Explainable Pruning-Focused Attention Network for High-Fidelity Livestock Facial Detection and Identification

Livestock facial detection plays a vital role in precision agriculture, supporting automated health monitoring, behavior analysis, and welfare assessment. However, unconstrained farm environments introduce significant challenges, including occlusion, pose variation, background clutter, and illumination inconsistency, which hinder reliable detection. Furthermore, ondevice deployment requires models that balance accuracy with computational efficiency. This work presents a novel Pruning-Guided YOLOv5(PG-YOLOv5), a lightweight detection framework optimized for livestock facial analysis. The framework integrates a StrucFocus-Channel Attention(SFCA)to enhance discriminative feature representation and a Attention-Induced Channel Pruning(AICP)mechanism to eliminate redundant channels. A Learnable Robust Attention Loss(LRAL)further promotes sparsity in attention responses, guiding pruning toward informative regions. For detection,PG-YOLOv5achieves 99.5% mAP50and 75.9% mAP50–95on the Pig Face Computer Vision Dataset (PCVD) with only 9.01 M parameters and 23.4 GFLOPs. When adapted for individual pig classification on the custom IPD-1k dataset, the model attains 98.21% Top-1 accuracy with a lightweight classification head, outperforming recent state-of-the-art methods while maintaining computational complexity suitable for real-time livestock monitoring. The code for the model is available at https://github.com/AbhiKandS/PGYOLOv5a.