基于改进YOLO v5复杂场景下肉鹅姿态的检测算法研究  

Research on detection algorithm about the posture of meat goose in complex scene based on improved YOLO v5

作　　者：刘璎瑛[1,2] 曹晅 郭彬彬 陈慧杰 戴子淳[3] 龚长万 LIU Yingying;CAO Xuan;GUO Binbin;CHEN Huijie;DAI Zichun;GONG Changwan(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210031,China;Key Laboratory of Breeding Equipment,Ministry of Agriculture and Rural Affairs,Nanjing 210031,China;Institute of Animal Science,Jiangsu Academy of Agricultural Sciences,Nanjing 210014,China)

机构地区：[1]南京农业大学人工智能学院,江苏南京210031 [2]农业农村部养殖装备重点实验室,江苏南京210031 [3]江苏省农业科学院畜牧研究所,江苏南京210014

出　　处：《南京农业大学学报》2023年第3期606-614,共9页Journal of Nanjing Agricultural University

基　　金：国家自然科学基金项目(31972551);江苏省博士后科研资助计划项目(2020Z396);国家现代农业产业技术体系岗位科学家资助项目(CARS-40-20)。

摘　　要：[目的]肉鹅姿态是预警肉鹅异常行为、评判肉鹅健康状态的重要指标,针对传统养殖场人工观察肉鹅姿态耗时费力且有很大主观性等问题,提出了一种基于深度学习模型自动识别肉鹅姿态的检测算法。[方法]利用YOLO v5模型对扬州鹅4种姿态(站立、休憩、饮水和梳羽)进行识别;对YOLO v5模型加入SENet、CBAM、ECA三种注意力模块改进网络结构,提高模型的识别能力;设计明暗试验和密集场景试验进一步验证模型在复杂场景下的鲁棒性。[结果]YOLO v5+ECA模型的平均检测精度(mAP)为88.93%,相比YOLO v5提升了2.27%。在识别精度(AP)上,站立姿态为91.85%,休憩姿态为93.42%,饮水姿态为90.02%,梳羽姿态为80.42%。在明暗试验和密集场景试验中,YOLO v5+ECA模型性能表现稳定,漏检现象和误检现象相对较少。[结论]该模型可以实现养殖场复杂场景下肉鹅姿态准确快速检测,为后续肉鹅行为监控和健康防疫提供数据支撑。[Objectives]The posture of meat goose is an important indicator for early warning of the meat goose’s abnormal behavior and evaluation of the meat goose’s health status.Since manual observation of the meat goose’s posture on traditional farms was time-consuming,laborious and highly subjective,a new method based on deep learning was proposed in this paper.The detection algorithm recognized the posture of meat goose automatically.[Methods]The YOLO v5 model was used to identify the four postures of Yangzhou goose(standing,resting,drinking and feather preening).Three attention modules including SENet,CBAM and ECA were added to the YOLO v5 model to improve the network structure and the recognition ability of the model.Light and dark experiments and dense scene experiments were designed to further verify the robustness of the model in complex scenes.[Results]The average detection accuracy(mAP)of the YOLO v5+ECA model was 88.93%,which was 2.27%higher than that of YOLO v5.In terms of recognition accuracy(AP),the standing posture was 91.85%,the resting posture was 93.42%,the drinking posture was 90.02%,and the feather preening posture was 80.42%.In the light and dark experiments and dense scene experiments,the performance of the YOLO v5+ECA model was stable,and the missing detection and false detection were relatively rare.[Conclusions]The posture of the meat goose can be detected rapidly and accurately in the farm complex scene with this model.The data can support the following research on the behavior monitoring and epidemic prevention of the meat goose.

关 键 词：深度学习 YOLO v5 扬州鹅 姿态识别 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S835[自动化与计算机技术—计算机科学与技术]