Inspection of surface defects on stay cables using a robot and transfer learning

In-service stay cables suffer from surface scratch and crack defects, which may cause corrosion inside cables, and fracture damage is likely to occur when those defects are exposed to long-term rain and sunshine environments. Current methods such as manual inspection and bridge inspection vehicles are inefficient, costly and risky. However, traditional image processing technologies (e.g., Canny) and convolutional neural networks may not be able to obtain accurate surface defect information. This paper proposes a novel and cost-effective method for identifying stay cable surface defects combining a cable inspection robot and transfer learning on a cascade mask region conventional neural network (Cascade Mask RCNN). This automatic procedure not only precisely iden-tifies the defects but also locates and measures the defects that can be used for further maintenance strategies. Comparison work and on-site testing were conducted to evaluate the proposed model performance, and the validity of cable defects identification and measurement. An automatic and cost-effective inspection method is proposed for cable surface defect detection. Transfer learning with Cascade Mask RCNN model is presented for defect identification and location. The IoU index can reach up to 0.743, comparison work with other networks and on-site test was implemented to validate the validity and accuracy of cable surface defect detection and measurement.