| 基于YOLOv11-VCL模型的田间辣椒苗实时检测方法 |
| 点此下载全文 |
| 引用本文：胡家睿,陆傲鹏,陈蒙,石航,韩长杰.基于YOLOv11-VCL模型的田间辣椒苗实时检测方法.植物保护学报,2026,53(1):154-163 |
| DOI：10.13802/j.cnki.zwbhxb.2026.2026810 |
| 摘要点击次数: |
| 全文下载次数: |
|
|
| 中文摘要:为精准监测辣椒苗移栽后的存活状态，以提升移栽成功率并评估其生长状态，以YOLOv11n为基础模型构建并优化辣椒苗快速检测模型，针对田间辣椒苗检测中常见的复杂背景与光照条件干扰问题，在骨干网络的跨阶段局部v3+2×2卷积核（cross-stage partial v3 with kernel 2，C3k2）模块中引入坐标注意力（coordinate attention，CA）机制；同时加入视觉变换器（vision transformer，ViT）注意力机制模块以优化多尺度特征融合的识别效率与精准度，采用高效交并比（efficient intersection over union，EIoU）损失函数替代传统定位损失模块，结合多任务权重自适应策略和数据增强手段优化训练过程，以提高定位精度。结果表明：改进的YOLOv11-VCL模型对于不同光照、复杂背景条件下的辣椒苗均有较好的检测效果，精确率、召回率和平均精度均值分别为95.8%、90.6%和95.1%，明显优于 YOLOv5n、YOLOv8n、YOLOv9n、YOLOv10n 和 YOLOv11n 模型。YOLOv11-VCL 模型的参数量为3.3 M，计算量为7.5 GFLOP，推理速度为每秒146帧，能够满足精准施药边缘计算设备的运行要求，有效助力辣椒的自动化田间管理。 |
| 中文关键词:辣椒苗 实时检测 注意力机制 YOLOv11-VCL 卷积神经网络 |
| |
| A real-time detection approach for pepper seedlings in farmland utilizing YOLOv11-VCL model |
| Author Name | Affiliation | E-mail | | Hu Jiarui | College of Mechanical and Electrical Engineering, Xinjiang Agricultural University, Urumqi 830052, Xinjiang Uygur Autonomous Region, China
Xinjiang Intelligent Agricultural Machinery Equipment Engineering and Technology Research Center, Urumqi 830052, Xinjiang Uygur Autonomous Region, China | | | Lu Aopeng | College of Mechanical and Electrical Engineering, Xinjiang Agricultural University, Urumqi 830052, Xinjiang Uygur Autonomous Region, China
Xinjiang Intelligent Agricultural Machinery Equipment Engineering and Technology Research Center, Urumqi 830052, Xinjiang Uygur Autonomous Region, China | | | Chen Meng | College of Mechanical and Electrical Engineering, Xinjiang Agricultural University, Urumqi 830052, Xinjiang Uygur Autonomous Region, China
Xinjiang Intelligent Agricultural Machinery Equipment Engineering and Technology Research Center, Urumqi 830052, Xinjiang Uygur Autonomous Region, China | | | Shi Hang | College of Engineering, Heilongjiang Bayi Agricultural University, Daqing 163319, Heilongjiang Province, China | | | Han Changjie | College of Mechanical and Electrical Engineering, Xinjiang Agricultural University, Urumqi 830052, Xinjiang Uygur Autonomous Region, China
Xinjiang Intelligent Agricultural Machinery Equipment Engineering and Technology Research Center, Urumqi 830052, Xinjiang Uygur Autonomous Region, China | hcj_627@163.com |
|
|
| Abstract:Accurate monitoring of the survival status of pepper seedlings post-transplanting is the key to improving the transplanting success rate and evaluating growth status, and also a prerequisite for realizing the automatic field management of peppers. To this end, this study proposes a fast detection method for the survival status of pepper seedlings named YOLOv11-VCL, which is based on the improved YOLOv11n. Firstly, aiming at the common interference problems of complex background and illumination conditions in field pepper seedling detection, the coordinate attention (CA) mechanism is introduced into the cross-stage partial v3 with kernel 2 (C3k2) module of the backbone network. Meanwhile, to further optimize the recognition efficiency and accuracy of multi-scale feature fusion, the vision transformer (ViT) attention mechanism module is added to enable the model to detect pepper seedlings more accurately in complex field scenarios. In addition, efficient intersection over union (EIoU) is adopted to replace the traditional localization loss, and the multi-task weight adaptive strategy and data augmentation methods are integrated to optimize the model training so as to improve the localization accuracy. The results showed that the YOLOv11-VCL network model achieved excellent detection performance for pepper seedlings under different illumination and complex background conditions, with the precision, recall and mean average precision (mAP) reaching 95.8%, 90.6% and 95.1%, respectively. The detection results were significantly superior to those of the YOLOv5n, YOLOv8n, YOLOv9n, YOLOv10n and YOLOv11n models, and the missed detection and false detection rates of the improved model were significantly reduced. The improved model has a parameter quantity of 3.3 M, GFLOPs of 7.5 and a frame rate of 146 frames per second (FPS), which can meet the operation requirements of edge computing devices for precision pesticide application, and effectively facilitate the realization of automatic field management. |
| keywords:pepper seedling real-time detection attention mechanism YOLOv11-VCL convolutional neural network |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
|
您是本站第 16178383 版权所有：植物保护学报 京ICP备05006550号-2 主管单位：中国科协 主办单位：中国植物保护学会、中国农业大学 地址：北京市圆明园西路2号 中国农业大学植物保护学院 植物保护学报编辑部 电话：010-62732528 电子邮件：zbxb@cau.edu.cn 技术支持： 北京勤云科技发展有限公司 |
