Multi-modal cascade detection of pipeline defects based on deep transfer metric learning

Pipeline defect detection technology plays an important role in pipeline maintenance and transportation. Defect detection based on machine learning methods has gained considerable attention in practical engineering. However, it is still challenging to provide an accurate diagnosis and defect size estimation due to the poor inter-class discriminability and intra-class concentration. Such as, it is difficult to distinguish the hole defect, which is similar in appearance to dent defect. For this purpose, a multi-modal cascade detection framework of pipeline defects based on Deep Transfer Metric Learning (DTML) is proposed for defect recognition and defect size estimation, which integrates with machine vision and Magnetic Flux Leakage (MFL). DTML model based on ResNet50 is designed to extract discriminative features from defect images obtained through vision sensor. To enhance the features of MFL signals, Gramian Angular Field (GAF) is used to achieve the two-dimensional feature extraction. After that, three ResNet101 models are developed to estimate the pipeline defect size of different types. The experimental results demonstrate that the proposed multi-modal cascade detection framework performs well in defect recognition and defect size estimation.