<p>Steel surface defect detection faces critical challenges including category diversity, extreme scale variations, and weakened morphological features in complex backgrounds. To address these issues, we propose MCD-RTDETR, an enhanced real-time detection algorithm based on RT-DETR-R18. The algorithm introduces three key innovations. First, a Cascaded Dynamic Multi-Scale Convolution Module achieves efficient multi-scale feature extraction through progressive channel partitioning and spatial-channel synergistic attention. Second, a Position-Aware Attention mechanism with Dynamic Position Bias explicitly models spatial relationships between defects to distinguish visually similar categories. Third, a Dynamic Adaptive Multi-Scale Feature Pyramid Network adaptively integrates heterogeneous features while preserving fine-grained details through learnable weighted fusion. Extensive experiments on GC10-DET demonstrate that MCD-RTDETR achieves 68.9% mAP, representing a 3.6% improvement over the baseline. The model reduces parameters by 37% to 12.52M and computational cost by 29.3% to 40.3 GFLOPs, while maintaining real-time performance at 45.2 FPS. Cross-dataset validation on NEU-DET and DeepPCB achieves 74.6% and 98.3% mAP, respectively, demonstrating excellent generalization capability. The proposed method effectively addresses multi-category defect detection challenges while significantly reducing computational burden, providing a practical solution for industrial deployment.</p>