基于注意力机制改进YOLOv5的茶叶病害检测方法
A tea disease detection method based on Yolov5 improved attention mechanism
针对茶叶病害在复杂自然环境下识别精度受限的问题,提出了一种融合SE注意力机制的改进YOLOv5目标检测方法YOLOv5 - SE.该方法在YOLOv5网络主干中引入SE注意力模块,增强特征通道的自适应加权能力,抑制冗余信息对检测精度的影响.同时,引入小目标检测头,提升对小尺度病害的检测效果,增强模型在高分辨率特征图上的检测能力.在包含8种茶叶病害及健康叶片的数据集上进行实验验证,结果表明,YOLOv5 - SE的检测精度达到93.2%,召回率达到95.0%,较改进前模型精度提高4.5个百分点,召回率提高3.8个百分点.进一步的消融实验表明,SE注意力机制能够在优化特征提取能力的同时,有效提升模型对复杂背景下茶叶病害的检测鲁棒性,减少误检与漏检的发生.
To address the limitations in the accuracy of tea disease recognition under complex natural environments, this paper proposes an improved YOLOv5 target detection method, YOLOv5 - SE, which integrates the SE attention mechanism. The proposed method incorporates the SE attention module into the YOLOv5 backbone to enhance the adaptive weighting of feature channels and suppress redundant information that may affect detection accuracy. Additionally, a small - object detection head is introduced to improve the detection performance of small - scale tea diseases and enhance the model′s capability in high - resolution feature maps. Experimental validation on a dataset containing eight types of tea diseases and healthy leaves demonstrates that YOLOv5 - SE achieves a detection accuracy of 93.2% and a recall rate of 95.0%, with an improvement of 4.5 percentage points in accuracy and 3.8 percentage points in recall compared to the previous model. Further ablation experiments indicate that the SE attention mechanism optimizes feature extraction while significantly enhancing the model′s robustness in detecting tea diseases under complex backgrounds, thereby reducing false positives and false negatives.
target detection / attention mechanism / YOLOV5 / tea disease
| [1] |
杜朝东,蒋智钢,杜泽锋, |
| [2] |
田甜,程志友,鞠薇, |
| [3] |
孙道宗,刘欢,刘锦源, |
| [4] |
芦兵,孙俊,杨宁, |
| [5] |
胡根生,谢一帆,鲍文霞, |
| [6] |
李子茂,徐杰,郑禄, |
| [7] |
李汝嘉.基于加权抽样分层分类学习方法的茶叶病害分类[D].昆明:云南农业大学,2023. |
| [8] |
李伟豪,詹炜,周婉, |
| [9] |
李吴洁,危疆树,王玉超, |
| [10] |
陈禹,吴雪梅,张珍, |
| [11] |
杨小伟.基于卷积神经网络的自然场景下茶叶病害识别[D].合肥:安徽大学,2021. |
| [12] |
|
| [13] |
|
| [14] |
刘超阳,范菁,曲金帅, |
| [15] |
|
教育部-新一代信息技术创新项目(2023IT077)
云南省教育厅科学研究基金(2025Y0670)
云南省教育科学规划项目(BC24012)
/
| 〈 |
|
〉 |