GLE-YOLO:A Lightweight and Efficient Defect Detection Model for Power Inspection

The research on defect detection in transmission lines and substations is critically important to ensure continuous power delivery. However, current research faces a series of obvious limitations: performance degradation caused by inter-class sample imbalance and intra-class sample variance, conflicts between detection accuracy and speed, and limited detectable defect categories. This study proposes a novel detection model called Generalized, Lightweight, and Efficient YOLO (GLE-YOLO). By including the Generalized Efficient Layer Aggregation Network (GEALN), the model can effectively capture multi-scale features from every sample to better deal with defect categories with limited samples or large sample variance. A lightweight decoupled detection head structure is designed to improve performance while reducing computational complexity. The model is further lightweight by the use of partial convolutions. Furthermore, the bounding box loss function is optimized to improve the convergence speed. To demonstrate effectiveness of the proposed model, we construct a rich-scenario, large-capacity, high-quality defect dataset, including ten categories for training and testing. Experimental results show that our model performs very well on our dataset, averaging 88.7% precision, 77.5% recall, and 90.4 FPS processing speed. This research provides an effective solution for the defect detection of substations and high-voltage transmission cables.