LCNet: balancing representation capacity and computational cost for concrete crack detection in complex backgrounds.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Concrete crack detection is vital for structural health monitoring but challenging due to complex backgrounds and limited edge device resources.
• Existing methods often struggle with visual interference and computational efficiency on resource-constrained platforms.

Purpose of the Study:

• To propose a lightweight crack segmentation network (LCNet) for accurate and efficient concrete crack detection.
• To enhance computational resource allocation and feature representation for improved performance in bridge inspection.

Main Methods:

• Developed a lightweight crack segmentation network (LCNet) featuring an asymmetric channel scaling strategy and a bottleneck enhancement module.
• Conducted extensive experiments on a self-constructed bridge crack dataset and real-world bridge inspection images.

Main Results:

• LCNet achieved superior performance compared to state-of-the-art methods (DeepLabV3+, UNet, UNet++, Swin-Unet) with an mIoU of 87.51% and Dice coefficient of 89.67%.
• The network demonstrates a favorable balance between accuracy and efficiency, with only 9.01M parameters and 19.03 GFLOPs.
• LCNet effectively suppressed pseudo-texture interference, preserved crack continuity, and accurately captured complex crack topology.

Conclusions:

• LCNet provides a robust and practical solution for concrete crack detection in challenging engineering scenarios.
• The network shows strong potential for UAV- and edge-device-based structural health monitoring applications due to its efficiency and accuracy.