Cable Defect Detection Using an Improved RT-DETR Model

Abstract

Small defect targets and complex backgrounds in transmission-line inspection images can limit detection accuracy, and existing detection models often incur substantial computational cost. To address these issues, this study proposes a cable defect detection method based on an improved RT-DETR. First, a DualConv module is introduced into the backbone to strengthen local feature representation while reducing parameter redundancy. Second, a ContextGuidedDown module is incorporated for downsampling to alleviate the loss of critical information. Finally, Dynamic Tanh replaces LayerNorm in the attention-based intra-scale feature interaction (AIFI) encoder to improve the regulation of high-level features. Experiments were conducted on the CableInspect-AD dataset and the RF100 Cable Damage dataset. In the main single-run comparison on CableInspect-AD, the proposed method achieves an mAP@0.5 of 92.4%, which is 2.7 percentage points higher than that of the original RT-DETR, while the number of parameters is reduced from 20.1 M to 18.6 M and the computational cost decreases from 58.3 to 53.4 GFLOPs. Across four paired random-seed runs, DCGD-RTDETR achieves a mean mAP@0.5 of 92.05 ± 0.34%, compared with 90.98 ± 1.10% for RT-DETR; however, the paired difference is not statistically significant (p = 0.086). In the same repeated-run analysis, AP@0.75 decreases from 65.01 ± 1.43% for RT-DETR to 63.26 ± 1.11% for DCGD-RTDETR (p = 0.005), indicating that localization performance at a stricter IoU threshold remains a limitation. Ablation experiments further confirm the complementary contributions of the three modifications. Overall, the proposed method provides a favorable detection-accuracy and model-complexity trade-off at IoU = 0.5, while further improvement is needed for stricter-IoU localization performance.