BC-DUnet-based segmentation of fine cracks in bridges under a complex background

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Bridge cracks are critical indicators of structural safety risks.
• Automatic detection and segmentation of bridge cracks are essential for civil engineers.
• Traditional methods struggle with complex backgrounds and small crack detection.

Purpose of the Study:

• To develop an advanced bridge crack segmentation method.
• To improve the accuracy and efficiency of automatic bridge crack detection.
• To overcome limitations of traditional methods in complex environments.

Main Methods:

• A densely connected feature extraction model (DCFEM) enhances small crack features.
• A background elimination module (BEM) filters irrelevant information.
• A cross-attention mechanism (CAM) improves pixel-level representation.

Main Results:

• Achieved 98.18% Pixel Accuracy, outperforming FCN and Unet.
• Increased Intersection over Union (IOU) by 14.12% over FCN and 4.04% over Unet.
• Demonstrated superior accuracy and generalization compared to non-traditional networks.

Conclusions:

• The BC-DUnet effectively eliminates background noise for accurate crack segmentation.
• The method enhances bridge crack detection efficiency and reduces costs.
• The proposed network holds significant practical application value for bridge maintenance.