A Small-Sized Defect Detection Method for Power Line Insulator Using Multiscale Feature and Lightweight Networks in UAV-Vision

Insulator defect detection (IDD) of transmission lines is crucial for power system safety and stability, with Un-manned Aerial Vehicle (UAV)-based machine learning methods becoming prevalent. However, these methods face challenges due to deep neural networks’ high complexity/low detection speed and the difficulty in detecting small defects within complex UAV backgrounds. To address these issues, this article proposes a Multiscale Insulator Defect Detect Network (MID-Net) for IDD, comprising a fully connected GhostNet (FCG-Net) backbone, a generalized multiscale spatial convolution (GMSC) module, and a category prediction head (CPH) module. First, the lightweight FCG-Net backbone network is designed to reduce the network’s computational cost. Second, the GMSC module enhances small defect feature representation and enables interactive multiscale feature fusion. Finally, the Focal-EIOU loss in the CPH module improves model convergence and robustness. Experimental findings demonstrate the proposed approach’s effectiveness in both detection accuracy and efficiency.