RDX-Net: an efficient model for rail defect detection

Abstract Rail transit rails often develop defects like cracks, corrosion, and fissures over prolonged use, posing significant threats to travel safety. To address the limitations of the traditional YOLO11 model in detecting complex backgrounds, multi-scale targets, and small defects, this paper proposes an optimized detection model, RDX-Net. The model improves the direction-awareness effect of complex defective regions by designing the asymmetric filled multi-branch feature fusion module (C3K2-DAP); proposes the double-layer routing attention enhancement module (CRA) to optimize the attention mechanism and significantly improve the model’s feature modeling capability of the defective regions in complex backgrounds; and improves the downsampling modules (ADown and DRFD) to effectively retain multi-scale information and enhance the detection capability of tiny targets through the multi-path feature fusion strategy.Experimental results indicate that RDX-Net achieves a [email protected] of 69.8% on public rail defect datasets, improving by 2.6% over the original model, with precision increased by 2.0%. Additionally, the model’s parameters and computational load are reduced by 20.2% and 14.3%, respectively, achieving a lightweight design while maintaining efficient detection speed. The enhanced RDX-Net model excels in detecting complex backgrounds, multi-scale targets, and small defects, making it suitable for real-time defect detection in rail transportation. It offers efficient and reliable technical support to ensure rail safety.