A dual-attention guided network for robust crack detection in civil infrastructure

Ensuring infrastructure safety requires reliable crack detection in concrete structures. Yet, conventional approaches often fail to meet the efficiency and accuracy demands of practical engineering environments. While deep learning has exhibited remarkable capabilities, models still face performance bottlenecks in precise crack detection. To address these challenges, this paper proposes a dual-attention guided network named YOLO-DAGNet. First, an Anisotropic Attention and Depthwise-Separable Conv Modulation (AADM) module captures the elongated geometric characteristics of cracks, enhancing sensitivity to linear features. Second, a Dynamic Gated Attention (DGA) module enhances holistic recognition via self-attention; its dynamic gating mechanism filters background noise and improves computational efficiency. Third, a Global Query Learning (GQL) module coordinates these components, establishing a synergistic mechanism for coherent feature enhancement. Evaluations on a concrete crack dataset reveal that YOLO-DAGNet attains a mAP50 of 92.8% and a mAP(50–95) of 71.6%, surpassing the baseline by margins of 2.4% and 2.8%, respectively. Crucially, this high precision is achieved alongside a rapid inference speed of 141 FPS. Comprehensive benchmarking across three distinct datasets confirms the model’s consistent superiority and adaptability to diverse surface textures. Consequently, the proposed framework offers a reliable and efficient solution for real-time, non-destructive structural health assessments.