Journal of Engineering Research

Coastal Bridges are prone to early rusting under the interaction of salt spray, vibration and load. Traditional manual detection is inefficient and inaccurate. To achieve visual identification and standardized classification of corrosion degrees, this paper proposes a visual inspection method that integrates multi-platform shooting, deep learning segmentation and three-dimensional index determination. A acquisition system composed of unmanned aerial vehicles (UAVs) and magnetic crawling robots was constructed. The Corro-U²-Net model was adopted to achieve high-precision mask extraction in the corrosion area. The rust proportion R, crater depth index D and surface topography entropy H were introduced to build a grade evaluation system, and the support vector machine was combined to achieve specification correspondence. In the actual measurement of the 200-meter test section of the Cross-Bay Bridge, the pixel-level IoU reached 0.87, and the grade consistency rate was 91%, which was significantly improved compared with the existing methods. Moreover, it demonstrated good engineering adaptability in terms of maintenance costs and decision-making efficiency.

Corrosion of steel bridges; Visual inspection; Lightweight network; Three-dimensional quantitative indicators; Maintenance decision-making closed loop