ID-RTDETR: An Efficient Cross-Scale Fusion Algorithm for Insulator Detection

Abstract

Insulator inspection is critical for power system maintenance, but existing methods have challenges in terms of efficiency and accuracy. We propose Insulator Detection Real-Time DEtection TRansformer (ID-RTDETR), an optimized RT-DETR framework with four key innovations for insulator detection. Firstly, Partial Reparameterized Block (PR Block) uses partial convolution with reparameterized in the backbone network, which reduces the computational cost and maintains the feature quality. Secondly, Focusing Diffusion Feature Pyramid Network (FDFPN) realizes the direct interaction of cross-scale features, which can effectively capture different defects. Thirdly, the Attention-based Intra-scale Feature Interaction efficient (AIFI-efficient) reduces the complexity of feature interaction through the additive attention mechanism. Fourth, the Generalized-Inner IoU (G-Inner IoU) loss function is proposed to combine global optimization with local refinement for accurate fault localization. Through the experiment, PR Block reduces parameters by 28.76% and computational cost by 24.78% while improving mAP50:95 by 1.62%. AIFI-Efficient enhances the inference speed of algorithms while maintaining accuracy. FDFPN enhances mAP50:95 by 3.10% through cross-scale feature interaction. G-Inner IoU increased by 2.74% in mAP50:95. The synergistic combination of all four modules achieves optimal performance. After experiments, compared with the baseline RT-DETR, ID-RTDETR improves the mAP50:95 by 1.8 percentage points, reduces the model size by 17.10%, and reaches 71.6 fps. ID-RTDETR provides an efficient and accurate solution for automatic insulator inspection.