Automated concrete crack detection enhanced by deep learning and generative adversarial networks

Area of Science:

• Civil Engineering
• Computer Vision
• Materials Science

Background:

• Traditional concrete crack inspection is subjective and poses safety risks.
• Automated methods struggle with limited data and complex environments.

Purpose of the Study:

• To develop an automated, high-precision method for concrete crack detection and measurement.
• To address challenges of limited annotated data and complex environments in automated crack analysis.

Main Methods:

• Utilized Conditional Generative Adversarial Networks (cGAN) for high-fidelity crack image augmentation.
• Employed an enhanced U-Net segmentation network with channel attention and dilated residual blocks.
• Applied a hybrid loss function (Dice + binary cross-entropy) to manage class imbalance.
• Incorporated fractal theory for quantitative geometric feature analysis of detected cracks.

Main Results:

• The cGAN-augmented dataset improved crack segmentation model performance, achieving 82.03% Average Precision (AP) and 83.24% F1 score.
• The fractal dimension-based measurement method demonstrated ≤4% error for complex crack networks.
• Experimental results verified the method's stability and engineering applicability.

Conclusions:

• The proposed automated method offers a reproducible and precise solution for concrete crack detection and quantification.
• This approach effectively overcomes limitations of traditional inspection and current automated techniques.