| 摘要: |
| 【目的】 为推动井冈蜜柚果园实现智能化喷药，提升资源利用效率，提出一种基于改进的YOLOv5的蜜柚果树检测方法。【方法】 文章采用无人机M30T搭载CMOS影像传感器，在蜜柚果园采集504张高分辨率图像，并使用LabelImg工具进行图像标注。引入ECA注意力模块对YOLOv5目标检测算法进行改进，提高蜜柚果树的检测精度和速度。【结果】 （1）在测试集上，改进的YOLOv5算法平均检测精度达90.09%，检测速度达到5.28 it/s，与SSD模型、YOLOv4 模型和YOLOv5模型相比精度分别提高了8.33%、12.74%和1.72%。（2）改进算法在不同光照条件和拍摄角度下均展现出良好的鲁棒性，综合漏检率仅为3.54%，较原始YOLOv5、YOLOv4和SSD模型分别降低了1.77%、16.81%和14.15%。（3）ECA模块实现精度—速度最优平衡，较原模型mAP提升1.72%至90.09%，优于其他注意力模块。【结论】 改进的YOLOv5算法能够实现蜜柚果树的精准检测，对井冈蜜柚果园的智能喷药的发展有重要意义。 |
| 关键词: YOLOv5 蜜柚果树识别 注意力模块 |
| DOI：10.12105/j.issn.1672-0423.20250103 |
| 分类号: |
| 基金项目:江西省重大科技专项课题“果园智慧管控技术与智能装备集成研究”（20203ABC28W014-5）；井冈山农高区省级科技专项揭榜挂帅项目“智慧农业综合服务平台关键技术与装备研发与应用”（20222-051255） |
|
| Identification method of honey pomelo tree based on improved YOLOv5 |
|
Zhang Jiahao1,2, Guo Yang1, Liu Jie1,2, Chen Guipeng1,2
|
|
1.Jiangxi Academy of Agricultural Sciences，Institute of Agricultural Economics and Information，Nanchang 330200，Jiangxi，China;2.East China Jiaotong University，School of Electrical and Automation Engineering，Nanchang 330000，Jiangxi，China
|
| Abstract: |
| [Purpose] In order to promote intelligent spraying in Jinggang honey pomelo orchards and improve the efficiency of resource use，this study proposes a honey pomelo tree detection method based on an improved YOLOv5 algorithm.[Method] The M30T drone equipped with a CMOS image sensor was used to collect 504 high-resolution images in honey pomelo orchards，which were annotated using the LabelImg tool. An attention mechanism module was introduced to enhance the YOLOv5 target detection algorithm to improve the precision and speed of pomelo tree identification.[Result] （1）On the test set，the improved YOLOv5 algorithm achieved a mean average precision（mAP）of 90.09% with a detection speed of 5.28 iterations per second（it/s）. Compared to the SSD，YOLOv4，and the original YOLOv5 models，the precision improved by 8.33%，12.74%，and 1.72%，respectively. （2）The improved algorithm demonstrated strong robustness under varying illumination conditions and camera angles，exhibiting a comprehensive missed detection rate of only 3.54%. This represented reductions of 1.77%，16.81%，and 14.15% compared to the original YOLOv5，YOLOv4，and SSD models，respectively. （3）The efficient channel attention（ECA）module achieved an optimal balance between precision and speed. Compared to the baseline YOLOv5 model，it improved the mAP by 1.72% to 90.09%，outperforming other attention mechanisms in comprehensive evaluations.[Conclusion] The improved YOLOv5 algorithm demonstrates the ability to achieve precise honey pomelo tree detection and holds significant implications for developing intelligent spraying systems in Jinggang honey pomelo orchards. |
| Key words: YOLOv5 honey pomelo tree identification attention mechanism |
