Lightweight multiscale behavior recognition for caged laying hens using an enhanced YOLOv8 framework.

This study presents a lightweight deep-learning framework for recognizing key health-related behaviors of caged laying hens, addressing the challenges of dense housing, frequent occlusions, and subtle action differences. Building upon YOLOv8n (serving as our base model), we introduce three major enhancements: (1) a C2f-FasterNet-EMA backbone that improves multiscale feature extraction; (2) a FasterNet-based neck combined with the Dysample upsampler to refine small-object localization while reducing computational cost; (3) an EMASlideLoss function that alleviates sample imbalance and stabilizes the training process. Evaluated on the Lukou Dataset, which includes four target behaviors (eating, open-mouth breathing, self-pecking, and mutual pecking), the improved model achieves mAP@50 scores of 98.15%, 81.03%, 93.65%, and 94.32% for each behavior, respectively. Overall, compared with the retrained baseline YOLOv8n under identical experimental settings, the proposed method attains a 2.26% improvement in overall mAP@50 while reducing the model size by 22.92% relative to the baseline.