LEM-YOLO: A Lightweight Cow Behaviour Detection Algorithm based on YOLOv8

Nowadays, precision livestock farming, driven by intelligent equipment, is becoming the mainstream model in modern animal husbandry. However, existing models struggle to balance complexity and accuracy: high-precision models are computationally intensive, while lightweight models tend to be less accurate. To address this, we propose LEM-YOLO, a lightweight cattle detection model based on YOLOv8. It reduces computational redundancy through an improved SCConv module, enhances feature extraction with the EMA multi-scale attention mechanism, and optimizes bounding box regression using MPDIoU loss. A multi-detection head structure further improves multi-scale target detection. Experimental results show that LEMYOLO reduces parameters and GFLOPs by $\mathbf{2 4. 2 \%}$ and $\mathbf{2 6. 8 \%}$, respectively, while increasing precision and recall. This demonstrates its effectiveness in achieving a better trade-off between performance and efficiency, making it ideal for real-time cow behavior recognition in resource-constrained environments.