基于复杂背景下改进YOLOv8的茶叶嫩芽检测
Tea Bud Detection based on Improved YOLOv8 in Complex Backgrounds
董春旺 1王孟杰 2陈之威 3丁泽中 2田征瑞2
作者信息
- 1. 山东省农业科学院 茶叶研究所,山东 济南 250100
- 2. 山东省农业科学院 茶叶研究所,山东 济南 250100;石河子大学 机械电气工程学院,新疆维吾尔自治区 石河子 832003
- 3. 石河子大学 机械电气工程学院,新疆维吾尔自治区 石河子 832003
- 折叠
摘要
茶树嫩芽检测是实现名优茶智能采摘的前提,但在复杂茶园环境中存在芽叶目标与背景相似、密集生长等问题,为目标检测精度与速度带来难度.为提升茶树嫩芽的识别精度和处理效率,研究提出一种基于改进YOLOv8 模型的茶树嫩芽检测方法.该模型用Ghost_Conv替换原始模型中的标准卷积,极大地减少了模型的参数;在Backbone的C2f模块后添加了GAM_Attention模块,以提高对嫩芽目标的检测能力;同时将SPPF应用在Head前,帮助模型更好地整合从Neck输入的深层语义信息.试验结果表明,改进模型实现了90.7%的精度(P)、86.0%的召回率(R)和92.8%的平均精度(mAP).与初始网络相比,改进后模型的P、R和mAP分别提高了9.6%、7.2%和5.7%.与其他检测算法模型相比,在P、R和mAP及模型复杂度上均有较好效果,为后期模型在实际场景的部署应用提供了理论支持.
Abstract
The detection of tea buds is a prerequisite for achieving intelligent harvesting of high-quality tea.However,in complex tea garden environments,there are problems such as similarity between bud and leaf targets and background,dense growth,which pose difficulties for the accuracy and speed of target detection.To improve the recognition accuracy and processing efficiency of tea buds,a tea bud detection method based on an improved yolov8 model is proposed.This model replaces the standard convolution in the original model with Ghost_Conv,which greatly reduces the parameters of the model;the GAM_Attention module is added after the C2f module in the Backbone to improve the detection ability of tender shoot targets;at the same time,SPPF is applied in front of the Head,which helps the model better integrate the deep semantic information input from the Neck.The experimental results show that the improved model achieves an accuracy(P)of 90.7%,a recall rate(R)of 86%and an average accuracy(mAP)of 92.8%.Compared with the initial network,the improved model has increased P,R,and mAP by 9.6%,7.2%,and 5.7%,respectively.Compared with other detection algorithm models,it has good performance in P,R,mAP,and model complexity,providing theoretical support for the deployment and application of the model in practical scenarios in the later stage.
关键词
茶/嫩芽检测/深度学习/Ghost_Conv/GAM_AttentionKey words
Tea/Bud detection/Deep learning/Ghost_Conv/GAM_Attention引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据