Scholarlink

In order to solve the problems of closed current television (CCTV) detection defects in underground pipeline networks, including insufficient automation and reliance on the technical level of professional personnel, an intelligent method has been successfully constructed using comprehensive image processing and deep learning technology, which can assist detection personnel in quickly and accurately identifying pipeline defect types. Ten typical defect images were collected and processed to create a sample set. Using deep convolutional networks AlexNet and ResNet50 for transfer learning, classification network parameters were optimized. The intelligent classification model's accuracy was verified through a test set and specific engineering examples. Results showed test set accuracies of 92.00% and 96.50%, and real case accuracies of 85.41% and 87.94%, with ResNet50 performing better. This method enhances automation and accuracy in pipeline defect classification, showing significant potential for promotion.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automatic Defect Classification Method of Drainage Pipe Based on Image Processing and Deep Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhen He,
  • Ningjun Dang,
  • Xinwei Zhou,
  • JinHui Li

摘要

In order to solve the problems of closed current television (CCTV) detection defects in underground pipeline networks, including insufficient automation and reliance on the technical level of professional personnel, an intelligent method has been successfully constructed using comprehensive image processing and deep learning technology, which can assist detection personnel in quickly and accurately identifying pipeline defect types. Ten typical defect images were collected and processed to create a sample set. Using deep convolutional networks AlexNet and ResNet50 for transfer learning, classification network parameters were optimized. The intelligent classification model's accuracy was verified through a test set and specific engineering examples. Results showed test set accuracies of 92.00% and 96.50%, and real case accuracies of 85.41% and 87.94%, with ResNet50 performing better. This method enhances automation and accuracy in pipeline defect classification, showing significant potential for promotion.