欢迎访问《中国农学通报》，
中国农学通报

中国农学通报 ›› 2023, Vol. 39 ›› Issue (22): 158-164.doi: 10.11924/j.issn.1000-6850.casb2023-0194

• 农业信息·科技教育 • 上一篇    

融合空洞卷积和特征金字塔的Faster R-CNN柑橘害虫检测方法

岑霄()   

  1. 广西农业职业技术大学信息工程学院，南宁 530007
  • 收稿日期:2023-03-14 修回日期:2023-05-12 出版日期:2023-08-05 发布日期:2023-07-28
  • 作者简介:

    岑霄，男，1991年出生，广西凭祥人，高级工程师/讲师，硕士，研究方向为目标检测、图像处理、无线网络、移动计算。通信地址：530007 广西南宁市西乡塘区大学东路176号 广西农业职业技术大学，E-mail：。

  • 基金资助:
    广西农业科技自筹经费项目“基于深度学习的图像分析技术在亚热带水果病虫识别中的应用研究”(Z2019100); 广西高校中青年教师科研基础能力提升项目“基于数据挖掘的广西农作物产量预测的应用研究”(2021KY1189); 广西农业职业技术大学科学研究与技术开发计划课题“基于Android的广西特色水果病虫害智能诊断系统的研究与实现”(YKJ2217); 广西农业职业技术大学科学研究与技术开发计划课题“基于‘物联网’技术智能移动式蔬菜灌溉系统的研究”(YKJ2132)

Faster R-CNN Detection Method for Citrus Pests Based on Dilated Convolution and Feature Pyramid

CEN Xiao()   

  1. College of Information Engineering, Guangxi Agriculture Vocational and Technical University, Nanning 530007
  • Received:2023-03-14 Revised:2023-05-12 Online:2023-08-05 Published:2023-07-28

RichHTML

2

PDF (PC)

33

摘要：

对柑橘种植过程中常见的4种害虫进行分析，针对害虫体型大小不一、低对比度特征多等检测难题，在Faster R-CNN模型基础上进行改进。针对池化层的下采样导致被检测图像的分辨率下降，进而导致图像中特征信息丢失的问题，采用空洞卷积法抓取图像中更多深层次的特征并增大感受野。结合特征金字塔网络FPN对数据中不同尺度的特征进行融合，增强特征的健壮性，解决原RPN网络只使用最终输出的单一图层进行检测，检测准确率不高的问题。通过与原Faster R-CNN模型、YOLOv4等经典目标检测模型进行对比，改进方案的mAP为91.72%，检测准确率得到一定的提升，实验结果证明所提出的改进方案能够适应自然环境下柑橘害虫识别的需求。

关键词: 空洞卷积, 特征金字塔网络, Faster R-CNN, 柑橘害虫, 目标检测

Abstract:

The four common pests in the citrus planting process were analyzed, and the detection problems of the pests with different sizes and low contrast features were improved on the basis of the Faster R-CNN model. Aiming at the problem that the down-sampling of the pooling layer leads to a decrease in the resolution of the detected image, which in turn leads to the loss of feature information in the image, the hole convolution method is used to capture more deep-level features in the image and increase the receptive field. Combined with the feature pyramid network FPN to fuse the features of different scales in the data, enhance the robustness of the features, and solve the problem that the original RPN network only uses a single final output layer for detection, and the detection accuracy is low. Compared with the original Faster R-CNN model, YOLOv4 and other classic target detection models, the mAP of the improved scheme is 91.72%, and the detection accuracy is improved. The experimental results prove that the proposed improvement scheme can meet the demand of identifying citrus pests in natural environments.

Key words: dilated convolution, FPN, Faster R-CNN, citrus pest, object detection