摘要
[目的]建立一种水果外观无损检测方法。[方法]以油桃为研究对象,使用IQQU3手机相机采集图像数据,并进行图像预处理。使用LabelImg图像标注软件标注数据。采用镜像翻转、左右翻转、平移等方法对数据进行扩充。将扩充后的图像按照8∶2划分为训练集和测试集。最后使用YOLOv8(n,s,m,l,x)5个模型对数据进行训练,训练150轮,比较分析5个模型的训练结果,选出最优的检测模型。[结果]构建了油桃数据集,共4205张图像;YOLOv8(n,s,m,l,x)训练集总损失值分别为2.275,1.778,1.482,1.880,1.401,测试集的总损失值分别为2.724,2.253,2.057,2.105,2.004;YOLOv8(n,s,m,l,x)的精确率分别为94.0%,98.0%,97.4%,97.3%,97.9%,召回率分别为95.4%,95.5%,95.9%,96.9%,96.9%。综合比较,YOLOv8s为较优的模型,其平均检测精确率达97.8%,正常、损伤、疤痕的平均精确率分别为96.2%,98.8%,98.4%,其推理时间、计算量(GFLOPs)分别为179.4 ms、28.4。[结论]YOLOv8能够有效地检测水果外观品质,可用于水果外观的无损检测。
[Objective]To establish a nondestructive detection method for fruit appearance.[Methods]Nectarines were used as the research subject.The IQQU3 smart phone camera was used to capture the picture data,which was then preprocessed.The image annotation program Labelimg was used to label the data.Panning,left-right flipping,and mirror flipping were used to enlarge the data.Using an ratio of 8∶2,the enlarged photos were split into training and test sets.Lastly,the data was trained for 150 epochs using five YOLOv8 models(n,s,m,l,x).The training results of the five models were compared and analyzed in order to determine which detection model was the best.[Results]The nectarine dataset was constructed,there were 4,205 total photos;YOLOv8(n,s,m,l,x)the total loss values in the training set were 2.275,1.778,1.482,1.880,and 1.401,respectively,The total loss values of the test set were 2.724,2.253,2.057,2.105,and 2.004,respectively;YOLOv8(n,s,m,l,x)precision were 94.0%,98.0%,97.4%,97.3%,97.9%,respectively,The recall were 95.4%,95.5%,95.9%,96.9%,and 96.9%,respectively.In a comprehensive comparison YOLOv8s was the better model,and the average detection accuracy mAP_0.5 was 97.8%.The average precision of fresh,bruise and scar were 96.2%,98.8%and 98.4%,respectively.The inference time and calculation amount(GFLOPs)were 179.4 ms and 28.4 respectively.[Conclusion]YOLOv8 can effectively detect the quality of fruit appearance,which can be used for non-destructive testing of fruit appearance,and this study can provide new ideas for non-destructive testing of fruits.
作者
唐兴萍
王白娟
杨红欣
杨正明
李汝嘉
吴文斗
TANG Xingping;WANG Baijuan;YANG Hongxin;YANG Zhengming;LI Rujia;WU Wendou(College of Food Science and Technology,Yunnan Agricultural University,Kunming,Yunnan 650201,China;College of Tea Science,Yunnan Agricultural University,Kunming,Yunnan 650201,China;College of Big Data,Yunnan Agricultural University,Kunming,Yunnan 650201,China)
出处
《食品与机械》
CSCD
北大核心
2024年第7期103-110,共8页
Food and Machinery
基金
云南省重大科技专项计划项目(编号:A303202324600101)
科技创新项目(编号:S9032023111)。
关键词
YOLOv8
油桃
外观品质
检测分类
YOLOv8
nectarine
fruit appearance quality
testing and classification
