LMSCD-Net: a lightweight multi-scale crack detection network for robust and efficient structural monitoring

Abstract Automated crack detection plays a crucial role in structural health monitoring and infrastructure maintenance. However, due to the irregular morphology of cracks and the complexity of the background, existing models often rely on deep network structures to extract high-level features while neglecting low-level information, resulting in blurred crack edges and loss of fine details. Additionally, the large number of parameters increases computational costs and introduces redundant features, making models more susceptible to interference in complex environments. To address these challenges, this paper proposes a lightweight multi-scale crack detection network (LMSCD-Net) that effectively balances the utilization of low-level details and high-level semantic feature extraction while reducing the overall model parameters. LMSCD-Net consists of three core modules: the Hyper Weave Downsampling Attention module, which enhances global feature representation by modeling long-range dependencies while compensating for the loss of low-level information; the Axial Decomposition Pooling module, which refines edge feature extraction to improve the structural integrity of detected cracks; and the Multi-scale Gated Enhancement module, which employs a gating mechanism to dynamically filter effective information, suppress redundant features, and enhance crack region responsiveness. To evaluate the effectiveness of LMSCD-Net in complex crack detection scenarios, extensive experiments were conducted on three publicly available datasets: GAPs384, CrackTree206, and DeepCrack. The results demonstrate that LMSCD-Net outperforms ten state-of-the-art crack detection models across multiple key metrics, achieving superior detection accuracy while maintaining a lower computational cost, thereby exhibiting enhanced robustness and generalization capabilities.