基于卷积神经网络的管道表面缺陷识别研究
2020-09-21袁泽辉郭慧周邵萍
袁泽辉 郭慧 周邵萍



摘 要: 针对传统管道表面缺陷检测方法存在效率低、准确率不高的问题,提出一种通过机器视觉检测管道表面缺陷的方法,在采集管道表面缺陷的图像信息后通过卷积神经网络的算法分类不同的缺陷。通过加入批量归一化层,改进低层和中层卷积核的构造,优化了GoogleNet的构造,提高了卷积神经网络的泛化性和收敛性。试验结果表明,应用卷积神经网络后对管道表面缺陷的识别率较高,显著提高了管道表面缺陷识别的效率和准确率,具有较好的工程意义。
关键词: 缺陷识别; 管道表面缺陷; 机器视觉; 卷积神经网络; 缺陷分类; GoogleNet构造优化
中图分类号: TN911.73?34 文献标识码: A 文章编号: 1004?373X(2020)17?0047?05
Abstract: As the traditional pipeline surface defect detection method has the problems of low efficiency and low accuracy, a method based on machine vision to detect pipeline surface defects is proposed, in which the algorithm of convolutional neural network is used to classify different defects after the image information of pipeline surface defects is collected, and the low?level and middle?level convolution kernels are improved by adding a batch normalization?level to optimize the GoogleNet structure and enhance the generalization and convergence of convolutional neural networks. The test results show that the method's recognition rate of pipeline surface defects is higher after the application of convolutional neural network, which significantly improves the recognition efficiency and accuracy of defects on the pipeline surface. Therefore, the method has good engineering significance.
Keywords: defect recognition; pipeline surface defect; machine vision; convolutional neural network; defect classification; GoogleNet structure optimization
0 引 言
由于管道相比于其他運输而言,具有成本低、效率高的特点,管道运输在现代经济发展中扮演着极其重要的作用。然而随着时间的推移,管道表面会出现多种缺陷,这严重影响了管道运输的效率,甚至会造成事故带来巨大的损失。因此对管道表面缺陷进行检测是非常必要的。
目前,我国对管道表面缺陷无损检测的方法有超声波检测法、漏磁检测法、涡流检测法。其中,基于超声波检测法首先发射超声脉冲,当遇到管壁时发生反射,通过计算发射脉冲与反射脉冲的时间差获得管壁表面的情况。另外,基于漏磁的方法检测管壁表面状况,通过靠外加磁场对管壁进行磁化,如果管壁表面正常磁力线完全通过管道,若管壁表面出现缺陷则磁力线发生泄露,通过磁力线的变化获得管道表面的情况。……
