基于改进YOLOv5s模型的柑橘病虫害识别方法
Improved YOLOv5s based identification of pests and diseases in citrus
郑宇达 1陈仁凡 1杨长才 2邹腾跃1
作者信息
- 1. 福建农林大学机电工程学院,福州 350002
- 2. 福建农林大学计算机与信息学院,福州 350002
- 折叠
摘要
针对现有检测模型不能满足在自然环境中准确识别多种类柑橘病虫害的问题,提出一种基于改进YOLOv5s模型的常见柑橘病虫害检测方法.改进模型引入ConvNeXtV2模型,构建一个CXV2模块替换YO-LOv5s的C3模块,增强提取特征的多样性;添加了动态检测头DYHEAD,提高模型对不同空间尺度、不同任务目标的处理能力;采用CARAFE上采样模块,提高特征提取效率.结果显示,改进后的YOLOv5s-CDC的召回率和平均精度均值分别为81.6%、87.3%,比原模型分别提高了4.9、3.4百分点.与其他YOLO系列模型在多个场景下的检测对比,具有更高的准确率和较强的鲁棒性.结果表明,该方法可用于自然复杂环境下的柑橘病虫害的检测.
Abstract
Accurately identifying pests and diseases in citrus can be used to timely reduce the econom-ic losses.A common method for detecting pests and diseases in citrus based on the improved YOLOv5s model was proposed to solve the problems that the existing models of detection cannot accurately identify multiple types of pests and diseases of citrus in the natural environment.The model was improved by intro-ducing the ConvNeXtV2 model and constructing a CXV2 module to replace the C3 module of YOLOv5s,enhancing the diversity of extracted feature.The dynamic detection head DYHEAD was added to improve the processing ability of the model for different spatial scales and task targets.The CARAFE upsampling module was used to improve the efficiency of extracting feature.The results showed that the improved YO-LOv5s-CDC had a mean recall rate and average precision of 81.6%and 87.3%,4.9 percentage points and 3.4 percentage points higher than that of the original model,respectively.Compared with the detection with other YOLO serial models in multiple scenarios,it had higher accuracy and stronger robustness.It is indicat-ed that this method can be used for detecting the diseases and pests of citrus in complex natural environ-ments.
关键词
深度学习/病虫害/YOLOv5s/目标检测Key words
deep learning/pests and diseases/YOLOv5s/target detection引用本文复制引用
基金项目
福建省自然科学基金(2019J01402)
出版年
2024
NETL
NSTL
万方数据