• Optimized YOLOv8n integrates AKConv, SE, and BiFPN for better feature extraction. • Precision improved to 92.7 outperforming baseline. • Model size reduced from 6.2MB to 5.4MB, decreasing computational demands. • The system supports precision feeding, reducing waste and promoting sustainability. Feed costs comprise over 40% of total aquaculture production expenses. However, traditional feeding methods remain heavily dependent on manual operations, leading to imprecise feeding control, significant waste (15–25%) and environmental pollution. To accurately identify fish feeding behavior and mitigate feed waste, along with associated water pollution issues, this study proposes an intelligent feeding behavior recognition system based on an optimized YOLOv8n model. A dedicated dataset comprising 5102 annotated images was constructed and processed through standardized acquisition, pre-processing, and labeling procedures. The enhanced YOLOv8n architecture incorporates three key modifications: (1) Adaptive Kernel Convolution (AKConv) in the Backbone for superior multiscale feature extraction, (2) a Squeeze-and-Excitation (SE) attention mechanism to enhance feature representation, and (3) a Bidirectional Feature Pyramid Network (BiFPN) replacing PANet for efficient feature fusion. Experimental results demonstrate substantial performance improvements: precision increased from 90.1% to 92.7% (+2.9%), recall improved from 83.2% to 88.8% (+6.7%), and mean average precision (mAP) rose from 87.0% to 91.7% (+5.4%) compared to the pre-optimized YOLOv8n. Additionally, the optimized model exhibits reduced computational demands, with parameters decreasing from 3.01M to 2.47M, FLOPs dropping from 8.2G to 5.2G, and model size shrinking from 6.2MB to 5.4MB. Ablation studies further validate the individual contributions of each enhancement. This approach not only enhances feeding precision and reduces waste but also promotes sustainable aquaculture practices, demonstrating significant potential for real-world deployment.
Shi et al. (Sun,) studied this question.