<p>Railway turnouts are of critical importance to the railway track system. Defects in them have the potential to result in severe safety incidents and significant property damage. Conventional detection techniques are prone to inaccuracy and lack the capacity for real-time performance due to the irregular distribution and varying dimensions of turnout defects, in addition to the influence of changing lighting conditions and complex backgrounds. In order to enhance the efficacy of railway turnout defect detection, this study puts forth a high-precision detection model, designated as SMP-CGFM-Inner-GIoU-DETR(SCG-DETR), which is founded upon the RT-DETR architectural framework. The initial stage of the process involved the development of the C2f-SMP feature extraction module, which was created using the SMPConv and C2f modules. This module optimises the aggregation of local and global information through successive convolutions and gating mechanisms, thereby enhancing the receptive field and feature extraction capabilities. Secondly, the CGFM module was introduced, incorporating the CAA attention mechanism to optimise the weighting of feature maps, thus enhancing fine-grained extraction and contextual information integration. Finally, the original loss function was replaced with an Inner-GIoU based on auxiliary bounding boxes, which accelerates model convergence and improves accuracy. The experimental results demonstrate that the proposed enhanced model achieves an mAP50 score of 71.0%, representing a notable increase of approximately 4.6% compared to the original model. Additionally, the detection speed reached 80.5 FPS, meeting the real-time requirements for engineering applications.</p>