To improve the accuracy of surface defect detection on steel while maintaining detection speed, this study proposes an enhanced RT-DETR detection model called FTD-DETR. First, images were obtained from a publicly available steel surface defect dataset, and data were partitioned and augmented, resulting in a steel surface defect dataset containing 2,000 images. The ResNet18 model, known for its low computational complexity and high detection accuracy, was chosen as the backbone feature extraction network. Then, a Faster-EMA module was introduced to update the basic blocks in ResNet18, enhancing the feature extraction speed of the model and improving inter-layer feature interaction. Finally, the AIFI module of RT-DETR was replaced with a Transformer with Deformable Attention encoder structure. This multi-head self-attention mechanism combined with dynamic attention further increases feature representation while reducing computational complexity. Experimental results show that FTD-DETR achieves a precision of 83.6%, recall of 67.7%, and mean average precision (mAP) of 79.3%. Compared to the baseline model RT-DETR, FTD-DETR significantly reduces parameters, floating-point operations, and memory usage while maintaining high accuracy. It features low complexity, high accuracy, and fast detection speed, providing technical support for steel surface defect detection.