Real-Time Tunnel Crack Detection Based on FEA Stress Heatmaps and Improved Faster R-CNN

This paper proposes a real-time crack detection method for complex tunnel environments. By integrating finite element analysis stress fields with visual data, three optimizations are made to Faster R-CNN: 1) Adoption of a computationally more efficient ResNeXt-101 backbone network; 2) Introduction of a stress heatmap-based attention mechanism in the region proposal network to prioritize high-stress regions; 3) Reconstruction of anchor parameters and adoption of depthwise separable convolutions to reduce computational complexity, tailored to the slender morphology of cracks. Experimental results on the Tunnel3D-Crack dataset demonstrate that the proposed method achieves a 94.6% , with a recall rate of 85.7% in high-stress regions, and realizes real-time detection performance of 16.3 FPS on the Jetson AGX Xavier edge device, significantly outperforming the baseline models.