| 肖权旌,王强,谷小红,许卫荣,刘剑锋,国树东.改进YOLOv5的高密度聚乙烯管热熔接头3D全聚焦成像缺陷识别分析*[J].,2025,44(1):88-96 |
| 改进YOLOv5的高密度聚乙烯管热熔接头3D全聚焦成像缺陷识别分析* |
| 3D total focusing method imaging defect identification analysis of high density polyethylene thermal butt fusion joint based on improved YOLOv5 |
| 投稿时间：2023-09-01 修订日期：2024-12-26 |
| 中文摘要: |
| 为保障聚乙烯管热熔接头缺陷识别效率和准确性，该文提出一种基于改进YOLOv5的高密度聚乙烯管热熔接头3D全聚焦成像缺陷识别方法。首先，在YOLOv5模型的特征提取网络中加入SE注意力机制模块，提高不明显缺陷的边缘特征提取，在预测网络中，将原始YOLOv5的损失函数改为SIoU损失函数，提高模型回归效率和收敛速度，更有利于模型的优化；其次，对高密度聚乙烯试块典型缺陷(φ1 mm、φ2 mm、φ3 mm)进行3D全聚焦成像检测实验，采集原始的3D全聚焦缺陷图谱，完成图像增广并建立数据集；最后，采用迁移学习策略对改进模型进行训练，获取最优模型并进行评价。结果表明：该方法与传统YOLOv5相比，其准确率提升了2.4%，召回率提升了3.1%，较好解决检测人员错检、漏检等情况，提高检测效率。 |
| 英文摘要: |
| In order to ensure the efficiency and accuracy of defects identification of polyethylene pipe butt fusion joints. This paper proposed a method of 3D total focusing method (TFM) imaging defect identification of butt fusion joints of high density polyethylene (HDPE) pipe based on improved YOLOv5. Firstly, the SE attention mechanism module was added to the feature extraction network of YOLOv5 model to improve the edge feature extraction of non-obvious defects. In the prediction network, the loss function of the original YOLOv5 was changed to SIoU loss function to improve the regression efficiency and convergence rate of the model, which is more conducive to the optimization of the model. Secondly, 3D-TFM imaging detection experiments were conducted on the defects of HDPE test blocks (φ1 mm, φ2 mm, φ3 mm), the original 3D-TFM defect map was collected, image augmentation was completed, and the data set was established. Finally, transfer learning strategy is used to train the improved model, and the optimal model was obtained and evaluated. The results show that compared with traditional YOLOv5, the precision of this method is improved by 2.4%, recall rate by 3.1%. It is better to solve the wrong detection and missed detection of inspectors, and improve the detection efficiency. |
| DOI：10.11684/j.issn.1000-310X.2025.01.008 |
| 中文关键词: 全聚焦成像 3D成像 高密度聚乙烯管 缺陷识别 深度学习 |
| 英文关键词: Total focus method 3D imaging High density polyethylene pipe Defects identification Deep learning. |
| 基金项目:国家重点研发项目（2023YFF0611600）,浙江省‘尖兵’‘领雁’研发攻关计划资助（2022C03179）,浙江省市场监管局科技计划项目（20210144） |
|
| 摘要点击次数: 11 |
| 全文下载次数: 4 |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
