Research on Crack Detection in Concrete Structures Based on Deep Learning Network Models

Abstract

As service life increases and concrete structures are exposed to variable loads and harsh environments, surface cracks inevitably develop. Cracks affect not only surface appearance but also the long-term durability of concrete components; therefore, accurate identification and real-time monitoring are essential throughout the engineering life cycle. To address the low efficiency of conventional manual inspection, the limited recognition accuracy for microcracks, and the inadequate feature-extraction and interference-suppression capabilities of the original YOLOv11 model in complex field environments, this study proposes an improved YOLOv11 model for concrete crack detection. First, the original backbone is replaced with FasterNet to reduce redundant computation and strengthen the extraction of shallow and deep crack features. Second, VoVGSCSP, VoVGSCSPC, and GSConv modules are used to reconstruct the neck network, replacing the default C3k2 and standard convolution blocks to improve multi-scale feature fusion for cracks of different widths. Third, TripletAttention is embedded in the detection pipeline to capture three-dimensional feature correlations, suppress complex background noise, and emphasise discriminative crack information. Experiments were conducted on a self-built dataset of 7,353 crack images, randomly divided into training, validation, and test subsets in an 8:1:1 ratio. Ablation and comparative experiments show that the improved model achieved higher precision and mAP@0.5 than the original YOLOv11. The method supports fine-grained detection of concrete microcracks under diverse and complex construction conditions and provides a feasible approach for automated structural-damage inspection.