Abstract: To address low efficiency and insufficient accuracy in traditional tunnel lining crack detection,a YOLOv10-CGBN network based on deep learning is proposed.The architecture incorporated a Context Guided Block (CG Block) that fused local and global information to enhance elongated crack detection accuracy.The integration of Bidirectional Feature Pyramid Network (BiFPN) into the feature fusion module enabled cross-scale bidirectional crack feature flow,capturing multi-scale crack information without substantially increasing computational costs.Experimental validation demonstrates that YOLOv10-CGBN achieves 90.5% AP50 and 67.1% AP50-95 on our tunnel lining dataset,surpassing baseline YOLOv10 by 5.5% and 4.4% respectively.The optimized network maintains 154.4 FPS inference speed with only 2.42M parameters,showing superior real-time performance compared to existing single-stage detectors.The results show that the method effectively balances accuracy and efficiency in practical tunnel inspection applications.