To address the challenges of multi-scale feature los and defect-background confusion in traditional steel surface defect detection under complex industrial scenarios, this paper proposes SMP-YOLO, a high-precision detection model based on the YOLOv11 framework. The model introduces three key innovations: (1) A spatial multi-state perception convolution module (SMSPDConv) that leverages a parallel pooling strategy—maximum, mean, and minimum pooling—combined with dynamic channel dimension adjustment, enabling robust extraction of both prominent defects and fine-grained textures; (2) A dual-domain spatial-channel mixed attention mechanism (DSCMA) applied in the neck to enhance defect saliency and improve feature fusion; and (3) A progressive bidirectional feature pyramid (PBiFPN) that incorporates cross-scale reuse and bidirectional connections to refine multi-level feature aggregation. Experiments show that SMP-YOLO achieves significant improvements, reaching 94.7% mAP@50 and 68.5% mAP@50-90, surpassing the YOLOv11 baseline by 3.4% and 6.4%, respectively, while boosting Precision by 6.9% to 92.4%.