摘要
[目的]本文针对传统农作物叶片病虫害识别模型YOLOv3存在的检测实时性与鲁棒性差以及漏检率高的问题,提出了一种改进的玉米叶片病虫害检测模型——YOLOv3-Corn。[方法]该模型采用Darknet-53作为特征提取网络,将网络输出的8倍特征图与新加入的4倍下采样特征图进行拼接,建立了104104尺度的检测层;在前期构建的包含6个类别玉米常见病虫害数据集中,利用K-means++聚类算法选取12个先验框并分别匹配到4个不同尺度的检测层中进行目标识别。[结果]在YOLOv3系列模型中,YOLOv3-Corn模型的检测精度均值(mAP)、召回率(Recall)达到了93.31%和93.08%,与YOLOv3模型相比分别提高了4.03%和9.78%。在非YOLO系列模型中,YOLOv3-Corn模型平衡了Faster R-CNN模型的检测速度不足和RetinaNet模型的召回率、精确度不足的问题。[结论]在保证提取相同特征参数、检测时效性好的前提下,YOLOv3-Corn模型有效提高了识别精度。
[Objectives]Aiming at the problems of YOLOv3 model in the detection of crop diseases and insect pests,such as poor real-time detection and robustness,and high missed detection rate,an improved corn pest detection model-YOLOv3-Corn was proposed in this paper.[Methods]In the model,Darknet-53 was used as a feature extraction network,and then a 104104-scale detection layer was built by splicing the 8-fold feature map output by Darknet-53 and the newly added 4-fold down-sampling feature map.Finally,in the pre-constructed data set containing 6 categories of common corn diseases and pests,12 anchor boxes were selected by using the improved K-means clustering algorithm(K-means++)and matched to 4 detection layers of different scales for target recognition.[Results]The experimental results showed that the mean detection accuracy(mAP),and recall rate(Recall)of the YOLOv3-Corn model were 93.31%and 93.08%,which were 4.03%and 9.78%higher than those of the YOLOv3 model in YOLOv3 series models,respectively.In non-YOLO series models,the YOLOv3-Corn model balanced the insufficient detection speed of the Faster R-CNN model with the insufficient recall rate and accuracy of the RetinaNet model.[Conclusions]Under the premise of ensuring the same feature parameters and detection timeliness,the precision rate of YOLOv3-Corn model is effectively improved.
作者
徐会杰
黄仪龙
刘曼
XU Huijie;HUANG Yilong;LIU Man(School of Management,Henan University of Science and Technology,Luoyang 471023,China;Information Engineering College,Henan University of Science and Technology,Luoyang 471023,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2022年第6期1276-1285,共10页
Journal of Nanjing Agricultural University
基金
国家自然科学基金项目(61901241)
河南省高等学校重点项目(21A510003)
河南省高校省级大学生创新创业训练计划项目(S202110464045)
河南科技大学SRTP项目(2021252)。
