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

Concrete crack detection in complex backgrounds remains challenging due to severe visual interference and limited computational resources on edge devices. To address these issues, this study proposes a lightweight crack segmentation network (LCNet) for accurate and efficient crack detection in real-world bridge inspection scenarios. LCNet adopts an asymmetric channel scaling strategy to improve computational resource allocation across network layers and introduces a bottleneck enhancement module to strengthen feature representation while maintaining low computational cost. Extensive experiments conducted on a self-constructed bridge crack dataset demonstrate that LCNet outperforms several state-of-the-art methods, including DeepLabV3+, UNet, UNet++, and Swin-Unet. Specifically, LCNet achieves an mIoU of 87.51% and a Dice coefficient of 89.67%, with only 9.01 M parameters and 19.03 GFLOPs, demonstrating a superior balance between accuracy and efficiency. Qualitative results further show that LCNet effectively suppresses pseudo-texture interference, preserves crack continuity, and accurately captures complex crack topology under challenging conditions. In addition, experiments on real bridge inspection images validate its robustness and practical applicability in engineering scenarios. Overall, LCNet achieves a favorable trade-off between segmentation accuracy and computational efficiency, demonstrating strong potential for UAV- and edge-device-based structural health monitoring applications.