Surface defects in steel impact durability and availability, challenging industrial production. The similarity between small defects and the background hinders current detection methods from handling multi-scale variations and accurately locating defects, complicating recognition. To address these issues, we propose DMFSO-YOLO, a novel steel defect detection method. Within its backbone, the Speed-Optimized Precision Module (SOPM) enhances computational efficiency, reduces memory consumption, and mitigates overfitting. Additionally, the Dynamic Multi-scale Feature Fusion module (DMFF), designed in the YOLOv8 neck, improves feature extraction across dimensions and layers. The Normalized Gauss-Wasserstein Distance (NWD) loss function provides stable gradient feedback by accurately measuring the difference between predicted and actual bounding boxes. Experiments on the NEU-DET dataset show that DMFSO-YOLO achieves a 7.9% mAP improvement over traditional methods and reaches 245.2 FPS, highlighting its potential as a robust and efficient solution for real-time defect detection in industrial applications.