Surface damage detection of cable stays based on PointRend model using unmanned aerial vehicles

Abstract

Cable stays are critical load-bearing components that are significant of cable-stayed bridges. Traditional detection methods have several shortcomings. To address these problems, this study proposes a detection method that uses unmanned aerial vehicles to shoot videos of cable stays and identifies surface damage through deep learning. To improve the robustness of the method for detecting cable damage, the proposed method consists of three phases: background removal, damage recognition, and planar unfolding. In the first phase, a Background Removal Model based on PointRend is used to remove complex backgrounds of cable images, which could also reduce computational costs for subsequent processing of non-damage images. In the second phase, a Damage Recognition Model based on PointRend performs pixel-level semantic segmentation of damage. In the third phase, cable surface images are unfolded to eliminate the image distortion. On a self-made dataset, the proposed method achieved a mIoU score of 89.90%. Experimental results demonstrate the effectiveness of the proposed method in detecting cable stays’ surface damage.

Keywords:

• Background removal
• cable stays
• computer vision
• damage recognition
• deep convolutional neural networks
• PointRend model
• unmanned aerial vehicles

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The authors are grateful for the financial support from Key Scientific and Technological Research Projects of Henan Province (212102310975, 222102320436).