MG-MetaDETR: Marrying Muti-Scale and Grouping Strategy With DETR for Few-Shot Insulator Defect Detection

Insulator defects pose a significant risk to the safe operation of transmission lines and can potentially lead to serious hazards within the power system. While UAV inspection technology has gradually replaced manual inspection of transmission lines due to its cost-effectiveness and efficiency, the limited number of samples for detecting insulator anomalies necessitates the development of a few-shot object detection algorithm with robust generalization capabilities, accuracy, and efficiency for effective power grid operation and maintenance. This paper presents the construction of a few-shot private dataset of insulator defects, based on insulator defect images obtained from a provincial power grid company during the period of 2020–2023. Secondly, the paper introduces a few-shot learning model, MG-MetaDETR, which is based on a multi-scale and grouping strategy for insulator defect detection. The proposed model leverages the end-to-end Transformer architecture DETR to streamline the object detection process. Additionally, it introduces the use of the multi-scale backbone network Res2Net-101 to enhance the detection of small target insulators. Finally, the paper employs a label grouping strategy to expedite the convergence of the model. The precision and recall rates of this method are reported as 72.4% and 75.4%, respectively, based on 50 training samples of each type.