摘要
养猪业是我国农业领域的重要组成部分,近年来规模化养殖场发展迅速。生猪姿态改变往往预示着健康状况变化或疾病发生,因此,实时监测生猪姿态可以帮助养殖户掌握猪只生长发育和健康状况,及时调整养殖方案或采取疾病防治措施,从而提高养殖效益并保障最终的猪肉产品质量,同时还可为生猪养殖产业分析研究提供数据支持。传统的监测方法主要依靠养殖户不定期的肉眼观察,耗时费力且无法满足实时需求,不适合规模化养殖场使用。计算机视觉技术的发展为实现生猪姿态的实时监测提供了技术手段。本研究基于YOLOv8s模型进行改进,提出一种生猪姿态识别模型RMAK-YOLOv8s。主要从三个方面进行改进:一是通过结构重参数化改进主干网络的C2f模块,实现隐式特征复用,达到模型轻量化及检测速度提高的目的;二是添加多尺度卷积注意力机制,用于捕捉多尺度特征图,加强有效特征的权重比例;三是使用可变核卷积代替标准卷积,获得更有效的特征信息,为平衡网络开销和性能提供更多选择。实验结果表明,与原始模型YOLOv8s相比,RMAK-YOLOv8s的参数量减少10.77%,计算量减少5.23%,平均精度均值mAP@0.5、mAP@0.5∶0.95分别达到93.7%、78.5%,分别提高1.7、1.3个百分点,能精确识别生猪姿态,可为实时监测生猪姿态及后续行为分析和健康管理等提供技术支撑。
The pig farming industry is an important component of China’s agriculture.In recent years,large-scale pig farms develope rapidly.Changes in the postures of pigs often indicate alterations in their health conditions or the occurrence of diseases.Therefore,real-time monitoring pig postures could help farmers mas-ter the growth and health status of pigs in order to adjust their farming plans or take disease prevention and control measures timely,thereby improving the efficiency of pig farming and ensuring the quality of the final pork products.It can also provide data support for analysis and research of the pig farming industry.Traditional monitoring methods mainly rely on the occasional visual observations of farmers,which are time-consuming and labor-intensive and can not meet the real-time requirements,making them unsuitable for large-scale pig farms.The development of computer vision technology has provided a technical means for the real-time monitoring of pig postures.This study improved the YOLOv8s model and proposed a pig posture recognition model called RMAK-YOLOv8s.The improvements were mainly made in three aspects:firstly,the C2f module of the back-bone network was improved through structural reparameterization to achieve implicit feature reuse,thereby a-chieving model lightweight and increasing the detection speed;secondly,a multi-scale convolutional attention mechanism was added to capture multi-scale feature maps and enhance the weight ratio of effective features;thirdly,variable kernel convolution was used instead of standard convolution to obtain more effective feature information,providing more options for balancing network cost and performance.The experimental results showed that compared with the original model YOLOv8s,RMAK-YOLOv8s reduced the number of parameters by 10.77%,reduced the computational cost by 5.23%,and achieved mAP@0.5 and mAP@0.5∶0.95 of 93.7%and 78.5%,which were 1.7 and 1.3 percentage points higher,respectively.It could accurately identify the postures of pigs and provide technical support for real-time monitoring of pig postures and subsequent be-havior analysis and health management.
作者
王鲁
朱永泉
王韵
刘瑞麟
唐辉
Wang Lu;Zhu Yongquan;Wang Yun;Liu Ruilin;Tang Hui(College of Information Science and Engineering,Shandong Agricultural University,Taian 271018,China;College of Animal Science and Technology,Shandong Agricultural University,Taian 271018,China)
出处
《山东农业科学》
北大核心
2025年第11期170-180,共11页
Shandong Agricultural Sciences
基金
山东省重点研发计划(重大科技创新工程)项目(2022LZGC003)
山东省自然科学基金面上项目(ZR2023MA011)。
关键词
生猪姿态识别
YOLOv8s
结构重参数化
多尺度卷积注意力
可变核卷积
Pig posture recognition
YOLOv8s
Structural reparameterization
Multi-scale convolutional attention
Variable kernel convolution
