Experiment configuration and model parameters.

Railway fasteners play a crucial role in ensuring track stability and safety; however, manual inspection is both inefficient and prone to errors. Deep learning models, such as YOLO, are widely utilized for defect detection. However, these methods demand comparatively larger model sizes, greater computational power, and more storage capacity, yet their defect feature extraction remains inadequate. To overcome these challenges, this research introduces RFD-DETR, an enhanced detection transformer model specifically optimized for real-time rail fastener defect identification. The model incorporates three distinct modules to facilitate multi-scale feature extraction, enhance model efficiency, and improve defect detection. Firstly, a wavelet transform convolution module (WTConv) is employed, which integrates a wavelet transform to enhance multi-scale feature extraction while reducing model computation. Secondly, a cross-scale feature fusion module (CSPPDC) is utilised, incorporating channel gated attention downsampling (CGAD) to refine defect detection. Finally, a wavelet transform feature upgrading (WFU) module is integrated within the neck module, enhancing feature fusion and contributing to the overall efficacy of the model. Experimental findings based on an expanded rail fastener dataset indicate that RFD-DETR achieves a 98.27% mean average precision (mAP) when evaluated at an IoU threshold of 0.5, outperforming the baseline model. Furthermore, it lowers computational expenses by 18.8% and reduces parameter count by 14.7%.