Real-time defect detection in concrete structures using attention-based deep learning and GPR imaging.

Area of Science:

• Civil Engineering
• Artificial Intelligence
• Non-Destructive Testing

Background:

• Subsurface defect detection in concrete structures faces accuracy and real-time efficiency challenges.
• Automated analysis of ground-penetrating radar (GPR) data is crucial for infrastructure health monitoring.
• Class imbalance in defect datasets hinders model performance.

Purpose of the Study:

• To enhance the accuracy and real-time efficiency of automated GPR defect detection in concrete structures.
• To develop a robust deep learning model for subsurface defect identification.
• To improve the reliability of non-destructive testing for civil infrastructure.

Main Methods:

• An enhanced YOLOv5 model integrated with an Efficient Channel Attention (ECA) mechanism was proposed.
• A Deep Convolutional Generative Adversarial Network (DCGAN) was used for data augmentation to address class imbalance.
• A specialized dataset of concrete defects was curated for training and validation.

Main Results:

• The YOLOv5+ECA model achieved the highest mean average precision (mAP) compared to baseline and other attention variants.
• The proposed model maintained real-time inference speeds, suitable for unmanned aerial vehicle (UAV) deployment.
• ECA's channel-specific feature recalibration significantly improved detection accuracy.

Conclusions:

• The enhanced YOLOv5+ECA model offers a precise and efficient solution for subsurface defect detection in concrete.
• This approach advances automated GPR analysis for infrastructure health monitoring.
• The method is applicable to critical concrete structures like tunnel linings and bridge decks.