CrackNet: A novel multi-scale architecture for crack segmentation.

Area of Science:

• Computer Vision
• Structural Engineering
• Artificial Intelligence

Background:

• Structural safety inspection relies heavily on accurate crack detection.
• Existing methods face challenges due to noise, illumination variations, and complex backgrounds.
• Automated crack detection is crucial for efficient and reliable structural health monitoring.

Purpose of the Study:

• To introduce CrackNet, a specialized segmentation network for concrete crack detection.
• To enhance the accuracy and robustness of automated crack detection systems.
• To provide a robust solution for real-world engineering applications.

Main Methods:

• Developed CrackNet, a segmentation network featuring a lightweight multi-scale convolution enhancement block (LightMSCBlock) for detailed feature extraction.
• Integrated a SAF attention module in skip connections for scale-aware feature fusion and edge refinement.
• Employed a multi-scale feature fusion (MSFF) module in the decoder to optimize feature integration and minimize information loss.

Main Results:

• CrackNet demonstrated superior performance on three public datasets (CFD, Crack500, DeepCrack) compared to state-of-the-art methods.
• Achieved significant improvements in F1 and IoU scores, outperforming methods like SegFormer and MobileNetV3-UNet.
• Ablation studies confirmed the effectiveness of individual modules (LightMSCBlock, SAF, MSFF).

Conclusions:

• CrackNet offers enhanced accuracy and robustness in concrete crack detection.
• The proposed network shows strong potential for practical implementation in structural safety inspections.
• The developed model contributes to advancing automated structural health monitoring techniques.