Abstract
Regular tunnel inspections are essential for ensuring structural safety, yet current convolutional neural networks - or Transformer-based approaches struggle to balance accuracy, speed, and adaptability to complex tunnel environments. This study proposes an anchor-free selective state space (AFS3) framework specifically designed for real-time tunnel structure health monitoring. First, an anchor-free detection mechanism that directly regresses object boundaries from center points, enabling flexible handling of multi-scale and irregularly distributed defects. Second, a residual mamba module-based backbone that integrates selective state space modeling with residual learning to enhance long-range dependency modeling and fine-grained feature extraction while maintaining low computational cost. Besides, a novel C3OK2 neck incorporating omni-dimensional dynamic convolution, which dynamically adjusts kernel weights across all dimensions to improve multi-scale feature fusion under complex tunnel conditions. Finally, extensive experiments on a custom-labeled multi-device tunnel defect dataset demonstrate that AFS3 outperforms the latest YOLO11 and other surpasses existing state-of-the-art models in both accuracy and inference speed, showing strong potential for mobile deployment. To foster further research, the tunnel structure image dataset can be accessed at https://github.com/Kzz1yt/Tunnel_structure_image_data.
Funding Information
  • The National Natural Science Foundation of China (52362048)

This publication has 25 references indexed in Scilit: