MB-YOLO: a deep learning-based network for detecting cracks in highway tunnel linings

Abstract Efficient and accurate detection of lining cracks can provide a basis for assessing tunnel structural safety. To address the issues of complexity and poor generalization capability in traditional crack detection methods, this study proposes a deep learning-based network, MB-You Only Look Once (YOLO). Considering the elongated morphology of cracks, the network introduces the mixed local channel attention (MLCA), which combines channel and spatial attention at both local and global levels to enables the model to fully learn the contextual information of the crack regions, to enhance crack detection performance. Meanwhile, the PMLCA module is proposed to replace the PSA module for focusing on the regions where the cracks are located, thus reducing computational redundancy. In addition, the network also integrates bi-directional feature pyramid network in the feature fusion section, whose multi-level FPN and bi-directional cross-scale feature fusion mechanism can allow the crack feature information to flow bi-directionally across different scales and capture the crack information at different scales without additional significant computational costs. Experimental results on the self-collected dataset reveal that the AP50 and AP50-95 of MB-YOLO can reach 92.8% and 71.3% respectively, which is improved by 7.8% and 8.6% compared to the baseline model YOLOv10. MB-YOLO outperforms other single-stage target detection networks, such as YOLOv3, YOLOv5, and YOLOv9, in terms of accuracy, and its inference speed reaches 138.7 fps, which can realize real-time detection of tunnel lining cracks.