Insulator Defect Detection Method Using Heterogeneous Dual-Stream Collaborative Network in Complex Scenes

The safe operation and maintenance of power systems rely heavily on accurate defect detection of key equipment like insulators. Traditional manual inspections suffer from low efficiency and high rates of missed detections. With the advent of deep learning, insulator detection accuracy has improved significantly. However, in complex outdoor scenarios, challenges such as background interference and object occlusion cause a notable decline in the performance of existing visual algorithms, leading to a missed detection rate for insulator defects as high as 42%. This study employs an enhanced Swin-Transformer architecture as the backbone network and proposes a heterogeneous dual-stream network integrating detection and segmentation functionalities. The dual-stream fusion module, designed using a Graph Neural Network, effectively enhances recognition accuracy under object occlusion. Experiments show that this method improves mAP by $\mathbf{1 1 \%}$ in complex scenarios, providing a robust solution for UAVbased power inspections. This advancement reduces the risk of power grid failures and promotes the engineering application of visual detection technology in the energy sector.