for better small target detection. The study demonstrates real-time detection potential, supporting precision feeding in aquaculture.