Multilevel Structural Components Detection and Segmentation toward Computer Vision-Based Bridge Inspection

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Bridge deterioration poses significant safety risks.
• Automated inspection methods are crucial for efficient and accurate structural assessments.
• Computer vision (CV) offers potential for automating bridge inspection tasks, particularly with unmanned aerial vehicles (UAVs).

Purpose of the Study:

• To propose and verify a multilevel bridge inspection framework utilizing CV technology and CNN models.
• To assess the performance of different CV models for recognizing bridge types and components at various distances.
• To investigate techniques for improving model accuracy, including image augmentation and hyperparameter tuning.

Main Methods:

• Developed a multilevel inspection framework using CV and CNNs.
• Trained Resnet50 on a dataset of 1200 images for long-distance bridge type classification.
• Employed YOLOv3 for medium-distance bridge component detection and Mask-RCNN for close-distance component segmentation.

Main Results:

• Achieved 96.29% accuracy in classifying bridge types (arched, cable-stayed, suspension).
• Obtained 93.55% detection accuracy for girders and 82.64% for piers at medium distances.
• Reached 90.8% bounding box and 87.17% segmentation accuracy for components at close distances.

Conclusions:

• The proposed CV framework effectively automates multilevel bridge inspection.
• Different CV models demonstrate varying performance based on distance and task (classification, detection, segmentation).
• The study provides valuable insights into CV model selection, data acquisition, and optimization for bridge inspection.