Real Time Cable Defects Recognition in Tunnels Using Improved YOLOv10n

Cable defects inspection in tunnels faces three challenges: scarcity of critical defect samples, feature degradation in low-light environments, and dense cable occlusion. This paper proposes a real-time recognition method based on improved YOLOv10n. Progressive Hard sample Mining (PHM) significantly enhances generalization performance for long-tail distributed critical defects. The Adaptive Dark Channel Prior (ADCP) is combined with an illumination-aware multi-scale Retinex algorithm to construct illumination-invariant features, overcoming texture loss in low-light conditions. A Cable-Oriented Feature Pyramid Network (CO-FPN) combined with axial deformable convolution strengthens linear structure modeling capability against occlusion interference. Verified by the test set, the model achieved an overall mean Average Precision at IoU threshold 0.5 (mAP @ 0.5) of 91.7%, with mAP @0.5 for loose connectors reaching 94.2%. The comprehensive False Positive Rate (FPR) for loose connectors dropped to 6.8% in low-light environments and 8.2% in high-occlusion scenarios. This method effectively enhances accuracy and robustness of cable defect recognition, providing a new solution for unmanned inspection in complex tunnels.