摘要
对经济林木虫害进行目标检测有助于及时发现虫情,从而更有针对性地控制虫害。首先采用加权双向特征融合技术丰富各级特征图的语义信息和修改自适应Anchor计算方法对YOLOv5主干网络模型进行改进,然后在含20种经济林木虫害的图像扩增数据集上进行试验,最后与多种基于深度学习的目标检测方法进行对比。试验结果表明:改进后的YOLOv5模型相对于YOLOv3、YOLOv4、YOLOv5、Faster-RCNN和CenterNet模型,其平均精度均值分别提升0.133、0.156、0.113、0.128和0.078,最优达到0.923,模型推断速度为64.9帧。因此,改进的YOLOv5模型对经济林木虫害检测已达到实际应用水平,可为经济林木虫害预警系统提供算法支撑。
Object detection of insect pests in economic forests helps to detect insects in a timely manner,so as to control the pests more pertinently.In this paper,the weighted bidirectional feature fusion technology was used to enrich the semantic information of feature maps at all levels and the adaptive Anchor calculation method was modified to improve the YOLOv5 backbone network model.Then experiments were carried out on the enhanced data set containing 20 kinds of economic forest pests,and finally compared with a variety of object detection methods based on deep learning in map and inference speed.The results showed that compared with YOLOv3,YOLOv4,YOLOv5,Faster-RCNN and CenterNet models,the average accuracy of the improved YOLOv5 model was increased by 0.133、0.156、0.113、0.128 and 0.078,respectively,with the optimal accuracy of 0.923,and the model inference speed was 64.9 frames.Therefore,the improved YOLOv5 model in this paper has reached the practical application level on detecting insect pests in economic forests,which can provide algorithm support for the early warning system of economic forest insect pests.
作者
何颖
陈丁号
彭琳
He Ying;Chen Dinghao;Peng Lin(College of Big Data,Yunnan Agricultural University,Kunming,650201,China)
出处
《中国农机化学报》
北大核心
2022年第4期106-115,共10页
Journal of Chinese Agricultural Mechanization
基金
国家自然科学基金(31960290)
云南省重大科技专项计划项目资助(202002AD080002—6)
云南省基础研究专项(202101AT070248)。