基于图像处理和机器学习的PE管道缺陷检测  

PE pipeline defect detection based on image processing and machine learning

作　　者：符前坤 李强 冉文燊 林楠 王洋[1] FU Qiankun;LI Qiang;RAN Wenshen;LIN Nan;WANG Yang(School of Intelligent Manufacturing Modern Industry(School of Mechanical Engineering),Xinjiang University,Urumqi 830046,China;Xinjiang Yian Special Inspection Engineering Co.,Ltd.,Urumqi 830000,China;Xinjiang Uygur Autonomous Region Special Equipment Inspection and Research Institute,Urumqi 830000,China;Department of Pressure Pipe,China Special Equipment Inspection and Research Institute,Beijing 100013,China)

机构地区：[1]新疆大学智能制造现代产业学院(机械工程学院),新疆乌鲁木齐830046 [2]新疆益安特检工程有限公司,新疆乌鲁木齐830000 [3]新疆维吾尔自治区特种设备检验研究院,新疆乌鲁木齐830000 [4]中国特种设备检验研究院压力管部,北京100013

出　　处：《现代电子技术》2024年第21期59-66,共8页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(11903072);新疆大学博士启动项目(620321029)。

摘　　要：对于聚乙烯(PE)管道,在运行中经常有不同程度的泄漏等异常,通过对管道中不同异常的实验模拟,收集数据并手动标记相应的数据集。为了提高管道缺陷图像的质量,首先采用加权平均法对图像进行灰度处理;然后,利用伽马变换改进管道背景与缺陷的对比度;最后,使用双重过滤来降低图像中的噪声。为了降低数据的复杂度,提高模型训练速度,采用改进的Sobel算法对管道缺陷图像进行边缘检测,采用自适应阈值分割算法分割缺陷图像的边缘,生成二值图像,用二值图像训练模型,减少了模型对颜色特征的依赖,加快了模型的收敛速度。为了提高管道缺陷检测的精度,引入CA注意力机制,提高目标检测特征提取能力。改进的YOLOv5模型的mAP和召回率分别为97.18%和98.03%。与原算法相比,mAP增加了1.33%,召回率增加了3.83%。Usually,anomalies such as different degrees of leakage may occur to the polyethylene(PE)pipes in operation.The data are collected and the corresponding data sets are labeled manually by experimental simulation of different anomalies of the pipelines.The weighted average method is used to process the image gray in order to improve the quality of defect images of the pipeline.The contrast ratio between the pipeline background and the defects is improved by applying gamma transform.Double filtering is used to reduce the noise in the image.In order to reduce the complexity of the data and improve the speed of model training,the improved Sobel algorithm is used to detect the edges of the defect image.The adaptive threshold segmentation algorithm is used to divide the edges of the defect image,so as to generate a binary image.The binary image is used to train the model,which reduces the model dependence on the color features and accelerates the model convergence.In order to improve the accuracy of pipeline defect detection,the coordinate attention(CA)mechanism is introduced and the feature extraction ability of object detection is improved.The mAP and recall rate of the improved YOLOv5 are 97.18%and 98.03%,respectively.In comparison with that of the original algorithm,its mAP is increased by 1.33%and its recall rate is increased by 3.83%.

关 键 词：缺陷检测 图像处理 机器学习 YOLOv5 注意力机制 二值图像 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]