基于YOLOv5改进模型的柑橘果实识别方法  

Citrus fruit recognition method based on the improved model of YOLOv5

作　　者：黄彤镔 黄河清 李震 吕石磊 薛秀云 代秋芳 温威 HUANG Tongbin;HUANG Heqing;LI Zhen;LÜShilei;XUE Xiuyun;DAI Qiufang;WEN Wei(College of Electronic Engineering(College of Artificial Intelligence),South China Agricultural University,Guangzhou 510642,China;Division of Citrus Machinery,China Agriculture Research System,Guangzhou 510642,China;Guangdong Engineering Research Center for Monitoring Agricultural Information,Guangzhou 510642,China;Guangdong Artificial Intelligence and Digital Economy Laboratory(Guangzhou),Guangzhou 510330,China;Engineering Fundamental Teaching and Training Center,South China Agricultural University,Guangzhou 510642,China)

机构地区：[1]华南农业大学电子工程学院(人工智能学院),广州510642 [2]国家柑橘产业技术体系机械化研究室,广州510642 [3]广东省农情信息监测工程技术研究中心,广州510642 [4]人工智能与数字经济广东省实验室(广州),广州510330 [5]华南农业大学工程基础教学与训练中心,广州510642

出　　处：《华中农业大学学报》2022年第4期170-177,共8页Journal of Huazhong Agricultural University

基　　金：国家重点研发计划(2020YFD1000107);国家自然科学基金项目(31971797);国家现代农业产业技术体系建设专项(CARS–26);广东省科技厅项目(2021A1515010923);广东省省级乡村振兴战略专项(粤财农[2021]37号);广东省大学生科技创新培养专项(pdjh2020a0083)。

摘　　要：为实现在自然环境下对柑橘果实的识别,提出一种基于YOLOv5改进模型的柑橘识别方法。通过引入CBAM(convolutional block attention module,卷积注意力模块)注意力机制模块来提高网络的特征提取能力,改善遮挡目标与小目标的漏检问题;采用α-IoU损失函数代替GIoU损失函数作为边界框回归损失函数,提高边界框定位精度。结果显示:本研究提出的模型平均精度AP值达到91.3%,在GPU上对单张柑橘果实图像的检测时间为16.7 ms,模型占用内存为14.5 Mb。结果表明,本研究基于YOLOv5的改进算法可实现在自然环境下快速准确地识别柑橘果实,满足实时目标检测的实际应用需求。The rapid and accurate identification of citrus fruit is of great significance for the realization of automatically picking citrus in orchards,the prediction of citrus yield and the intelligent management of citrus industry.A citrus recognition method based on the improved model of YOLOv5 was proposed to realize the recognition of citrus fruits in natural environment.The feature extraction ability of the network and the problem of missed detection of occluded targets and small targets was improved by introducing the CBAM attention mechanism module.Theα-IoU loss function instead of the GIoU loss function was used as the bounding box regression loss function to improve the positioning accuracy of the bounding box.The results showed that the average accuracy AP value of the proposed model reached 91.3%,with the detection time of a single citrus fruit image on the GPU of 16.7 ms and the model occupying 14.5 Mb of memory.It is indicated that the algorithm improved can quickly and accurately identify citrus fruits in the natural environment,meeting the practical application requirements of real-time target detection.It will provide new ideas for the intelligent citrus industry.

关 键 词：YOLOv5 柑橘识别 自动采摘 CBAM 损失函数 注意力机制 α-IoU 

分 类 号：S661.1[农业科学—果树学] TP391.41[农业科学—园艺学]