A Code-Conforming Computer Vision Framework for Visual Inspection of Reinforced and Prestressed Concrete Bridges

TL;DR

This paper introduces a new AI tool called VIADUCT to help inspect concrete bridges by detecting various types of damage using computer vision.

Contribution

The study introduces a code-conforming framework that detects a wide range of bridge defects using multimodal attention mechanisms and deep learning.

Findings

The VIADUCT tool uses YOLOv8n and attention mechanisms to detect bridge defects with promising precision.

Multimodal attention mechanisms improve detection by focusing on relevant bridge areas and suppressing background.

The framework is adaptable to newer object detection models as they become available.

Abstract

The assessment of structural degradation in reinforced concrete bridges is a crucial task for infrastructure maintenance and safety. Traditional inspection methods are often time-consuming, dependent on expert interpretation and weather conditions. This study explores the potential of artificial intelligence to support inspectors in the detection of typical deterioration patterns in reinforced (RC) and prestressed concrete (PRC) bridges, developing the VIADUCT (Visual Inspection and Automated Damage Understanding by Computer vision Techniques) software tool. Unlike previous studies, focusing only on a limited variety of possible defects (e.g., cracks, water stains), this study aims to train a deep learning model to be able to recognise a larger range of defects, such as those foreseen by the current Italian code for the assessment of existing bridges. The methodology relies on the YOLOv8n object detection model, which was trained, validated, and tested using a dataset including 1045 either wide-angle or detailed photographs taken during routine inspections. With these kinds of images being challenging for object detection algorithms (they include large parts of the background), multimodal attention mechanisms were implemented in the Graphical User Interface (GUI) through the semantic segmentation of the bridge surface using both the SAM and the U-Net model, as well as a tile reduction approach. These attention mechanisms allow the object detection model to focus on the relevant portions of the image (i.e., the bridge), while suppressing background information. Despite the limitation of the small size dataset used for training, results showed promising detection capabilities and precision. Furthermore, VIADUCT is ready to accept and use newer and more efficient versions of the object detection model, as soon as they become available.

Figures