Pigs’ facial regions encode a wealth of biological trait information; detecting their facial poses can provide robust support for individual identification and behavioral analysis. However, in large-scale swine-herd settings, variable lighting within pigsties and the close proximity of animals impose significant challenges on facial-pose detection. This study adopts an aerial-inspection approach—distinct from conventional ground or hoist inspections—leveraging the high-efficiency, panoramic coverage of unmanned aerial vehicles (UAVs). UAV-captured video frames from real herding environments, involving a total of 600 pigs across 50 pens, serve as the data source. The final dataset comprises 2800 original images, expanded to 5600 after augmentation, to dissect how different facial expressions reflect pig physiological states.ion. 1. The proposed Detect_FASFF detection head achieves mean average precision at 50% IoU (mAP₅₀) of 95.5% for large targets and 95.9% for small targets, effectively overcoming cross-scale feature loss and the accuracy shortcomings of the baseline YOLOv8s in detecting pig facial targets. 2. Addressing the excessive computation and sluggish inference of standard YOLOv8, we incorporate a Partial_Conv module that maintains mAP while substantially reducing Runtime. 3. We introduce an improved exponential moving-average scheme (iEMA) with second-order attention to improve small-target accuracy and mitigate interference from the piggery environment. This yields an mAP₅₀ of 96.4%. 4. Comprehensive comparison–The refined YOLOv8 is benchmarked against traditional YOLO variants (YOLOv5s, YOLOv8s, YOLOv11s, YOLOv12s, YOLOv13s), Rt-DeTR, and Faster-R-CNN. Relative to these models, the enhanced YOLOv8 shows a statistically significant increase in overall mAP. These results highlight the potential of the upgraded model to transform pig facial-expression recognition accuracy, advancing more humane and informed livestock-management practices.