MSDNet: A Multi-Scale Feature Representation Network Model for Tunnel Bolt Detection

Bolts are critical components of tunnel linings, essential for ensuring safe tunnel operation. Due to their susceptibility to corrosion and detachment, it is vital to closely monitor them during maintenance. However, identifying a large number of corroded bolts presents challenges, particularly due to interference from complex background noise and the inherent limitations of CNNs in extracting local features. To address these issues, we introduces a novel multi-scale feature extraction detection network (MSDNet) designed to improve tunnel bolt maintenance by reducing false positives and missed detections. We incorporate a pyramid structure within the visual transformer framework to overcome CNNs’ limitations, enabling the generation of multi-scale corroded bolt feature maps that emphasize global information. Additionally, we develop a feature enhancement module (FEM) to capture more detailed features in small or localized areas during the search stage. Lastly, we design an efficient feature aggregation modules (FAM) to fuse coarse-level semantic corroded bolt information with fine-level features in a top-down pathway. We collected and organized a dataset of corroded bolts, and conducted comparative and ablation experiments to demonstrate the effectiveness of our proposed model. The results show that our model performs better in handling the challenges posed by this dataset.