基于改进YOLOv5n模型的农作物病虫害识别方法
Identification method of crop diseases and insect pests based on improved YOLOv5n model
承达瑜 1赵伟 1何伟德 1武择鹏 1王建东2
作者信息
- 1. 河北工程大学矿业与测绘工程学院,河北 邯郸 056038
- 2. 中国农业科学院农业环境与可持续发展研究所,北京 100081
- 折叠
摘要
针对模型对复杂场景下农作物病虫害的识别精度低、模型参数量大的问题,本研究对轻量级YOLOv5n模型进行改进.首先,在YOLOv5n模型的骨干网络中加入坐标注意力模块,使模型关注检测目标及其位置,减少复杂背景对模型的影响.其次,引入加权的双向特征融合金字塔网络(BiFPN),减少小目标信息丢失,提高了模型的特征学习能力.最后,用损失函数SIoU代替损失函数CIoU,在不改变模型参数量的情况下,提升了目标检测精度.在无人机采集到的玉米病虫害数据集上,本研究提出的AgriPest-YOLOv5n模型的mAP@0.50 达 81.32%,在Jetson Xavier开发板上检测速度达到 77 FPS,模型大小为 1.63 MB.改进后的YOLOv5n模型能够满足轻量化的要求,能够实时、准确地识别复杂背景下的农作物病虫害,本研究结果可为病虫害精准防治提供技术支持.
Abstract
In order to solve the problems of low recognition accuracy for crop diseases and insect pests in complex scenes and large model parameters of the model,the lightweight YOLOv5n model was improved in this study.Firstly,a co-ordinate attention module was added to the backbone network of YOLOv5n model to make the model focus on the detection target and its location and reduce the influence of complex background on the model.Secondly,the weighted bi-directional feature fusion pyramid network(BiFPN)was introduced to reduce the information loss of small targets and improve the model's feature learning ability.Finally,the loss function SIoU was used to replace the loss function CIoU,which improved the target detection accuracy without changing the parameters of the model.In the dataset of corn pests and diseases collect-ed by unmanned air vehicle,the AgriPest-YOLOv5n model mAP@0.50 proposed by this study reached 81.32%,and the detection speed reached 77 FPS on the Jetson Xavier development board.The size of the model was 1.63 MB.The improved YOLOv5n model can meet the requirement of light weight,and can identify crop diseases and insect pests in real time and accurately under complex background.The results of this study provide technical support for the precision control of crop diseases and insect pests.
关键词
农业病虫害/目标检测/轻量级模型/注意力机制Key words
agricultural pest and disease/object detection/lightweight model/attention mechanism引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据