[Paper Review] Bridge Damage Detection using a Single-Stage Detector and Field Inspection Images

Chaobo Zhang, C. C. Chang+1 more|arXiv (Cornell University)|Dec 19, 2018

Infrastructure Maintenance and MonitoringEngineering51 references18 citations

TL;DR

This paper proposes an improved single-stage YOLOv3 detector for automated bridge damage detection using field inspection images. By integrating batch renormalization and focal loss, the method achieves 80% average precision—13% higher than original YOLOv3—enabling faster, more accurate detection of four concrete damage types: crack, pop-out, spalling, and exposed rebar.

ABSTRACT

Detecting concrete surface damages is a vital task for maintaining the structural health and reliability of highway bridges. Currently, most of these tasks are conducted manually which could be cumbersome and time-consuming. Recent rapid advancement in convolution neural network has enabled the development of deep learning-based visual inspection techniques for detecting multiple structural damages. However, most deep learning-based techniques are built on two-stage, proposal-driven detectors and using less complex image data, which is not promising to promote practical applications and integration within intelligent autonomous inspection systems. In this study, a faster, simpler single-stage detector is proposed based on YOLOv3 for detecting multiple concrete damages of highway bridges. To realize this, a bridge damage dataset consisting of field inspection images is prepared. These images have large variations in damage appearance and monitoring scene and are labeled with four types of concrete damages: crack, pop-out, spalling and exposed rebar. The original YOLOv3 is further improved by introducing a novel transfer learning method, Batch Renormalization and Focal Loss. The improved YOLOv3 is evaluated in terms of average precision and speed. The results show that the improved YOLOv3 has a detection accuracy of up to 80%. Its performance is about 13% better than the original YOLOv3.

Motivation & Objective

• Address the limitations of manual bridge inspection, which is time-consuming and error-prone.
• Overcome the inefficiency of two-stage detectors in real-time structural health monitoring applications.
• Develop a faster, single-stage deep learning detector suitable for integration into autonomous inspection systems.
• Create a diverse, real-world dataset of field inspection images with four types of concrete damage for training and evaluation.
• Improve YOLOv3’s performance on complex, variable bridge damage images through novel training techniques.

Proposed method

• Constructed a bridge damage dataset from real field inspection images with high variability in lighting, scale, and damage appearance.
• Labeled images with four damage types: crack, pop-out, spalling, and exposed rebar.
• Enhanced YOLOv3 with batch renormalization to stabilize training and improve generalization under distribution shifts.
• Applied focal loss to address class imbalance, particularly for rare or small damage instances.
• Used transfer learning with ImageNet pretraining to accelerate convergence and improve detection performance.
• Trained and evaluated the model on the custom dataset using average precision and inference speed as metrics.

Experimental results

Research questions

• RQ1 Can a single-stage detector like YOLOv3 achieve high accuracy in detecting multiple types of concrete bridge damage from real field images?
• RQ2 How does the integration of batch renormalization and focal loss improve YOLOv3’s performance on complex, variable bridge inspection data?
• RQ3 To what extent does the proposed method outperform the original YOLOv3 in terms of detection accuracy and inference speed?
• RQ4 Can the improved detector be effectively deployed in practical, real-time bridge inspection systems?
• RQ5 How robust is the model to variations in damage appearance and image quality in real-world field conditions?

Key findings

• The improved YOLOv3 achieved an average precision of 80% on the bridge damage detection task.
• The model's performance was 13% higher than the original YOLOv3, demonstrating significant accuracy gains.
• The use of focal loss effectively mitigated class imbalance, improving detection of rare or small damage instances.
• Batch renormalization enhanced training stability and model generalization across diverse image conditions.
• The single-stage architecture enabled faster inference, making it suitable for real-time deployment in autonomous inspection systems.
• The dataset of field inspection images with four damage types provides a valuable benchmark for future research in bridge health monitoring.