基于改进YOLO v8s的母猪分娩结束识别方法

Method for Identifying End of Sow Farrowing Based on Improved YOLO v8s Model

摘要:

在母猪分娩自动监测技术中,一般采用识别最后一头仔猪产出为标志的间接方法来判断分娩结束,该方法在实际应用中存在实时性较差、识别精度不高和易受遮挡干扰等问题,难以满足生产需求。 针对上述问题,以胎衣为直接识别目标,提出了一种基于改进 YOLO v8s 的分娩结束识别方法。 通过引入 Focus 模块、 SE 注意力机制、 C2f-SCConv 模块及 BiFPN-P2 结构,对模型在特征表达、多尺度信息融合及轻量化设计等方面进行了优化。 消融试验结果表明,SE 注意力机制与 C2f-SCConv 模块显著提升了小尺度、低对比度目标的检测精度,BiFPN-P2 结构在保证精度的同时有效降低了参数量和计算复杂度。 最终,融合多模块的改进 YOLO v8s 模型精确率、召回率和 mAP 分别达到 98. 1% 、94. 9% 和 98. 8% ,参数量仅 7. 33 × 10 6 ,推理速度仍保持 61. 71 f / s。 与 NanoDet、RT-DETR、 Faster R-CNN、YOLO v5 及原始 YOLO v8s 等主流模型相比,本方法在检测精度方面表现最优,同时展现了良好的实时性。 提出了基于视频帧间隔和帧级计数器的时序判断机制,在真实分娩视频中实现了分娩结束事件的准确识别。 在 25 f 间隔条件下,平均时间误差为 5. 98 s,在 5 f 间隔条件下进一步缩小至 1. 84 s,显著提高了识别时效性与准确性。 研究表明,本研究提出的方法突破了依赖仔猪目标的间接推断模式,实现了由图像级检测向视频级时间节点识别的拓展。

Abstract:

In automatic monitoring of sow farrowing, the end of parturition is usually determined indirectly by identifying the birth of the last piglet. However, this approach suffers from poor real-time performance, low detection accuracy, and strong susceptibility to occlusion, making it unsuitable for production needs. To address these issues, a farrowing-end detection method that directly identified the placenta was proposed based on an improved YOLO v8s model. By incorporating the Focus module, SE attention mechanism, C2f-SCConv module, and BiFPN-P2 structure, the model was optimized in feature representation, multi-scale information fusion, and lightweight design. Ablation experiments showed that the SE attention mechanism and C2f-SCConv module significantly improved the detection accuracy of small-scale and low-contrast targets, while the BiFPN-P2 structure effectively reduced the number of parameters and computational complexity without sacrificing precision. The improved YOLO v8s model achieved 98. 1% precision, 94. 9% recall, and 98. 8% mAP, with only 7. 33 × 10 6 parameters and an inference speed of 61. 71 f / s. Compared with mainstream models such as NanoDet, RT-DETR, Faster R-CNN, YOLO v5, and the original YOLO v8s, the proposed method achieved the best detection accuracy while maintaining excellent real-time performance. Furthermore, a temporal judgment mechanism based on video frame intervals and frame-level counters was developed to accurately recognize the end of farrowing in real farrowing videos. Under a 25 f interval, the average time error was 5. 98 s, which was further reduced to 1. 84 s at a 5 f interval, significantly improving both timeliness and accuracy. The proposed method overcame the limitations of piglet-based indirect inference and extended from image-level detection to video-level temporal event recognition.