基于YOLOV3优化模型的复杂场景下茶树嫩芽识别  

Tea buds recognition under complex scenes based on optimized YOLOV3 model

作　　者：张晴晴 刘连忠[1] 宁井铭[2,3] 吴国栋 江朝晖[1] 李孟杰[1] 李栋梁 ZHANG Qingqing;LIU Lianzhong;NING Jingming;WU Guodong;JIANG Zhaohui;LI Mengjie;LI Dongliang(College of Information and Computer Science,Anhui Agricultural University,Hefei 230036,China;College of Tea and Food Science and Technology,Anhui Agricultural University,Hefei 230036,China;State Key Laboratory of Tea Plant Biology and Resource Utilization,Hefei 230036,China)

机构地区：[1]安徽农业大学信息与计算机学院,安徽合肥230036 [2]安徽农业大学茶与食品科技学院,安徽合肥230036 [3]茶树生物学与资源利用国家重点实验室,安徽合肥230036

出　　处：《浙江农业学报》2021年第9期1740-1747,共8页Acta Agriculturae Zhejiangensis

基　　金：茶树生物学与资源利用国家重点实验室开放基金(SKLTOF20160202);安徽省高校自然科学研究项目重点项目(KJ2017A151);安徽省高校自然科学研究重大项目(KJ2019ZD20);国家重点研发计划(2016YFD0200900)。

摘　　要：茶叶智能采摘的关键技术之一是待采摘嫩芽的识别,而嫩芽大小、环境光照、拍摄角度等因素都会给嫩芽的精准识别带来困难。针对复杂场景下传统茶树嫩芽识别方法准确率低的问题,文章提出一种基于YOLOV3深度卷积模型的识别方法,并通过增加SPP模块优化模型,提高模型对茶树嫩芽的识别能力。实验结果表明,YOLOV3模型和YOLOV3优化模型均能在复杂场景下实现茶树嫩芽识别,且YOLOV3优化模型的平均精度均值mAP比YOLOV3模型提高3.5百分点,达到91%,说明YOLOV3优化模型能够更好地应用于自然场景下的茶树嫩芽识别。Tea buds recognition is one of the key technologies for intelligent tea buds picking,and the size of buds,environmental lighting,imaging angle and other factors will bring difficulties for precise buds identification.In order to solve the problem of low accuracy of traditional tea buds recognition in complex scenes,a tea buds recognition method based on YOLOV3 deep convolution model was proposed.By adding a SPP module into the YOLOV3 model,an optimized YOLOV3 model was designed to further improve the recognition ability of tea buds.The results showed that YOLOV3 and optimized YOLOV3 models could both realize tea buds recognition in complex scenes,and the average accuracy mean(mAP)of the optimized YOLOV3 model reached 91%,which was 3.5 percentage points higher than YOLOV3 model,indicating the optimized YOLOV3 could be well applied to the tea buds identification in natural environments.

关 键 词：深度学习 卷积神经网络 茶树嫩芽识别 

分 类 号：S571.1[农业科学—茶叶生产加工] S24[农业科学—作物学]