基于SPR-Net的哺乳母猪分娩前行为识别方法

Recognition of Pre-farrowing Behavior in Sows Based on SPR-Net

摘要:

母猪分娩前行为精准识别对于实现分娩预测有着重要意义。 提出了一种面向母猪姿态识别的 SPR-Net 网络,该网络使用 HRNet 作为骨干网络,在颈部融合特征金字塔网络与基于令牌的两阶段交叉注意力模块,同时在损失函数中引入关键点头损失作为归纳偏置,并通过非对称加权对分类头施加约束。 在测试集上,该网络对母猪 4 种姿态分类 Top-1 准确率达到 95. 6% 。 网络完成性能验证后,对不同胎次母猪分娩前行为进行探究。 研究发现,低胎次组和高胎次组母猪分娩前姿态转变频率呈现阶段性增长趋势;且低胎次组母猪在 Early 阶段、Late 阶段的姿态转变频率显著高于高胎次组母猪,而在 Mid 阶段两组的姿态转变频率差异不显著。 此外,低胎次组母猪和高胎次组母猪在分娩前表现出不同的姿态转变频率变化范围,这些行为特征可作为一种临产特征,帮助养殖场及时采取相应措施。 SPR-Net 网络在猪只分娩前行为识别中表现出优良性能,并通过视频分析揭示出不同胎次母猪的产前行为规律,为实现哺乳母猪分娩精准预测提供有力的技术支持和理论依据。

Abstract:

Accurate recognition of prepartum behaviors in sows is of great significance for farrowing prediction. An SPR-Net network was proposed for sow posture recognition. The network employed HRNet as the backbone and integrated a feature pyramid network with a token-based two-stage cross- attention module in the neck. Meanwhile, a keypoint head loss was introduced into the loss function as an inductive bias, and asymmetric weighting was applied to constrain the classification head. On the test set, SPR-Net achieved a Top-1 accuracy of 95. 6% for the classification of four sow postures. After performance validation, the network was further used to investigate prepartum behavioral patterns of sows with different parities. The results showed that the posture transition frequency of both low-parity and high-parity sows exhibited a staged increasing trend before farrowing. Moreover, the posture transition frequency of low-parity sows was significantly higher than that of high-parity sows during the early and late prepartum stages, whereas no significant difference was observed between the two groups during the middle stage. In addition, low-parity and high-parity sows showed different ranges of posture frequency variation before farrowing. These behavioral characteristics can serve as indicators of impending parturition, helping farms take timely management measures. Overall, SPR-Net demonstrated excellent performance in prepartum behavior recognition of sows, and video-based analysis revealed parity-related prepartum behavioral patterns. The research result can provide strong technical support and theoretical evidence for achieving accurate farrowing prediction in lactating sows.