A Defect Detection Method for Catenary Components Based on an Improved YOLOv7

The convenience and comfort of high-speed rail have significantly improved long-distance travel experiences. However, faults in the catenary system can disrupt rail services, cause economic losses, and pose potential safety hazards. Given the complex field environment and the small size of the catenary components to be inspected, new challenges arise in the detection process. To address this issues, with the goal of locating U-shaped hoop and detecting faults in cotter pins, this paper proposes an improved YOLOv7 network. The ConvNeXt Block is introduced to replace the ELAN module in the backbone network, enhancing feature extraction capabilities. The SPD-Conv module is employed to reconstruct the model's downsampling, preserving more fine-grained information of small targets. Finally, the SimAM attention mechanism module is integrated to enable the model to focus on more meaningful feature maps, reducing redundant feature extraction. The proposed method was tested on a self-built dataset. Compared to the baseline YOLOv7 model, the proposed method achieved a 1.98% improvement in mAP, a 2.26% increase in precision, and a 1.53% enhancement in recall. This method maintains a good balance between computational efficiency and recognition accuracy.