RGS-DETR: an improved DETR framework for defect detection in power components

Defects in essential power infrastructure components, particularly insulators and wind turbine blades, compromise grid integrity and may cause large-scale failures. Nevertheless, DETR-based detection systems remain plagued by misidentification, omission errors, and heavy computational overhead. To address these drawbacks, this paper presents RGS-DETR, a defect detection algorithm integrating Receptive Field Attention based Generalised Efficient Layer Aggregation Network (RFA-GELAN) and Spatial Channel Reconstruction Convolution (SCConv). First, RFA-GELAN combines Receptive Field Attention (RFAConv) and Generalised Efficient Layer Aggregation Network (GELAN) to enhance feature extraction via receptive field optimization. Second, a Hierarchical Visual Context Aggregation (HVCA) neck framework with an explicit visual center module is constructed to strengthen small target recognition by synergising global and local contextual cues. Finally, the Spatial Channel Receptive Field Attention (SCRFA) module alleviates spatial-channel feature redundancy through decoupled refinement. A self-established insulator dataset with 4,832 annotated images is adopted for validation. Compared with RT-DETR, RGS-DETR achieves 84.4% mAP@0.5 and 85.5% accuracy, with 3.3% and 4.8% improvements, while reducing parameters by 35.7 M and GFLOPs by 107.6. Further evaluations on public benchmarks demonstrate robust performance, attaining 99.3% mAP@0.5 on the CPLID insulator dataset and 83.5% mAP@0.5 on the wind turbine blade dataset. These results verify that RGS-DETR achieves comprehensive improvements in the reliability and efficiency of surface defect identification.