Abstract: The catenary system, as an essential component of electrified railways, provides energy to trains and ensures their normal operation. Damage to catenary system components poses a threat to train safety, making it crucial to regularly inspect the condition of these components. In recent years, UAVs have been widely used in monitoring the condition of critical catenary components. However, due to the complex and variable backgrounds, significant scale changes, and the presence of many small targets in the catenary images captured by UAVs, existing detection algorithms frequently suffer from false detections and missed detections of catenary components. To address this issue, this paper proposes a catenary component target detection method based on an improved YOLOv7 algorithm. By introducing an enhanced receptive field module, the network's feature extraction capability is strengthened, leading to more discriminative target feature representations. Additionally, an improved coordinate attention mechanism is incorporated during the fusion of adjacent scale feature maps to highlight the target features of catenary components and suppress redundant background information. The bounding box loss function is optimized using the Wasserstein distance, effectively improving detection accuracy. Experiments on the catenary component dataset show that the improved YOLOv7 algorithm can accurately detect various catenary components in drone-captured images, achieving a mean average precision of 97.27%, which is 3.83% higher than before the improvement. The proposed algorithm enhances the high-precision and rapid detection capabilities of drones for critical catenary components, providing technical support for achieving more intelligent drone inspections.

Key words: object detection, YOLOv7, contact network component recognition, feature extraction, attention mechanism, loss function