关键词: 绝缘子, InsuDet, EMSCP, RFCAConv, 多尺度特征融合

Abstract: Insulator defect detection is of great significance to ensuring the safe and stable operation of power transmission and distribution networks. In response to the shortcomings of traditional methods—such as poor performance in small target detection, insufficient multi-scale feature fusion, and a lack of synergy between classification and localization tasks—this paper proposes a model for insulator defect detection called InsuDet. First, to enhance detection performance for small targets, such as micro-cracks or localized damage, the RFCAConv module was introduced. This module integrates long-range information with spatial features within the receptive field, allowing the network to more precisely capture detailed features. Second, to address the issues of information loss during multi-scale feature fusion and inconsistent performance across different target sizes, we designed the EMSCP module to enhance the model’s perceptual and expressive capabilities, particularly for images with features at multiple scales. Furthermore, to optimize the enhancement of critical features and determine defect locations and extents through global context, the CAA_HSFPN neck network was proposed. Finally, to improve the interaction between classification and localization tasks and to adapt to the detection requirements in complex backgrounds with diverse target shapes, the Task-Aligned Dynamic Detection Head (TADDH) was designed. Experimental results demonstrate that InsuDet significantly outperforms existing methods on the REC-IDE dataset, achieving mAP50 and mAP50:95 values of 89.5% and 66.1%, respectively, while reducing the number of parameters to 1.81M. This model provides an effective and practical solution for insulator defect detection.

Key words: Insulator, InsuDet, EMSCP, RFCAConv, Multi-scale feature fusion