Transfer Learning-Based Defect Detection for Natural Gas Pipelines

In this paper, the problem of defect detection is investigated for natural gas pipelines. Combining with YOLO v5 network model, the transfer learning method is introduced to train the pre-training model, where the sewer pipe data set is employed as the source domain data. The weights of the underlying feature extraction network are transferred to the target model to help the natural gas pipeline detection model training. In addition, to verify the effectiveness of the proposed transfer learning-based defect detection method, the training model with the introduction of transfer learning has been compared with the conventional model on the natural gas pipeline defect dataset. Finally, the comparison results are presented to show the feasibility and effectiveness of the proposed method.