Research on the Efficiency of Bridge Crack Detection by Coupling Deep Learning Frameworks with Convolutional Neural...

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Bridge health monitoring is crucial for infrastructure safety.
• Automated bridge crack detection using deep learning presents significant challenges.
• Existing methods require optimization for accuracy and efficiency.

Purpose of the Study:

• To investigate the effectiveness of coupling deep learning frameworks (DLF) with convolutional neural networks (CNN) for bridge crack detection.
• To identify optimal DLF and CNN model combinations for enhanced bridge health monitoring.
• To evaluate various CNN architectures and deep learning frameworks for crack detection performance.

Main Methods:

• A dataset of 2068 bridge crack images was utilized.
• Images were randomly split into training, verification, and testing sets (8:1:1 ratio).
• Experiments were conducted using Faster R-CNN, SSD, YOLO-v5(x), U-Net, and PSPNet models across PyTorch, TensorFlow2, and Keras frameworks.

Main Results:

• Faster R-CNN and SSD (Keras) achieved high F1 values (0.76, 0.67) in object detection.
• YOLO-v5(x) (TensorFlow2) reached an F1 value of 0.67.
• U-Net (PyTorch) achieved 98.37% accuracy (AC) in semantic segmentation.
• PSPNet (TensorFlow2) achieved 97.86% AC in semantic segmentation.

Conclusions:

• Optimal coupling parameters for DLF and CNN in bridge crack detection were identified.
• The study yielded more accurate and efficient DLF and CNN models for practical bridge health detection.
• The findings offer significant practical value for infrastructure maintenance and safety.