| 李宗柱, 宋绍剑*, 李修华.一种基于YOLOv5的农业害虫检测方法[J].植物保护,2025,51(1):111-122. |
| 一种基于YOLOv5的农业害虫检测方法 |
| A YOLOv5-based method for agricultural pest detection |
| 投稿时间：2024-02-21 修订日期：2024-03-20 |
| DOI：10.16688/j.zwbh.2024089 |
| 中文关键词: 害虫识别 YOLOv5 数据增强 注意力机制 多尺度特征提取 |
| 英文关键词:pest identification YOLOv5 (You Only Look Once version 5) data augmentation attention mechanism multi-scale feature extraction |
| 基金项目:国家自然科学基金(31760342) |
|
| 摘要点击次数: 489 |
| 全文下载次数: 106 |
| 中文摘要: |
| 虫害是影响农作物产量的重要因素之一, 害虫种类的精确识别已成为农业领域目标检测的重要研究课题?但由于害虫样本存在类间相似, 标注的害虫样本尺度多样?背景复杂和类别分布不均匀等问题, 使害虫的精准识别面临严峻挑战?为此, 本文提出一种基于YOLOv5改进模型的农业害虫检测新方法?首先, 引入了一种新型特征金字塔(feature pyramid attention, FPA)模块, 用于替换基准YOLOv5主干网络的空间金字塔池化(spatial pyramid pooling, SPP)模块?该模块能够进行不同尺度的特征提取, 并将提取的特征拼接作为注意力机制指导网络进行细粒度特征提取?然后, 在YOLOv5主干网络输出层的特征提取过程中插入全局注意力上采样(global attention upsampling, GAU)模块, 用高级特征的全局信息来指导模型从复杂背景中提取特征, 使得模型能够从低级特征中更精准地提取类别定位细节特征, 进而提高模型的识别精度?本文在IP102害虫数据集上进行算法验证, 结果表明, 与现有的多尺度注意学习网络(multiscale attention learning network, MS-ALN)相比准确率提升了3.21百分点? |
| 英文摘要: |
| Pest attack is a critical factor that affects agricultural crop yields, and the accurate identification of pest species has become an important research topic in the field of target detection in agriculture. However, the accurate identification of pests still face critical challenges due to the issues, such as complex backgrounds, interclass similarities, multiple scales of annotated samples, and uneven distributions among different category samples. Therefore, this paper proposed a new method for detecting crop pests based on an improved YOLOv5 model. First, we introduced a feature pyramid attention (FPA) module to replace the spatial pyramid pooling (SPP) module in the backbone of YOLOv5, which enhances the network performance in extracting features of different scales and concatenates the extracted features as an attention mechanism to guide the network to extract fine-grained features. Then, a global attention upsampling (GAU) module was inserted into the output layers of the backbone to guide the model to extract features from complex backgrounds. Experiments conducted on the IP102 dataset showed that the accuracy was improved by 3.21 percent point compared to the multiscale attention learning network (MS-ALN), which achieved state-of-the-art performance on the IP102 dataset. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
