A lightweight convolutional neural network-based model and system for defect detection and navigation on bridge road surface

A lightweight convolutional neural network-based model and system for defect detection and navigation on bridge road surface

WoS Article

The Faster Region-based Convolutional Neural Network (Faster R-CNN) is widely used for detecting defects on road surface. However, its effectiveness in this task is limited by its large model size and slow detection speed. To address these challenges, two versions of the Faster R-CNN model-small and large-were developed. First, the models were structurally optimized by integrating inverted residual blocks, depthwise separable convolutions, and attention mechanisms to improve efficiency and performance. The large version also incorporated multiscale feature extraction for enhanced detection capabilities. Second, model pruning was applied to further compress the networks. Extensive ablation experiments were conducted to investigate the relationship between the model's internal structure and its impact on crack detection accuracy and efficiency. The experimental results demonstrate that the proposed models outperform general CNN-based models in bridge surface defect detection, achieving superior detection speed while maintaining high accuracy. The large version exhibits better performance but at the cost of increased model complexity. Testing was conducted on a real-life bridge in Nanjing, China. Additionally, a software application, integrated with a laptop and a smartphone, was deployed to identify defects and map their locations on the bridge, streamlining the detection process. The source code of this software is freely available at https://github.com/DUYA686686/detection-software.git

The Faster Region-based Convolutional Neural Network (Faster R-CNN) is widely used for detecting defects on road surface. However, its effectiveness in this task is limited by its large model size and slow detection speed. To address these challenges, two versions of the Faster R-CNN model-small and large-were developed. First, the models were structurally optimized by integrating inverted residual blocks, depthwise separable convolutions, and attention mechanisms to improve efficiency and performance. The large version also incorporated multiscale feature extraction for enhanced detection capabilities. Second, model pruning was applied to further compress the networks. Extensive ablation experiments were conducted to investigate the relationship between the model's internal structure and its impact on crack detection accuracy and efficiency. The experimental results demonstrate that the proposed models outperform general CNN-based models in bridge surface defect detection, achieving superior detection speed while maintaining high accuracy. The large version exhibits better performance but at the cost of increased model complexity. Testing was conducted on a real-life bridge in Nanjing, China. Additionally, a software application, integrated with a laptop and a smartphone, was deployed to identify defects and map their locations on the bridge, streamlining the detection process. The source code of this software is freely available at https://github.com/DUYA686686/detection-software.git

Bridge defect detection, Real-time detection, Faster R -CNN optimization, Software, Defect navigation, smartphone

Bridge defect detection, Real-time detection, Faster R -CNN optimization, Software, Defect navigation, smartphone

FU, R.; ZHANG, Y.; STRAUSS, A.; NOVÁK, D.; CAO, M.

01.10.2025

ADVANCES IN ENGINEERING SOFTWARE

208

United Kingdom of Great Britain and Northern Ireland

1

21

21

https://www.sciencedirect.com/science/article/pii/S0965997825001103?pes=vor&utm_source=clarivate&getft_integrator=clarivate