1.华南农业大学电子工程学院(人工智能学院)，广州 510642;2.广东省农情信息监测工程技术中心，广州 510642;3.华南农业大学珠江学院，广州 510900
谢家兴，E-mail：xjx1998@scau.edu.cn
刘洪山，E-mail：hugouliu@scau.edu.cn
TP391.4;S436.629
教育部产学合作协同育人项目（230702595161723）；国家现代农业产业技术体系专项（CARS-32-09）；华南农业大学新农村发展研究院农业科技合作共建项目（2021XNYNYKJHZGJ032）；广东省重点领域研发计划项目（2023B0202090001）；广东省现代农业产业技术体系创新团队建设专项（2023KJ108）；广东省科技创新战略专项资金项目（pdjh2023a0074）；大学生创新创业训练计划项目（S202410564127）
- XIE Jiaxing 1,2
- LIAO Fei 1
- WANG Weixing 3
- GAO Peng 1,2
- HU Kai 1
- WU Peiwen 1
- DENG Zhengqi 1
- LIU Hongshan 1,2
1.College of Electronic Engineering (Artificial Intelligence)， South China Agricultural University， Guangzhou 510642， China;2.Guangdong Engineering Research Center for Agricultural Information Monitoring， Guangzhou 510642， China;3.Zhujiang College， South China Agricultural University，Guangzhou 510900， China
- 摘要 |
- 图/表 |
- 访问统计 |
- 参考文献 |
- 相似文献 | | |
- 文章评论
针对荔枝园复杂背景下荔枝小目标病虫害检测困难的问题，提出一种基于改进Faster R-CNN的荔枝病虫害检测方法。以Faster R-CNN为基础，使用特征提取能力更优秀的Swin Transformer代替原有主干网络VGG16；通过特征金字塔网络（feature pyramid network， FPN）提升Faster R-CNN模型的多尺度特征融合能力，均衡提高每一类荔枝病虫害的识别精确率；引入感兴趣区域对齐（region of interest align， ROI Align）策略提升模型的候选框定位精度，进一步提升模型的整体检测效果。结果显示，改进后的模型平均精度均值达到92.76%，相较原始Faster R-CNN检测器提升了30.08百分点，在5类荔枝病虫害图像（藻斑病、炭疽病、煤烟病、毛毡病、叶瘿蚊）中的检测精度分别为93.05%、94.81%、96.57%、87.03%和92.34%，平均精度均值比SSD512、RetinaNet、EfficientDet-d0和YOLOv5s模型分别提高了20.50、5.70、13.08和3.26百分点。结果表明，改进后的Faster R-CNN模型能准确检测复杂背景下的荔枝病虫害目标，具有较高的应用价值，能为农作物病虫害快速、准确识别研究提供参考。
A method of detecting diseases and insect pests in litchi based on improved Faster R-CNN was proposed to solve the problems of detecting small targets of diseases and pests in complex backgrounds of litchi orchards. Swin Transformer with superior capabilities of extracting feature was used to replace the original backbone network VGG16 based on Faster R-CNN. The feature pyramid network （FPN） was used to enhance capability of the multi-scale feature fusion in the Faster R-CNN model， thereby improving the precision of identifying each type of diseases and insect pests in litchi in a balanced manner. The ROI Align strategy was introduced to refine the precision of the candidate box localization in the model， leading to the enhancement in the performance of overall detection in the model. The result showed that the average accuracy of the improved model was 92.76%， 30.08 percentage points higher than that of the original Faster R-CNN detector. The precision of detecting images of five types of diseases and insect pests including algal leaf spot， anthracnose， sooty mold， felt disease， and leaf gall in litchi was 93.05%， 94.81%， 96.57%， 87.03%， and 92.34%， respectively. The average precision was improved by 20.50， 5.70， 13.08， and 3.26 percentage points compared with that of SSD512， RetinaNet， EfficientDet-d0， and YOLOv5s model， respectively. It is indicated that the improved Faster R-CNN model can accurately detect diseases and insect pests in litchi with complex backgrounds， and has high value of application. It will provide a reference for studying the rapid and accurate identification of diseases and insect pests in crops.
引用本文
谢家兴,廖飞,王卫星,高鹏,胡凯,吴佩文,邓钲奇,刘洪山.基于改进Faster R-CNN的荔枝病虫害检测[J].华中农业大学学报,2025,44(1):62-73
复制
文章指标
- 点击次数:4
- 下载次数: 22
- HTML阅读次数: 3
历史
- 收稿日期:2024-09-11
- 在线发布日期: 2025-03-03
文章二维码
