Two-step rapid inspection of underwater concrete bridge structures combining sonar, camera, and deep learning

Abstract

Underwater defects in piers pose potential hazards to the safety and durability of river-crossing bridges. The concealment and difficulty in detecting underwater defects often result in their oversight. Acoustic methods face challenges in directly achieving accurate measurements of underwater defects, while optical methods are time-consuming. This study proposes a two-step rapid inspection method for underwater concrete bridge piers by combining acoustics and optics. The first step combines macroscopic sonar scanning with an enhanced YOLOv7 to detect and locate piers and defects. Second, the camera approaches the defects for image acquisition, and an enhanced DeepLabv3+ is used for defect identification. The results demonstrate an average mean average precision@0.5 of 95.10% for defect and pier detection, and an mean intersection over union of 0.914 for exposed reinforcement and spalling identification. The method was applied to a real river-crossing bridge and reduced inspection time by 51.2% compared to traditional methods for assessing a row of 11 piers.