Enhance the Concrete Crack Classification Based on a Novel Multi-Stage YOLOV10-ViT Framework

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Traditional concrete crack detection is labor-intensive and time-consuming.
• Vision-based deep learning offers efficient automated solutions for structural health monitoring.
• Early identification of concrete cracks is crucial for preventing structural failure.

Purpose of the Study:

• To develop and evaluate a novel multi-stage deep learning framework for concrete crack detection and multi-class type classification.
• To enhance the accuracy and efficiency of identifying normal, simple crack, and multi-branched crack types.
• To provide an automated system for early warning of potential structural deformation in concrete.

Main Methods:

• A multi-stage framework combining YOLOV10 for defect region detection and a modified Vision Transformer (ViT) for crack classification.
• Training YOLOV10 on a dataset of 1116 concrete images with bounding boxes.
• Training the ViT model on a 12,000-image dataset for classifying three crack types (normal, simple, multi-branched).

Main Results:

• The proposed multi-stage YOLOV10-ViT model achieved 90.67% precision, 90.03% recall, and 90.34% F1-score for crack type classification.
• The multi-stage framework significantly outperformed the individual ViT model, showing improvements of up to 19.99% in recall.
• The model demonstrated high accuracy in detecting and classifying various concrete crack types.

Conclusions:

• The developed multi-stage deep learning framework offers a highly accurate and efficient solution for concrete crack detection and classification.
• This YOLOV10-ViT model can be integrated into construction systems for real-time structural health monitoring and early warning.
• The findings contribute to advancing automated inspection techniques in civil engineering and structural maintenance.