PSNet: Local Context Perception and Global Semantic Information Extraction Based Detector for Power Scenario Defects Detection

The rapid development of power systems has made inspection tasks increasingly challenging. Effective intelligent detection methods for power system defects can significantly enhance the efficiency of power inspections. The complexity of power scenario images, the diversity of defects, and the low distinction between foreground and background all contribute to the difficulty of power system defect detection. Addressing these issues, this paper analyzes the characteristics of power scenario images collected by Unmanned Aerial Vehicles (UAVs) and designs a defect detection network for the power scenario (PSNet). To mitigate the challenge of some objects being small and highly correlated with their surroundings in power scenario images, a Local Context Enhancement Block (LCEB) based on dilated convolutions is proposed. Given the typically complex backgrounds in power scenario images, a Global Weighted Attention Block (GWAB) is designed. Additionally, an Electric Power Scenario Defects Detection Dataset (EPSD) comprising images collected by UAVs is introduced, which contains defect information in various common power scenarios. Our method demonstrates superior performance in power scenario defect detection tasks compared to current advanced detection methods. We also conducted ablation experiments, and the results demonstrate that each proposed module enhances the model's performance. Additionally, experiments on public dataset further demonstrated the generalization performance of the proposed method.