Railway fastener defect detection using RFD-DETR: A lightweight real-time transformer-based approach

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Manual inspection of railway fasteners is error-prone and inefficient.
• Existing deep learning models (e.g., YOLO) have limitations in defect feature extraction, model size, and computational demands.

Purpose of the Study:

• To introduce RFD-DETR, an optimized detection transformer model for real-time rail fastener defect identification.
• To enhance multi-scale feature extraction, model efficiency, and defect detection accuracy.

Main Methods:

• Developed RFD-DETR with three novel modules: Wavelet Transform Convolution (WTConv) for multi-scale feature extraction and computation reduction.
• Integrated Cross-Scale Feature Fusion (CSPPDC) with Channel Gated Attention Downsampling (CGAD) for refined defect detection.
• Incorporated Wavelet Transform Feature Upgrading (WFU) in the neck module for enhanced feature fusion.

Main Results:

• RFD-DETR achieved 98.27% mean average precision (mAP) at an IoU threshold of 0.5 on an expanded dataset.
• The model demonstrated superior performance compared to baseline models.
• Reduced computational expenses by 18.8% and parameter count by 14.7%.

Conclusions:

• RFD-DETR offers a highly effective and efficient solution for real-time rail fastener defect detection.
• The proposed model addresses the limitations of existing methods, improving accuracy and reducing resource requirements.
• This advancement contributes to enhanced railway track stability and safety through automated inspection.