欢迎访问《茶叶科学》,今天是

茶叶科学 ›› 2022, Vol. 42 ›› Issue (4): 549-560.doi: 10.13305/j.cnki.jts.2022.04.009

• 研究报告 • 上一篇    下一篇

基于改进YOLOv4-tiny的茶叶嫩芽检测模型

方梦瑞1, 吕军1,*, 阮建云2, 边磊2, 武传宇3, 姚青1   

  1. 1.浙江理工大学信息学院,浙江 杭州 310018;
    2.中国农业科学院茶叶研究所,浙江 杭州 310008;
    3.浙江理工大学机械与自动控制学院,浙江 杭州 310018
  • 收稿日期:2022-05-09 修回日期:2022-06-09 出版日期:2022-08-15 发布日期:2022-08-23
  • 通讯作者: *lv_jun@zstu.edu.cn
  • 作者简介:方梦瑞,男,硕士研究生,主要从事农业智能信息研究,fmengrui@163.com。
  • 基金资助:
    财政部和农业农村部:国家现代农业产业技术体系(CARS-19)、浙江省领雁计划项目(2022C02052)

Tea Buds Detection Model Using Improved YOLOv4-tiny

FANG Mengrui1, LÜ Jun1,*, RUAN Jianyun2, BIAN Lei2, WU Chuanyu3, YAO Qing1   

  1. 1. School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
    2. Tea Research Institute, Chinese Academy of Agricultural Sciences, Hangzhou 310008, China;
    3. School of Mechanical Engineering and Automation, Zhejiang Sci-Tech University, Hangzhou 310018, China
  • Received:2022-05-09 Revised:2022-06-09 Online:2022-08-15 Published:2022-08-23

摘要: 精准检测茶叶嫩芽是茶叶机械智能采摘的重要前提。针对茶叶大小不一、遮挡造成的小尺度嫩芽特征显著性弱、漏检率高等问题,提出一种基于改进YOLOv4-tiny的茶叶嫩芽检测模型。该模型在颈部网络添加52×52的浅层特征层以提高YOLOv4-tiny网络对小目标嫩芽的关注度,通过引入卷积块注意力机制(Convolutional block attention module,CBAM)以抑制背景噪声,提高嫩芽特征的显著性,采用双向特征金字塔网络(Bidirectional feature pyramid network,BiFPN)以融合不同尺度的特征信息,从而提出一个高性能轻量化的茶叶嫩芽检测模型YOLOv4-tiny-Tea。对同一训练集与测试集进行模型训练与性能测试,结果表明YOLOv4-tiny-Tea模型检测精确率和召回率分别为97.77%和95.23%,相比改进之前分别提高了5.58个百分点和23.14个百分点。消融试验验证了网络结构改进对不同尺度嫩芽检测的有效性,并将改进后的YOLOv4-tiny-Tea模型与3种YOLO系列算法进行对比,发现改进后的YOLOv4-tiny-Tea模型F1值比YOLOv3、YOLOv4、YOLOv5l模型分别提高了12.11、11.66和6.76个百分点,参数量仅为3种网络模型的13.57%、13.06%和35.05%。试验结果表明,YOLOv4-tiny-Tea模型能有效提高不同尺度下嫩芽检测的精确率,大幅度减少小尺寸或遮挡嫩芽的漏检情况,在保持轻量化计算成本的基础上获得较为明显的检测精度,能够满足农业机器人的实时检测和嵌入式开发的需求,可以为茶叶嫩芽智能采摘方法提供参考。

关键词: 茶叶, 嫩芽检测, YOLOv4-tiny, 注意力机制, 双向特征金字塔

Abstract: Precise detection of tea buds is a prerequisite for intelligent mechanical picking of tea. Aiming at the problems of poor salience and high missed detection rate of small-scale buds caused by different sizes of tea leaves and the cover of other tea leaves, this paper proposed a kind of tea buds detection model based on improved YOLOv4-tiny. In this model, a 52×52 shallow feature layer was added in the neck network to promote the attention of YOLOv4-tiny network to small target buds. A convolutional block attention module (CBAM) was introduced to suppress the background noise and improve the salience of buds, and a bidirectional feature pyramid network (BiFPN) was used to integrate characteristic information of different scales, so as to propose the YOLOv4-tiny-Tea, a high performance light weight tea buds detection model. The results of model training and performance testing on the same training set and test set show that for the YOLOv4-tiny-Tea model, the detection precision and recall rate were 97.77% and 95.23% respectively, which were 5.58% and 23.14% higher than those before modification. An ablation experiment verified the effectiveness of the modified network structure in detecting different scales of buds, and a comparison of YOLOv4-tiny-Tea model with three YOLO algorithms found that the F1 value of YOLOv4-tiny-Tea model was 12.11%, 11.66% and 6.76% higher than F1 values of YOLOv3, YOLOv4 and YOLOv5l models respectively. The number of parameters in YOLOv4-tiny-Tea model was merely 13.57%, 13.06% and 35.05% of the three network models. The experimental results demonstrate that the method proposed in this paper effectively improved the detection precision of buds under different scales, greatly reduced the missed detection rate of buds for small size or under shading, and significantly bettered the detection precision based on a lightweight computation overhead. Therefore, the method can meet the needs of agricultural robots for real-time detection and embedded development, thus providing a reference for intelligent tea buds picking.

Key words: tea, tea buds detection, YOLOv4-tiny, attention mechanism, bidirectional feature pyramid

中图分类号: