Small Target Detection for Power Line Components in Complex UAV Imagery

Transmission lines play a critical role in the stable operation of the power grid, and the proper functioning of their components is equally essential. However, these lines are often situated in harsh environments and continuously exposed to external challenges such as lightning strikes, heavy rain, and environmental contamination. Accurate detection of component damage is crucial to prevent potential safety hazards and equipment failures caused by transmission line faults. To address the limitations of existing algorithms—particularly their low efficiency in detecting small objects and poor accuracy in complex backgrounds—we propose a defect detection model for transmission line components based on YOLOv5. First, to enhance the model’s ability to detect small and low-resolution targets, we incorporate SPDConv into the backbone of the feature extraction network, significantly improving the extraction of small object features. In the Neck, we replace the original C3 module with the lightweight C3-Faster module to strike a balance between accuracy and model efficiency, thereby reducing latency without sacrificing performance. Moreover, the original YOLOv5 detection head integrates classification and localization, which helps reduce parameter count but introduces mutual interference between the two tasks, potentially compromising accuracy. To mitigate this issue, we redesign the decoupled detection head to decouple classification from localization. As a result, our improved model achieves a 3.88% increase in mAP, a 4.85% boost in precision, and a 3.96% improvement in recall compared to the baseline YOLOv5. Under identical conditions, our algorithm outperforms existing methods in terms of accuracy and exhibits superior small-target detection capability in complex backgrounds.