基于Faster R-CNN复杂背景的茶芽检测  

Tea Bud Detection Based on Faster R-CNN in Complex Background

作　　者：李赫 王玉[1] 范凯 丁兆堂[1,2] LI He;WANG Yu;FAN Kai;DING Zhaotang(College of Horticulture,Qingdao Agricultural University,Qingdao 266109,China;Tea Research Institute,Shandong Academy of Agricultural Sciences,Jinan 250108,China)

机构地区：[1]青岛农业大学园艺学院,山东青岛266109 [2]山东省农业科学院茶叶研究所,山东济南250108

出　　处：《青岛农业大学学报（自然科学版）》2022年第3期212-216,共5页Journal of Qingdao Agricultural University(Natural Science)

基　　金：山东省农业技术创新重大应用项目(SD2019ZZ010);山东省现代农业产业技术体系(SDAIT-19-01);山东省泰山学者专项基金项目(ts201712057);青岛市民生工程项目(19-6-1-64-nsh);山东省农业科技基金项目(2019LY002;2019YQ010;2019TSLH0802)。

摘　　要：茶芽检测是判断茶树农艺性状的基础,也是研发基于计算机视觉采茶机器人的基础。针对复杂背景中传统茶芽检测方法准确率低、稳定性差等问题,提出一种基于深度学习的茶芽检测方法。以Faster R-CNN(region-convolutional neural network)算法为框架,比较AlexNet、ResNet50、VGG193种网络模型茶芽检测性能,寻找最佳网络模型。结果表明,使用VGG19的茶芽检测准确率为86.3%,召回率为96.1%,F1分数为0.909,综合检测效果最优。该方法可很好地应用于复杂背景茶芽检测。Tea bud detection is the basis for judging the agronomic traits of tea trees,and it is also the basis for developing a tea picking robot based on computer vision.Aiming at the problems of low accuracy and poor stability of traditional tea bud detection methods in complex background,a tea bud detection method based on deep learning is proposed.Taking the Faster R-CNN(region-convolutional neural network)algorithm as the framework,the tea bud detection performance of AlexNet,ResNet50 and VGG19 models were compared to find the best model.The results showed that the accuracy rate of tea bud detection using VGG19 model was 86.3%,the recall rate was 96.1%,the F1 score was 0.909,and the comprehensive detection effect was the best.This method can be well applied to tea bud detection with complex background.

关 键 词：茶树 茶芽检测 深度学习 Faster R-CNN VGG19 

分 类 号：S609.1[农业科学—园艺学] TP391.4[自动化与计算机技术—计算机应用技术]