Image-based detection of bolts and bolt-missing defects in multi-angle and complex background scenarios

Area of Science:

• Structural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Manual inspection of bolted connections is inefficient and time-consuming.
• Automated defect detection is hindered by image variability (angles, lighting, occlusion, background).
• Deep learning offers potential but requires robust models for diverse conditions.

Purpose of the Study:

• To develop and validate a robust deep learning model for automated bolt defect detection.
• To address challenges posed by variable imaging conditions in structural health monitoring.
• To improve the accuracy and efficiency of inspecting critical infrastructure joints.

Main Methods:

• Constructed a diverse bolt image dataset from bridges, towers, and lab models.
• Applied image enhancement and Generative Adversarial Networks (GANs) for data augmentation.
• Compared YOLOv5, YOLOv8, and YOLOv10, then enhanced YOLOv8 with a Swin-Transformer backbone and a novel Multi-Scale and Detail-Enhanced Module (MEDM).

Main Results:

• The enhanced YOLOv8 model achieved superior performance (mAP=0.91, recall=0.85, precision=0.9).
• Demonstrated consistent high accuracy across various viewing angles (>100%), illumination levels (>94%), and complex backgrounds (>97.2%).
• Achieved a 98.94% detection rate in practical deployment, identifying missing bolts effectively.

Conclusions:

• The proposed enhanced deep learning model significantly improves automated bolt defect detection.
• The approach is effective for real-world structural health monitoring in diverse and challenging environments.
• This technology enhances the efficiency and reliability of infrastructure inspection.