Lightweight Dual-Backbone Detection Transformer for Infrared Insulator Detection

To address the challenges of high computational complexity and complex background interference in infrared insulator detection systems, especially on resource-constrained devices, this study proposes a lightweight dual-backbone detection transformer (LWDB-DETR). To extract target features from complex infrared backgrounds more accurately while reducing computational cost, we propose a multi-scale attention network (MSANet) as the adaptive backbone, which employs a multi-branch parallel structure and dynamically adjusts feature weights through a multi-scale gated attention mechanism. Additionally, we propose a contextual and geometric pyramid network (CGPN) to handle multi-scale feature fusion more effectively, which integrates context-aware modules with dynamic interpolation and implements geometric calibration through rectangular calibration modules. This enhances the fusion of low and high-resolution features while optimizing feature representation in infrared images with complex backgrounds. Furthermore, the lightweight context gate block (CGBlock) is introduced into the model, improving the model’s spatial awareness and detection accuracy by integrating gated convolution mechanisms and adopting adaptive feature channel weighting strategies. To validate our approach, an extensive infrared insulator dataset encompassing 3,790 images under various environmental conditions is constructed. Experimental results indicate that compared to RT-DETR-R18, the LWDB-DETR model reduces parameters by 53.2% and GFLOPs by 70.7% on the infrared insulator dataset. Moreover, mAP50 reaches 97.2%. On the CPLID dataset, it retains 94.3% mAP50 while maintaining computational efficiency. Deployment on the NVIDIA Jetson Xavier NX platform achieves 32.5 FPS, highlighting its effectiveness for real-time infrared insulator detection in power system monitoring applications.