改进YOLOv5在胡瓜钝绥螨品质管控系统中的设计与应用  

Design and application of improved YOLOv5 in the quality control system of Amblyseius cucumeris Oudemans

作　　者：李建兴[1,2] 刘振宇 马莹 张艳璇[3,4] 宋江 纪茂源[1] 旷树森 LI Jianxing;LIU Zhenyu;MA Ying;ZHANG Yanxuan;SONG Jiang;JI Maoyuan;KUANG Shusen(School of Electronics,Electrical and Physics,Fujian University of Technology,Fuzhou 350118,China;Fujian Industrial Integrated Automation Industry Technology Development Base,Fuzhou 350118,China;Institute of Plant Protection,Fujian Academy of Agricultural Sciences,Fuzhou 350013,China;Fujian Yanxuan Biological Control Technology Co.,Ltd.,Fuzhou 350013,China)

机构地区：[1]福建工程学院电子电气与物理学院,福建福州350118 [2]福建省工业集成自动化行业技术开发基地,福建福州350118 [3]福建省农业科学院植物保护研究所,福建福州350013 [4]福建艳璇生物防治技术有限公司,福建福州350013

出　　处：《南昌大学学报（工科版）》2023年第1期85-95,共11页Journal of Nanchang University(Engineering & Technology)

基　　金：高校产学合作项目(2021N5004);福建省自然科学基金项目(2020J01876)。

摘　　要：当前工业化养殖胡瓜钝绥螨的产量巨大,导致现有人工品质管控方式无法及时准确地对螨虫养殖品质进行监测。设计研发一款螨虫监控系统,包含螨虫麦麸分离装置、显微视觉装置、改进的YOLOv5螨虫检测模型。通过对YOLOv5模型的网络结构进行分析,在原YOLOv5模型上将嵌入ECANet注意力机制的MobileNetv3-Large网络作为主干特征提取网络,加强网络提取特征能力,进一步通过裁剪模型预测头的冗余部分,提升模型对中小目标检测性能。实验结果表明:改进后的M3ECA-YOLOv5-2H模型对比原YOLOv5模型,在平均检测精度mAP 50和mAP 75上提高了0.68和4.62个百分点;在模型大小上降低18.9 MB,占用更低内存;在单张图片推理速度上提升4 ms。此外对比其他5种通用目标检测模型,M3ECA-YOLOv5-2H模型在检测精度和检测速度等指标上整体表现更佳,具有很好的应用价值。The huge output of the current industrialized farming of Amblyseius cucumeris(Oudemans)makes the existing manual quality control methods unable to monitor the quality of mite breeding in a timely and accurate manner,a mite monitoring system was designed and developed,including a mite wheat bran separation device,a microscopic vision device,and improved YOLOv5 mites detection model.By analyzing the network structure of the YOLOv5 model,on the original YOLOv5 model,the MobileNetv3-Large network embedded with the ECANet attention mechanism is used as the backbone feature extraction network to enhance the network’s ability to extract features.The redundant part of the prediction head is further trimmed and improved.The detection performance of the model for small and medium targets.The experimental results show that compared with the original YOLOv5 model,the improved M3ECA-YOLOv5-2H model improves the average detection accuracy mAP 50 and mAP 75 by 0.68 and 4.62 percentage points;it reduces the model size by 18.9 MB and occupies lower memory;The image inference speed is improved by 4 milliseconds.In addition,compared with the other five general target detection models,the M3ECA-YOLOv5-2H model has better overall performance in terms of detection accuracy and detection speed,and has good application value.

关 键 词：螨虫 监控系统 注意力机制 检测计数 GUI监控界面 改进的YOLOv5 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S476[自动化与计算机技术—计算机科学与技术] S126[农业科学—农业昆虫与害虫防治]