基于改进YOLOv8n轻量化的人参外观质量精准识别  

Accurate recognition of ginseng appearance quality based on improved YOLOv8n lightweighting

作　　者：张丽娟 游浩海 李芝贻 魏湛郴 贾浩杰 于跃 李东明 ZHANG Lijuan;YOU Haohai;LI Zhiyi;WEI Zhanchen;JIA Haojie;YU Yue;LI Dongming(School of Internet of Things Engineering,Wuxi University,Wuxi 214105,China;School of Information Technology,Jilin Agricultural University,Changchun 130118,China;College of Instrument Science and Electrical Engineering,Jilin University,Changchun 130012,China)

机构地区：[1]无锡学院物联网工程学院,无锡214105 [2]吉林农业大学信息技术学院,长春130118 [3]吉林大学仪器科学与电气工程学院,长春130012

出　　处：《农业工程学报》2024年第24期274-282,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金(62206257);无锡学院引进人才研究创业基金(2023 r004,2023 r006,2023 r005);吉林省科技发展计划重点研发项目(20210204050YY)。

摘　　要：针对人参分级特征微小差异对专业人员依赖性强的问题,同时为了降低人工劳动强度,提高人参外观质量检测精度,实现将模型方便快速部署到移动端,该研究提出了一种基于改进YOLOv8n的轻量级方法。该模型(简称为CGC-YOLOv8模型)首先将YOLOv8n的骨干网络的卷积替换为条件参数化卷积(conditional convolutional networks),这使模型能够根据输入数据的特征进行调整;其次,在颈部网络,引入一种细颈组合(GSconv+VoVGSCSP)降低模型的参数量和模型尺寸的同时,进一步提升模型检测性能;最后,加入坐标注意力机制(coordinate attention)模块,有效地整合了空间坐标信息到生成的注意力图中,使模型细化特征更加关注人参外观特征,从而提高了特征提取的能力。试验证明,在使用分级良好的人参批次采集的数据集上,该研究的CGC-YOLOv8模型在关键评价指标上,精确度、召回率和平均精度均值(mAP50、mAP50-95),分别达到85.70%、91.23%、94.69%、72.43%,与原模型YOLOv8n相比较分别提升了2.74、4.64、3.71和4.9个百分点,与原YOLOv8n模型相比有显著提高,而且在模型大小和计算浮点数明显减少。研究结果可为人参等药材外观质量精准识别提供参考。Ginseng has been regarded as the top of conventional Chinese medicine.Its root can be widely used as an herbal medicine with a long history and multiple medicinal values;Thus,it is very crucial to monitor and manage ginseng quality.In this study,a lightweight CGC-YOLOv8 model was introduced into the detection system,in order to classify and assess the appearance quality of ginseng.Ginseng was harvested in Fusong County,Jilin Province,China.An experimental sample was then selected in the current year.Data enhancement techniques were used to simulate the various typical environments.In the basic network design of the model,the regular convolutional layer was replaced with two conditional convolutional layers(CondConv)in the backbone,according to the architecture of the YOLOv8 model.The appearance feature of input ginseng was highlighted to enhance the accuracy and efficiency of feature extraction.In addition,a thin neck combination(GSconv+VoVGSCSP)in the neck structure was applied to lighten the model.The coordinate Attention mechanism(CA)was incorporated to detect the head of the model.The performance was effectively improved without increasing the computational burden,particularly suitable for resource-limited devices.The CGC-YOLOv8 model was then optimized by ablation experiments.The better performance was achieved,with 85.70%precision,a recall of 91.23%,IoU=0.50 mean average precision(mAP50)of 94.69%,and IoU=0.95 mean average precision(mAP50-95)of 72.43%.Furthermore,the CGC-YOLOv8 improved the precision by 2.74 percentage points,recall by 4.64 percentage points,mAP50 by 3.71 percentage points,and mAP50-95 by 4.09 percentage points,respectively,compared with the YOLOv8n.The improved model outperformed the original one when detecting images without training.In addition,the number of parameters was reduced by 6.39%in the CGC-YOLOv8 model,compared with the original.The weight size was only 6.0 MB,fully meeting the lightweight condition and easy to deploy on mobile devices.A series of experiments were conducted to c

关 键 词：人参 外观质量 YOLOv8 目标检测 注意力机制 深度学习 图像识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]