DMDSNet: A Computer Vision-based Dual Multi-task Model for Tunnel Bolt Detection and Corrosion Segmentation

The inspection of corroded bolts is a significant issue in the Structure Health Monitoring (SHM) of subway tunnels. However, detection-only methods may result in missed detection of corroded bolts due to the small rust area. To address this challenge, the present study ingeniously divides the task into two parallel tasks: bolt detection and pixel-level rust segmentation. By taking the intersection of the two tasks, we achieve the goal of improving performance. Specifically, we propose a Dual Multi-task Detection and Segmentation Network (DMDSNet) for tunnel bolt maintain, which helps to reduce false and missed detection rates. The detection branch incorporates a coordinate attention module that enhances the focus on bolts in tunnel patches, while the segmentation branch adopts a cross-stage partial-based decoder that increases the accuracy of determining whether a pixel belongs to the corrosion area. Both branches share the same backbone, which simplifies the model. The superiority of our proposed algorithm is demonstrated through sufficient comparisons and ablation experiments using our corroded bolt dataset, which was captured from a real subway tunnel and is publicly available at https://github.com/StreamHXX/Tunnel-lining-disease-image.