Climbing robot-assisted quantitative identification of sheath surface defects of stay cables

Sheath surface defect identification of stay cables using climbing robots or drones is gaining popularity. However, effective quantitative identification approaches remain challenging. This study proposes a cable climbing robot assisted quantitative identification method to address sheath surface defects combining the deep learning method and image processing techniques. First, standard and high-resolution cable sheath images, collected by an improved cable climbing robot with exclusive scales indicating the defect size, were categorised into three datasets according to defect types. Subsequently, a mask region-based conventional neural network (Mask R-CNN) model was developed for defect capture, with a connected domain algorithm quantifying defect areas. Perspective projection was applied to the defect masks to reduce the image distortion effect. The test results show that the model can achieve over 89% precision, recall and F1-score for three defect types, with masks presenting overlapping rates reach up to 87% compared to the ground-truth regions. The quantitative identification results are promising to facilitate stay cable assessment and maintenance. Future study will be focused on extending the volume and categories of the dataset and improving the performance of the model.