HEYOLO-IDD: An efficient insulator defect detection network based on the improved YOLOv11n

The automated detection of insulator defects is a crucial component in driving the intelligent development and construction of new power systems. To address limitations such as accuracy and generalisation in insulator defect detection, this paper proposes HEYOLO-IDD, a high-efficiency detection model based on YOLOv11n. Firstly,KernelWarehouse-Conv(KWConv)is introduced as a downsampling module to reduce redundant parameters in the backbone network and lower computational complexity. Secondly, to address the characteristics of defects, a Partial Multi-Scale Conv(PMC) module is designed to replace the Bottleneck in CSP. Combined with TripletAttention, this forms the CSP-PMC-TripletAttention(CPT) module, which enhances multi-scale and small-object feature extraction capabilities without increasing the parameter burden. This enables an innovative fusion of the backbone network based on KWConv and CPT. Finally, the bi-directional feature pyramid network (BiFPN) neck network is adopted to enhance feature fusion and simplify the architecture. On our self-built HEData-2025 dataset, compared to YOLOv11n, mAP(0.5) improves by 7.2% whilst the number of parameters is reduced by 23.2%, demonstrating superior overall performance. The generalisation capability is further validated on the public VisDrone-2019 dataset.