Smart City Infrastructure Monitoring with a Hybrid Vision Transformer for Micro-Crack Detection.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Current structural health monitoring (SHM) methods are inefficient, labor-intensive, and error-prone, especially for detecting micro-cracks.
• Subtle structural anomalies require advanced detection techniques for infrastructure safety and longevity.
• Deep learning offers potential for automated and precise defect identification in urban infrastructure.

Purpose of the Study:

• To develop a novel deep-learning framework for reliable structural health monitoring (SHM).
• To enhance defect detection capabilities, particularly for subtle anomalies like micro-cracks.
• To establish a new state-of-the-art in automated structural defect identification.

Main Methods:

• A modified Detection Transformer (DETR) architecture was developed, integrating a Vision Transformer (ViT) backbone.
• A Local Feature Extractor (LFE) module was designed to enhance the extraction of complex local spatial features.
• The Vision-Local Feature Detector (ViLFD) model was trained and validated on benchmark and custom datasets.

Main Results:

• The ViLFD model demonstrated superior performance over existing DETR and YOLO variants.
• Achieved high accuracy (95.0%), precision (0.94), recall (0.93), F1-score (0.93), and mAP@0.5 (0.89).
• Successfully and reliably detected subtle structural defects in experimental validation.

Conclusions:

• The proposed ViT-based DETR framework with LFE module significantly advances SHM capabilities.
• The ViLFD model offers a precise and reliable solution for automated detection of structural defects.
• This represents a significant step towards safer and more dependable urban infrastructure.